TinyML-enabled fuzzy logic for enhanced road anomaly detection in remote sensing

Abstract

Advanced techniques for detecting and classifying road anomalies are crucial due to road networks’ rapid expansion and increasing complexity. This study introduces a novel integration of Tiny Machine Learning (TinyML), remote sensing, and fuzzy logic through a fully connected U-Net architecture, TinyML-U-Net-FL, tailored for anomaly detection in resource-constrained environments. Our framework addresses critical gaps in existing methodologies, such as high computational demands and limited real-time processing capabilities, by leveraging model compression, quantization, and pruning techniques. These enhancements facilitate efficient real-time analysis directly on edge devices. In rigorous evaluations using the DeepGlobe and Dubai aerial imagery datasets, our framework achieved a notable recall of 92.4%, precision of 78.2%, and an F1-Score of 84.7%, demonstrating superior performance compared to contemporary methods, including DCS-TransUperNet, GOALF, GCBNet, DiResNet, and ScRoadExtractor. Incorporating fuzzy logic significantly improves the robustness of anomaly detection, enabling more precise and reliable classification. This research contributes substantially to intelligent transportation systems by facilitating precise, energy-efficient, timely detection and classification of road network irregularities, enhancing infrastructure management road safety, and supporting autonomous navigation applications.

Introduction

Tiny Machine Learning (TinyML) describes the ability to perform machine learning on microcontrollers and other tiny devices with limited resources, such as power consumption, memory size, etc. This combination allows for utilizing machine learning compactly and efficiently¹. The importance of this convergence is growing as the Internet of Things (IoT) expands, necessitating edge processing solutions that address the inherent limitations of latency, privacy, and connectivity in centralized systems². Extracting road networks from remote sensing images is vital in various applications, including urban planning, traffic management, and disaster response. The utilization of TinyML in identifying irregularities within road networks represents a significant advancement for Intelligent Transportation System (ITS)³, expanding their functionality beyond conventional vehicular scenarios⁴, as can be elaborated in Fig. 1.

Fig. 1

TinyML-enabled physical and digital AI.

Fig. 1 visually illustrates the integration of TinyML into physical and digital Artificial Intelligence (AI) frameworks. It depicts transitioning from traditional cloud-based systems to more localized, edge-based processing using TinyML, significantly reducing latency. This figure highlights how TinyML allows for the real-time processing of data collected from sensors directly on edge devices, enabling quicker and more efficient road condition assessments. The visual representation emphasizes the role of TinyML in allowing smart devices to perform complex computations locally rather than relying on distant cloud servers. This capability is crucial for applications where rapid decision-making is essential, such as in the monitoring and management of road networks⁵. Nevertheless, the conventional approaches employed for extracting road networks from remote sensing images frequently encounter challenges due to the intricate nature of the task, necessitating substantial human supervision^6,7. Although machine learning techniques possess advanced capabilities, they are not exempt from certain limitations, such as the need for significant computational resources and difficulties interpreting their outcomes. Despite advancements in road anomaly detection using machine learning techniques, several limitations remain unaddressed in prior studies. Specifically, conventional approaches often require extensive computational resources and high-powered processing, making real-time deployment challenging, especially on resource-constrained edge devices.

Furthermore, existing methods frequently struggle with ambiguity and uncertainty inherent in remote sensing data, limiting their effectiveness in practical, diverse scenarios. This research introduces a novel integration of TinyML, remote sensing, and fuzzy logic, significantly reducing computational complexity and enabling real-time anomaly detection directly on edge devices to bridge these gaps. The proposed TinyML-U-Net-FL architecture explicitly addresses issues of computational overhead, ambiguity handling, and efficient model deployment, thereby advancing the capabilities of ITS. Recent advancements such as cloud-edge collaborative frameworks have explicitly addressed limitations related to computational resources and real-time deployment, enhancing anomaly detection performance significantly⁸. Furthermore, integrated fuzzy logic with machine learning methods has explicitly addressed uncertainties in remote sensing data, further improving reliability in complex urban environments. These insights inform our explicit integration of TinyML, fuzzy logic, and remote sensing technologies. Therefore, this research’s primary objective is to develop and validate the performance of this integrated approach, demonstrating its efficacy for practical real-time road anomaly detection in diverse and resource-constrained environments. The field of TinyML tackles these challenges by facilitating computation on edge devices, promoting energy efficiency, minimizing latency, and enhancing privacy. The remarkable aspect of this technology lies in its capacity to respond effectively to the ever-changing conditions of road environments in real-time. The vast amount and complexity of remote sensing imagery make machine learning a valuable tool for inferring from data to predict future events that have not been encountered before⁹. Applying TinyML necessitates tuning models for the severe constraints of small devices and enables real-time anomaly detection and classification over diverse data sources^10,11,12. For example, the TinyML framework can independently identify road configuration in satellite or aerial imagery, making it likely that lightweight and low-power devices can do analysis instantly. Meanwhile, we learned that developing deep learning, such as Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), remarkably improves recognition on road networks^11,13,14.

These developments have been a significant milestone for improved methods of analyzing two-dimensional (2D) and, more recently, three-dimensional (3D) data. It has been continuing with the development of advanced architectures and attention mechanisms, which in turn is resulting in constant improvements here. As shown in various studies, integrating fuzzy logic inference with TinyML approaches has proven to address the noisy nature of remote sensing images efficiently^15,16. The model can be compressed and well-suited to on-device settings with potent hardware optimization. This integration is an excellent example of TinyML, an emerging technology for real-time on-device processing¹⁷. The collaborative progress of TinyML, remote sensing, and fuzzy logic applied to failure detection/categorization in road networks. Enabling algorithms locally installed on devices with limited resource levels reduces the reliance on cloud computing, which results in a more energy-efficient system, decreased delay time, and increased privacy¹⁸. Remote sensing is an approach for capturing various complex geographical images. In contrast, fuzzy logic manages the inherent imprecision¹¹ in images by using fuzzy sets and rules^13,19. This method improves the road inference process. The consequent approach is slightly different as most practices are still labor-intensive, which we want to eliminate. Influence of TinyML on the overall system in our results show that by incorporating it as a support approach for our Fuzzy Inference System (FIS) and thereby including deep learning techniques, we can achieve better accuracy and reliability compared with current methodologies^7,20. It highlights how such integrated technology can impact ITS. TinyML and FIS are proposed to develop the detection. Road networks are derived from remote sensing, illustrating potential applications for anomaly detection in Fig. 2.

Fig. 2

Road anomalies detection methods.

Fig. 2 comprehensively illustrates various road anomaly detection methodologies in ITS, categorizing them into traditional, deep learning, and hybrid approaches incorporating fuzzy logic and TinyML. It specifically highlights the pathway employed in this research, combining remote sensing, fuzzy logic inference, and TinyML-enabled neural network architectures. This visual summary aids in contextualizing our methodological innovation within existing techniques, demonstrating our model’s unique integration and superior effectiveness in managing ambiguity and real-time constraints. The diagram breaks down the detection process into stages, starting with remote sensing for data acquisition, then threshold-based, feature extraction, deep learning techniques for data processing, and anomaly detection, concluding with model generation and validation steps. This detailed flow helps to outline the end-to-end process for identifying road anomalies, emphasizing the importance of integrating these technologies to enhance accuracy and efficiency in detection systems. The all-encompassing process of road network anomaly detection includes incorporating TinyML applications of geographic information system remote sensing and fuzzy reasoning in geosciences. This integration identifies road types and surface damages, such as potholes, cracks, or ruts^7,21. The use of remote sensing technology enabled the extraction of high-resolution imagery that can be captured using aerial platforms like aircraft or satellites and detecting road networks due to their linear patterns, which contrast with the surrounding natural environment. Edge task-oriented devices with limited computational resources require TinyML models, which are highly efficient versions after being trained on a more extensive dataset^4,22,23.

There needs to be a balance between complexity, accuracy, and robustness against future lighting or weather conditions changes for maximum model performance. One of the common uses it is classifying road types and having better classes that all depend on fuzzy logic using predefined sets or thresholds, where FIS contribute to managing uncertainty or imprecision in data so it can be used for defining more refined roads classifications, which are underestimated by classical methodologies^24,25. Fuzzy sets and rules for detecting the variability in road attributes have been successfully incorporated into such systems to make complex segments easier²², facilitating an interpretation of ambiguous or poorly contrasted features, as mentioned above. Compounding these challenges is the pressure to use more complex models and techniques like knowledge distillation or neural architecture search. These are ways to maximize TinyML model efficiency^23,26. These technologies integrate to provide prompt and energy-efficient detection and classification of anomalies along with fuzzy systems for improved resiliency and adaptability in suboptimal conditions²⁷. Recent improvements in this field include automatic model design and transfer learning, which helped the increased use of TinyML across different application areas. This integration provides significant advantages over traditional methods, allowing for real-time data processing at or near the data being created.

Literature review

Road extraction from remotely sensed images has many potential applications in urban planning, traffic management, and emergency response. It is known that the images of roads are intricate, and diverse features exist on roads in different contexts, which makes this work becomes tough²⁸. Nevertheless, it is imperative not to underestimate the importance of the task. This literature review comprehensively examines the progression of methodologies in analyzing road networks. The gaps intended to be addressed by TinyML and FIS are emphasized. The early methods for analyzing road networks used spectral information and geometric characteristics to extract pertinent features. The methodologies for examining road networks were first developed in the 1990s. The effectiveness of these methods was found to be high in urban areas that exhibit a clearly defined spatial organization. However, their efficacy was limited in settings exhibiting more incredible intricacy and complexity. Traditional techniques encompass various methods, such as edge detection and morphological operations. Machine learning encompasses multiple strategies, such as supervised and unsupervised learning, alongside more sophisticated approaches like semi-supervised and reinforcement learning²⁹. The methods can be broadly classified into conventional and machine learning. Mnih and Hinton produced research showcasing a significant field progression by employing deep learning techniques and CNNs³⁰. These advancements have greatly facilitated the processing of high-resolution remote sensing data for road detection. Deep learning exhibits superiority in feature representation and learning automation, thereby preventing the necessity for manual feature engineering³¹. Nevertheless, these models necessitate a significant quantity of annotated data and exhibit a high degree of sensitivity to both noise and domain shifts. The high computational requirements of these models pose a challenge to achieving real-time analysis. The emergence of TinyML can be attributed to the computational challenges that have arisen in recent times. This technology facilitates the execution of machine learning models on devices characterized by limited power consumption and low latency, such as microcontrollers³². As elucidated by Banbury, the paradigm shift enables the local processing of data on the device, thereby preventing reliance on cloud resources and facilitating real-time road network analysis³³. TinyML presents the potential for unsupervised learning^29,34, cognate to the TEDA algorithm, enabling the direct identification of road anomalies using data acquired by sensors-embedded vehicles. This intervention enhances the level of road safety while concurrently reducing the expenditure associated with road maintenance. Within the realm of TinyML, various techniques have been devised to enhance the efficiency of models for deployment on resource-constrained devices. These techniques encompass model compression, quantization, and pruning, all of which aim to optimize the performance of models within the limitations imposed by these devices²³ as expounded upon the techniques in question^35,36. As mentioned, the methods aim to achieve a harmonious equilibrium between accuracy and effectiveness while upholding the models’ privacy and the data’s reliability. Despite the considerable advancements achieved in this domain through conventional and machine learning techniques, integrating TinyML and fuzzy logic systems presents a fresh and inventive strategy to tackle the existing limitations in road network extraction³⁷.

This phenomenon emerges due to integrating these systems, which effectively amalgamates the advantages of conventional methodologies and machine learning techniques. The present study thoroughly examined the evolution of methods employed in extracting road networks. Furthermore, a comprehensive assessment was conducted to ascertain the potential of the proposed integration. The investigation has identified particular areas of research that necessitate further examination to advance the field. Integrating TinyML, fuzzy logic, and remote sensing presents a promising prospect for enhancing the classification and recognition of diverse road types and anomalies^38,39. The statement above presents evidence indicating that this phenomenon holds substantial potential for facilitating future research and advancements in the field, thus enabling subsequent academic investigations. Recent literature significantly contributes to current advancements. Optimized binary neural networks explicitly demonstrate substantial efficiency improvements, enabling effective deployment on resource-constrained edge devices³⁹. Hybrid approaches integrating fuzzy logic explicitly enhance accuracy and robustness in anomaly detection by handling data uncertainty effectively⁴⁰. Additionally, incremental learning algorithms tailored explicitly for TinyML have significantly improved anomaly detection capabilities in dynamic environments and limited-resource scenarios^41,42. The application of deep learning methodologies explicitly highlights advanced methods for accurately detecting infrastructure anomalies like road cracks from remote sensing data⁴³. Comparative analyses of anomaly detection methods in IoT networks explicitly underline the robustness and applicability of various machine learning techniques in real-world scenarios⁴⁴.

Recent advancements in road network anomaly detection using advanced computational techniques have shown promising results. For example, recent studies have explored deep learning methodologies optimized for real-time road condition analysis using lightweight models suited for edge deployment⁴⁵. Other contemporary research efforts have investigated integrating cloud-edge collaborative frameworks, significantly enhancing anomaly detection accuracy and reducing response latency. Additionally, emerging fuzzy logic and hybrid machine learning approaches have demonstrated improved capabilities in handling uncertainties in remote sensing imagery, instrumental in diverse and complex urban environments⁴⁶.

Research methodology

Various methods are emerging for road detection, such as extracting features from images, both manually and automatically. Yet, achieving complete extractions of required features in road network anomalies from crack detection to intrinsic and occluded object, still in progression by using various techniques. A significant challenge here includes occlusions and intrinsic objects that cause uncertain and incomplete detections. The comprehensive depiction of the suggested framework’s systematic structure is elaborated extensively in Fig. 3.

Fig. 3

Proposed framework of TinyML with remote sensing and fuzzy logic.

This Fig. 3 demonstrates acquiring aerial/satellite imagery with obstructions and intrinsic objects. The first step within the cloud phase is pre-processing, which entails image rectification for further analyses to enhance data quality. This step is paramount because subsequent stages guarantee detecting and classifying features within the imagery. Segmentation follows pre-processing. The Gabor Filter, a well-known edge and texture analysis, performs segmentation. This step assists in feature identification, including occlusions, intrinsic objects, and other road anomalies. The images are first processed and then further analyzed using TensorFlow, which is known as one of the most potent machine-learning tools. In this phase, TensorFlow forms and sharpens the data for classification capabilities. TinyML, a machine learning system built for small handheld devices with limited power and processing capabilities, receives the processed data in the edge phase^4,47,48.

It guarantees the agile performance of the system in constrained environments with limited resources. A U-Net framework with fuzzy logic at the classification component of the anomaly detection system implements the step, which serves as the core anomaly detection mechanism. The combined approach enables accurate segmentation and classification of road anomalies in complex or straightforward road network imagery. The system then consolidates the classification results with tensor voting to increase the dependability of the results and accuracy by cross-sourcing multiple votes. Subsequently, the system runs a FIS to predict and analyze the data⁴⁹. When a set condition is fulfilled within the fuzzy inference system marked with a ’Y’, they are assigned for dimensioning, and the output produced is categorized as detected anomalies. On the other hand, when in fulfillment marked with ’N’ of the condition set, the process stalls until some adjustments are made or loops back the contact return to further analysis for exact results. Fuzzy logic shows how such techniques can cope with the inaccuracies of remote sensing data containing occlusions and intricacies of nature that can conceal essential features^2,50,51,52. This broad underpinning prepares raw data for advanced processing techniques, which are described in more detail in Fig. 4. Our optimization approach incorporates techniques informed by recent advancements, such as optimized binary neural networks designed explicitly for resource-constrained anomaly detection³⁹. Incremental learning algorithms for TinyML environments guide our model’s adaptability and real-time responsiveness⁴¹. In contrast, recent hybrid machine learning methodologies provide robust and precise anomaly detection performance tailored for automated road infrastructure inspection⁴³. As depicted in Fig. 3, the basic data processing steps and the anomaly detection process are completed. The data undergoes additional refinement and validation, as shown in Fig. 4. This progression from coarse-scale detection to fine-scale grained classification and optimization strategically shifts the focus from a TinyML and fuzzy logic theoretic approach to real-world practical application.

Fig. 4

TinyML integrated process for road network anomalies detection.

Integrating processes in this study with data collection, geometric and radiometric correction, quantization, and optimization, as illustrated in Fig. 4, enhances the automated workflow analysis, clearly outlining the interdependencies among multi-step procedures.

Such workflows demonstrate the impact and contribution of TinyML in conjunction with fuzzy logic and anomaly detection algorithms in TensorFlow upon detecting and classifying road anomalies. The figure illustrates how these technologies work together to enhance the precision of detecting and classifying road anomalies in diverse environmental settings. The process illustrated in Fig. 4 starts with collecting remotely sensed data, which is refined through a defined pre-processing work sequence that includes image corrections to enhance data quality and reduce noise. This step is crucial to ensure the data set is helpful for analysis. After this, the data set is split, TensorFlow is used for the model training validation, and then the data set is in the processing phase. At this stage, the model’s architecture must be robust to ensure accurate detection and identification of data anomalies in the following steps.

In the workflow quantization step, TinyML is leveraged to minimize the computation requirements of the system while still keeping the multi-functional system operating at peak performance. During this stage, performance criteria are guaranteed precision metrics, and classification results are fine-tuned with fuzzy logic⁵³. While integrating TinyML with fuzzy logic helps refine performance, more importantly, it guarantees dependable functionality in resource-constrained considerably lower resource contexts. Finally, the last step in the evaluation process applies a FIS to perform estimation and provide a classification for categorical anomaly detection in the optimization phase. It ensures high accuracy not only in identifying the anomalies but also in the categories of the anomalies so that they can serve practical purposes. All the operations from acquisition to optimization form a single workflow, which integrates theory and practice in an empirical approach to tackle the problem of road anomaly detection. The high-level framework of the work is depicted in Figs. 3 and 4 to encapsulate the whole strategy for detecting road anomalies. It starts from the outline and moves towards the detailed decomposition of work at the different levels of granularity. This blend exemplifies the engineering creativity of the research and the practicality of the solutions developed throughout the research. Following the application of fuzzy logic for the initial steps of classification, tensor voting further increments the granularity of detection. Tensor voting augments anomaly detection processes by enhancing the integration of peripheral local and continuous features, which are critical for overcoming occlusions and the self-similar intricate patterns that dominate remote sensing data. Combining these procedures ensures dependable detection and classification of road anomalies during extreme weather conditions and enhances the efficiency of ITS.

Acquisition and pre-processing

Obtaining and pre-processing remotely sensed images of road networks generally encompasses multiple stages. Image Acquisition involves capturing road network images using remote sensing techniques such as aerial photography or satellite imagery. The images are usually captured in different spectral bands, including visible, near-infrared, and thermal, depending on the requirements of the study. Pre-processing is typically done on the captured images to eliminate noise or artifacts that might diminish the image quality. This step involves several pre-processing techniques, such as radiometric and geometric corrections, atmospheric and terrain corrections, and image mosaicking. Several variables can impact radiometric circumstances, including variations in imaging seasons or dates, sun elevations and angles, meteorological conditions, and cloud cover, rain, or snow. Moreover, it is plausible that the precision of most change detection systems could be compromised. The pre-processing step is critical to ensure the images are high quality and suitable for further analysis. While the specific needs for picture pre-processing may differ across various change detection approaches, it is universally acknowledged that multi-temporal radiometric corrections and image enhancement are fundamental stages. Crucial radiometric and geometric corrections are meticulously applied during pre-processing to ensure data integrity. These corrected data sets then serve as the foundation for our subsequent machine-learning processing, where techniques such as tensor voting play a pivotal role in enhancing classification accuracy.

Image correction

Radiometric correction is typically deemed necessary before conducting change detection. The radiometric correction endeavors are implemented to address or remove disparities that may arise between the measurements acquired by sensors and the actual reflectivity and radiation brightness of objects’ spectra. This correction involves absolute and relative radiometric corrections⁵⁴. The main goal of absolute radiometric correction is to rectify any radiation distortions that do not pertain to the radiation properties of the surface of the object being corrected. These distortions are typically generated by sensor conditions, solar illumination, dispersion, air absorption, etc. Conventional approaches primarily involve calibrating radiation measurements to a standardized value using atmospheric radiation transmission codes, spectral curves in laboratory settings, dark object and radiation transmission codes, and scene rectification by removing dark objects, among other techniques. Implementing an absolute radiometric correction in most real cases poses a significant challenge due to the high cost and impracticality of surveying the air parameters and ground objects associated with the current data^17,50.

Furthermore, surveying previous data is nearly impossible, further complicating the implementation process. During the relative radiometric correction process, a reference image is utilized. Subsequently, a separate image’s radiation characteristics are modified to the equation with those of the preceding image. The primary techniques encompass histogram regularization-based correction and fixed object-based correction. This adjustment can mitigate or minimize the impact of atmospheric, sensor, and other noise sources. Moreover, it possesses a straightforward method. The illicit use of this phenomenon has become very prevalent. Currently, the dominant radiation algorithms employed in the pre-processing of change detection primarily include the image regression method, pseudo-invariant features, dark set and bright set normalization, no-change set radiometric normalization, histogram matching, and second simulation of the satellite signal in the solar spectrum, among other techniques²¹. It is crucial to recognize that many change detection methods do not require radiometric adjustment. There is an ongoing academic discussion over the significance of radiometric corrections in analyzing land cover change using several sensors. Nevertheless, empirical research has demonstrated that undertaking atmospheric adjustment before change detection through post-classification comparison is unnecessary if the spectral signal used for classification is derived directly from the images. Applying radiometric correction is often superfluous in change detection algorithms that depend on comparing features and objects. Image enhancement techniques are utilized to increase the visual quality of images. The methods mentioned earlier encompass contrast stretching, histogram equalization, and filtering⁵⁵.

The objective of image enhancement is to make the essential features of the road network more visible, making it easier for further analysis. Once the image is preprocessed and enhanced, object tagging is performed to identify the different objects in the image. For object tagging, the user must individually classify each object in the image, such as roads, buildings, trees, and waterbodies. The automatic tagging of objects is crucial for analyzing remotely sensed images since it enables identifying and exploring new or predefined features. Images following object tagging are then passed onto image coordinates processing to extract the necessary information. It may include object-based image analysis, classifying pixels into objects concerning their spectral, spatial, and temporal properties. Other means of processing could be machine learning algorithms like neural networks or decision trees.

Multiple road types dataset construction

The research relied on high-resolution aerial footage of Dubai, taken from satellites at the Mohammed Bin Rashid Space Centre (MBRSC). The imagery has been manually pixel-wise semantically segmented to label six classes: roads, buildings, nature (forest), vehicles, water bodies, and bare ground¹⁷. Advanced machine learning involves identification precision and the segmenting process for anatomizing urban infrastructure anomalies. This dataset is a significant part of our approach since it will help us create and approve models for recognizing anomalies based on TinyML with fuzzy logic. We used a fine rarefaction to choose road images with diverse and intricate background conditions. The collected remote sensing images were preprocessed using radiometric calibration¹¹, ortho-rectification, atmospheric correction, and mosaicking. It was followed by annotating road labels into integrated remote-sensing images to determine how well they represented their respective real-world road networks. The training and testing process uses the DeepGlobe dataset for the experiment⁵⁶.

The Dubai Path Dataset was used to evaluate both the refurbished model and its generalization methods like simulation, prototyping, and observation, to name a few, are very distinct from one another but effective for knowledge elicitation. The model’s performance was examined across various datasets. The DeepGlobe dataset is a highly regarded and extensively employed compilation of remote sensing imagery specifically curated for road extraction and network analysis research. The assemblage consists of 6226 aerial images, each with dimensions of 1024×1024 pixels and a spatial resolution of 0.5 meters. As mentioned, the images were later utilized for training and testing to evaluate the model’s effectiveness in accurately identifying and extracting road elements from real-world surroundings.

TinyML integration for processing

The proposed framework for segmentation uses Gabor filtering as pre-processing to validate the feasibility. The process is divided into TinyML implementation for macro to micro metrics evaluation in classification U-Net using TensorFlow and Keras road extraction anomalies³⁶. In the end, fuzzy logic is applied to predict the tested and validated image prediction for detecting road types and decreasing anomalies compared to existing methods. This study involved conducting comparative experiments using identical sets of training and testing samples alongside other pertinent models. The comparison experiments were performed using identical sets of training and test samples. The experimental computer operating system was Windows 11 Professional, with an Intel(R) Core(TM) i9 CPU configuration. The graphics card model utilized was the NVIDIA GeForce RTX 4060, with a video memory capacity of 16G. The CUDA version employed was 11.1, and the PyTorch version used was 1.7.0. Table 1 displays the experimental parameters.

Table 1 Parameters and values for road anomalies detection model.

This table summarizes the key parameters and their respective values for the road anomaly detection model. Important hyperparameters such as learning rate, optimizer, and loss function are listed. The approach employed in this study involves treating road extraction as a multi-class classification issue. A stochastic gradient descent (SGD) loss function was selected to optimize the model. The mathematical representation of this loss function is presented in Eq. (1), where $\alpha _{new}$ represents the parameters of the model as the actual value of the road, $\gamma$ is the learning rate, a scalar that determines the step size during the optimization process of the predicted value of the road, and $\nabla _{\alpha } l(\alpha )$ is the gradient of the loss function representing the pixel position.

$$\begin{aligned} \alpha _{new} = \alpha - \gamma \nabla _{\alpha } l(\alpha ) \end{aligned}$$

(1)

Results and analysis

It is complex to select an appropriate Knowledge Elicitation Method (KEM) for interactive sessions, including case studies, interviews, and supporting events organized in collaboration with experts, to ensure interaction-based information/knowledge integration. Additionally, the direct and indirect measures involve a scripted assessment administered by an experienced practitioner through a standard questionnaire format session. The indirect-based KEM requires a validity session to get meaningful data and knowledge. It also integrates some techniques from the KEM by leveraging established interfaces and protocols. Improved results require a conscientious choice of key exchange mechanisms. Object-based modeling is a knowledge-based expert system that recognizes and captures. Development of the KEM consists of pre-processing natural language text to a cognitive network language based on human picture interpretation, including object detection⁵⁷. Object-Oriented Modeling (OOM) is widely referred to as a KEM that can capture the cognitive activities subsumed in how humans understand interactions within real-world system contexts, which are realized through fully/semi-automated models. OOM is inspired by the human manner of detecting objects in remote sensing images. The OOM algorithm works on a classification basis to identify different objects within the image. It is designed around achieving accuracy that is as close to that of our perception. Out-of-order matching⁵⁸ is a computational process to rank images for semantic entities instead of the pixels within an image. The OOM system knows things and how these objects are related.

The segmenting input images involves utilizing a multi-resolution segmentation approach based on a region-merging and region-growing algorithm. Firstly, it is essential to note that every pixel is considered an independent entity in image processing. Secondly, these individual pixels are combined by merging, forming a larger polygonal object called a segment. In the process of multi-resolution segmentation, various factors are employed to facilitate the segmentation of the image. The parameters encompassed in this study comprise scale, form, and compactness. The idea of scale holds significant meaning as it establishes a direct correlation between the magnitude of a segment and its associated scale. The input images have been split at a scale of 30. Smoothness feeds into the two shape-heterogeneity measures ${SSH}_{\textrm{smooth}}$ and ${SSH}_{\textrm{compact}}$. which together form the overall shape heterogeneity H_shape. The combined Spectral–Shape Heterogeneity (SSH) as described in equation (2) to determine them together.

$$\begin{aligned} SSH = \bar{\omega}\cdot H_{\textrm{spectral}} + (1 - \bar{\omega}) \cdot H_{\textrm{shape}} \end{aligned}$$

(2)

where $\bar\omega\in[0,1]$ balances spectral versus shape contributions. The range of spectral heterogeneity ranges⁵⁹ between zero and one in equation (2).

$$\begin{aligned} H_{\mathrm{spectral}} = \sum_{r=0}^{B-1}\bar\omega_{r}^{\textrm{band}}\,\alpha_{_r}^{\textrm{band}} \end{aligned}$$

(3)

where,

• $\bar\omega^{\mathrm{band}}$ represents the weights of several layers

• B denotes the total number of spectral bands

• $\alpha^{\mathrm{band}}$ is the spectral bands values

Upon merging two adjacent segments illustrated in equation (3) the merged region’s spectral heterogeneity is recomputed by subtracting the original segments’ band‐wise contributions$.$.

$$\begin{aligned} H_{\mathrm{spectral}}^{\mathrm{merged}} = \sum_{r=0}^{B-1}\bar\omega_{r}^{\textrm{band}}\,\Bigl( n_{r}^{\textrm{merged}}\,\alpha_{r}^{\textrm{merged}} - n_{r}^{1}\,\alpha_{r}^{1} - n_{r}^{2}\,\alpha_{r}^{2} \Bigr). \end{aligned} $$

(4)

where, 𝑛_𝑟¹ and 𝛼_𝑟¹ denote the number of pixels and the mean value in band 𝑟 of the first pre-merge segment, respectively. 𝑛_𝑟² and 𝛼_𝑟²denote the number of pixels and the mean value in band 𝑟 of the second pre-merge segment, respectively. Then, the Shape Operator (SO) of a segment is calculated⁶⁰ as observed in Equations (5) and (6).

$$\begin{aligned} SO_{\textrm{compact}}= & \frac{p}{p_{\textrm{MBR}}} \quad \end{aligned}$$

(5)

$$\begin{aligned} SO_{\textrm{smooth}}= & \frac{p}{\sqrt{t_{\textrm{A}}}} \quad \end{aligned}$$

(6)

where,

• $p = \frac{Area}{Perimeter}$

• $p_{\textrm{MBR}} = \text {Perimeter of Minimum-Bounding Rectangle}$

• $t_{\textrm{A}} = \text {Total Area}$

$$\begin{aligned} SOH_{\textrm{smooth}}=& t_{\textrm{A}}^{\textrm{merged}} \cdot \frac{p_{\textrm{merged}}}{\sqrt{t_{\textrm{A}}^{\textrm{merged}}}} - \left( t_{\textrm{A1}} \cdot \frac{p_1}{\sqrt{t_{\textrm{A1}}}} + t_{\textrm{A2}} \cdot \frac{p_2}{\sqrt{t_{\textrm{A2}}}} \right) \end{aligned}$$

(7)

$$\begin{aligned} SOH_{\textrm{compact}}= & t_{\textrm{A}}^{\textrm{merged}} \cdot \frac{p_{\textrm{merged}}}{p_{\textrm{MBR}}^{\textrm{merged}}} - \left( t_{\textrm{A1}} \cdot \frac{p_1}{p_{\textrm{MBR1}}} + t_{\textrm{A2}} \cdot \frac{p_2}{p_{\textrm{MBR2}}} \right) \end{aligned}$$

(8)

In these formulas, 𝑛 _{𝑟 , 1} and 𝛼 _{𝑟 , 1} denote the pixel‐count and mean band- 𝑟 value of the first (pre-merge) segment, while 𝑡 _{𝐴 1} , 𝑝₁ and pM B R1 are its total area, area-to-perimeter ratio, and minimum-bounding-rectangle perimeter, respectively. Likewise, 𝑛_{𝑟 , 2} and 𝛼 _{𝑟 , 2} denote the pixel-count and mean band- 𝑟 value of the second pre-merge segment, with 𝑡_{𝐴 2} , 𝑝₂ ​ and p_MBR2 ​giving its area, area-to-perimeter ratio and MBR-perimeter. Equations (5) and (6) first introduce the two shape‐operators, which measure, respectively, a segment’s compactness and smoothness in isolation. Equations (7) and (8) then expand those operators into heterogeneity increments upon merging.  Consequently, Eqs. (6), (7) resulting in Eq. (8) give the overall shape heterogeneity calculation as depicted in Eq. (9). So that you can tune the relative weight of compactness versus smoothness in the final heterogeneity score. Equ. (9) recombines those two heterogeneity contributions into a single shape heterogeneity measures. 

$$\begin{aligned} SO_{\textrm{shape}} = \bar{\omega} _{\textrm{compact}} \cdot SO_{\textrm{compact}} + (1 - \bar{\omega} _{\textrm{compact}}) \cdot SO_{\textrm{smooth}} \end{aligned}$$

(9)

where 𝜔_compact ∈ [0,1] balances the compactness versus smoothness contribution, where the weight assigned to the compactness operator. Equation (9) recombines the two shape‐heterogeneity increments into a single overall shape‐heterogeneity measure. Each segment integrity is crucial for primary training in these several parameters in the multi-resolution segmentation²⁷. The compactness value is set as 0.5; instead, it was kept as 0.1. The nearest neighbor algorithm, as a supervised classification method, is applied to the training database for classification.

Performance evaluation and analysis

After pre-processing the remotely sensed image of a road network, the image may still contain some noise, artifacts, or other imperfections that can interfere with the detection and segmentation of the road network. Mathematical morphology is a valuable technique for further enhancing the image to improve the accuracy of the subsequent processing steps. Mathematical morphology is a set of image-processing techniques based on the mathematical theory of morphological operations. These operations involve structuring elements, which are small binary images that define the shape and size of the features to be detected or removed from the original image. The most basic morphological operations are dilation and erosion, with the possibility of merging pixels to give a complex object into an image. This part also includes reprocessing the preprocessed images to accomplish picture enhancement through mathematical morphology. The operation consists of finding a structure that can be sliced through the centerlines according to the size and shape of features in an image. For example, a line-shaped structuring element can select and combine erode dilation operations on preprocessed image road line shores, which will be thinned. This tactic can fix minor issues that don’t fit the noise bill or take care of short-circuiting road infrastructure deficits. After that, more morphological operations like dilation or erosion are applied to remove the small object and fill some holes in an image. Ultimately, they have to check whether or not improving an image brings good results using the right metrics (like accuracy, precision, and recall). After this, if any modification is required to the processing settings, it should be done accordingly.

Mathematical morphology enhanced the preprocessed images to achieve the best accuracy and precision for road network detection and segmentation. This enhancement technique could be exploited in many fields, such as traffic oversight, urban planning, and disaster response. The strategy proposed in the study uses Gabor filters to locate road features at margins during the first phase. The filter is applied to the aerial image at multiple scales and angles to obtain different road orientations and width scales. First, the filtered image is subjected to hysteresis thresholding, which gives us a preliminary road map. The suggested methodology integrates techniques such as Gabor filtering, hysteresis thresholding, road filtering based on form characteristics, and tensor voting. These techniques extract a refined and precise road centerline from high-resolution aerial images. The experimental findings provide evidence of the proposed approach’s efficacy compared to existing methodologies.

Gabor filtering

A proficient technique for detecting edges is necessary. The Gabor filter is often regarded as a highly suitable approach, owing to its resemblance to the human visual brain. This filter is advantageous in extracting road parameters from images of Very High Resolution (VHR), even in the presence of spectral diversity. Road edge characteristics exhibit variations in both scale and direction. Selecting road features responsive to the Gabor filter involves conducting tests at various scales while considering the significance of direction in analyzing edge features. The ensuing illustration demonstrates in Eq. (10) the utilization of the Gabor filter function for the identification of edge features in the spatial domain:

$$\begin{aligned} G(\alpha , \beta , \theta , \varphi ) = \exp \left( - \frac{x'^2 + \lambda y'^2}{2\beta ^2} \right) \cos \left( 2\pi \frac{x'}{\alpha } + \varphi \right) \end{aligned}$$

(10)

Where the parameters are defined as follows:

• $\lambda = 0.5$ constant aspect ratio determining ellipticity

• $\beta$ controls the size of the Gabor envelope

• $\alpha$ is the wavelength of the Gabor function

• $\theta \in [0, \pi ]$ determines the orientation angle

• $\varphi \in [-\pi , \pi ]$ is an offset parameter controlling symmetry

The rotated coordinates $(x', y')$ are given by:

$$\begin{aligned} x'&= x \cos \theta + y \sin \theta \end{aligned}$$

(11)

$$\begin{aligned} y'&= -x \sin \theta + y \cos \theta \end{aligned}$$

(12)

To detect road features from an input image, consider input-output mappings:

$$I_i(x,y) \rightarrow \text {Output Image}, \quad O_i(x,y) \rightarrow \varphi , \quad \text {where}~\varphi \in [0, \frac{\pi }{2}]$$

The purpose of this technique is to detect edges with varying orientations. Orientation angles $\theta _i$ are systematically adjusted and computed as:

$$\begin{aligned} \theta _i = \frac{i\pi }{N}, \quad i = 0, 1, 2, \ldots , N - 1 \end{aligned}$$

(13)

This investigation has demonstrated the adequacy of the eight distinct orientations for extracting road features illustrated in Eq. (13), as depicted in Fig. 5.

Fig. 5

Gabor-filtering on urban and rural images with the variation of kernels.

Fig. 5 visually exemplifies the impact of applying Gabor filters with different orientations and scales on urban and rural images. This visualization demonstrates how the filters enhance the detection of road edges, aligning with our theoretical discussion on the flexibility and effectiveness of Gabor filters in handling various road scenarios. During each orientation session, g(x, y), the creation of the image occurs at various values; subsequently, the input image undergoes convolution g(x, y). The output image $O_{i}(x,y)$ generation is achieved by summing the outputs of all orientations and using the L2-norm. Following applying Gabor filtering to VHR images for road edge detection, the subsequent step involves classifying road and non-road components. The utilization of hysteresis thresholding efficiently achieves this classification task. Using upper and lower thresholding values is employed $Th_{high}$. The pixels having responses between $Th_{low}$ and road parts are regarded as remaining components, whereas the non-road component is deleted. The assigned values are established in the present study as 0.2 and 0.5. It displays the primary road obtained through implementing the proposed methodology. The road network produced in the first stage of the proposed research has non-road components that deviate from the geometric characteristics of the road network. Moreover, the preliminary road map exhibits deficiencies in gaps and a disrupted road network resulting from the obstruction caused by the shadows cast by trees. It is imperative to address the tasks like filling holes, removing non-road elements, and reconstructing damaged sections of the road to improve the quality of road extraction. This task is accomplished by implementing filtering techniques that rely on morphological operations and form features. The closure operation fills voids in morphological filtering, while the opening calculation eliminates certain non-road elements. It is achieved by employing a structuring element with a dimension ranging from 2 to 5 pixels. The filtration process for shape characteristics depends on utilizing Geometric Feature Key (GFK) values across several parameters established using Connected Component Analysis (CCA). Each component can be described as a bounding rectangle⁶¹. The length of the bounding rectangle is denoted as L, the width as B, and the area of the connected component as A_c . The diverse values of GFK can be computed utilizing the subsequent formulae in Eqs. (14), (15) and (16):

$$\begin{aligned} GKF_1= & \frac{A_c}{PM^2} \end{aligned}$$

(14)

$$\begin{aligned} GKF_2= & \frac{L}{B} \end{aligned}$$

(15)

$$\begin{aligned} GKF_3= & \frac{A_r}{A_c} \end{aligned}$$

(16)

where, $PM ={Perimeter}$, $A_{r}= {Area}$. In these formulas, PM denotes the perimeter of the minimum‐enclosing rectangle; L and B are that rectangle’s length and width, respectively; Ar is the pixel‐area of the connected component; and Ac is the area of its convex hull.The smallest enclosing rectangle encompasses both the perimeter and the area of the object. Based on the geometric characteristics of the road network, any entity that meets the criteria outlined in Eq. (17) is classified as a valid road component. Conversely, an entity failing to meet these criteria is deemed invalid and rejected. It also displays the road network derived through shape feature filtering techniques on VHR aerial images⁶².

$$\begin{aligned} GKF_1 > 50, \quad GKF_2< 4, \quad GKF_3 < 1 \end{aligned}$$

(17)

Extracting road centerlines from images with VHR has been a highly active field of academic investigation. This work uses an Euclidean distance transformation approach to acquire road centerlines. Additionally, mathematical morphology approaches are employed to minimize undesirable minor branches. This approach necessitates a reduced amount of computational resources and time compared to alternative methodologies. The method successfully extracts a centerline that exhibits accuracy, one-pixel thickness, and a smooth appearance.

Model quantization and optimization

Tensor voting

Road reconstruction involves the restoration and improvement of existing road infrastructure. The centerline acquired during the process has multiple fractured portions due to occlusions caused by shadows, foliage, and vehicles. The reconstruction of the broken road centerline is necessary, and tensor voting technology efficiently accomplishes this objective by improving the precision of the demarcated road centerline. As a medium representing images’ saliency structure, tensor voting utilizes a voting mechanism to gather information from the surrounding environment. At the same time, the computation of the vector field is employed to determine the orientation information. The initial centerline, denoted as CL, is represented by the tensor field T_f, which can be decomposed in the following manner in Eq. (18):

$$\begin{aligned} T_f = (\lambda _1 - \lambda _2) e_1 e_2^{T_f} + \lambda _2 (e_1 e_1^{T_f} + e_2 e_2^{T_f}) \end{aligned}$$

(18)

where, $\lambda _1$ and $\lambda _2$ are eigenvalues $e_1$ and $e_2$ are eigenvectors, $e_1 e_1^{T_f}$ is the stick tensor, ${\lambda_1 - \lambda_2}$ represents the saliency of the stick tensor, $e_1 e_1^{T_f} + e_2 e_2^{T_f}$ is the ball tensor, with saliency $\lambda _2$. In the above, λ₁ ≥ λ₂ are the eigenvalues of the local structure tensor, and e₁, e₂ are the corresponding orthonormal eigenvectors. Once the process of encoding voting is completed, tensors transmit their information to adjacent nodes. At the same time, the decay function determines the degree of influence exerted by the vote in Eqs. (19) and (20). Here, P is the vote’s arclength (distance from the voting point), K is the local curvature, σ is the voting scale, c is the weighting constant.

$$\begin{aligned} DF(P,K) = e^{- \frac{(P^2 + cK^2)}{\sigma ^2}} \end{aligned}$$

(19)

where, $P = {arc length}$, $K = {curvature}$, $\sigma = {scale\ of\ voting}$ ( $\infty$ size of broken centerline in this study) and c is a constant related to $\sigma$ and can be calculated as $c=-16ln(0.1)(\sigma -1/2)$. After voting on each point, $C_{L}$ it $T_{f}$ generates a new tensor field $T_{fs}$, and it can decompose as follows:

$$\begin{aligned} T_{fs} = (\lambda _{s1} - \lambda _{s2}) e_{s1} e_{s2}^{T_f} + \lambda _2 (e_{s1} e_{s1}^{T_f} + e_{s2} e_{s2}^{T_f}) \end{aligned}$$

(20)

The decomposition $T_{fs}$ generates a stick tensor saliency map $\lambda _{s1}-\lambda _{s2}$. The utilization of centerline line reconstruction as a Boundary Field (BF) with a pre-established constant value can be implemented. It suggests that the tensor field can be considered at a tiny scale and neglected for values beyond the BF. The present study presents the findings obtained through experimentation and subsequent analysis, followed by a comprehensive explanation. The section presents a series of experiments using various photographs to evaluate the proposed method’s performance in delineating road centerlines. In addition, the outcomes of the proposed methodology equation with those of established techniques for road centerline extraction, as seen by their agreement on multiple quantitative metrics. The road network was derived by implementing the suggested methodology on various VHR aerial photographs, as depicted in Fig. 6.

Fig. 6

Rural and urban anomaly in road network detection.

The multi-layered filtering technique effectively removes noise and deceptive road components. However, tree shadows pose challenges during road extraction, leading to a fragmented road network, as illustrated in the Fig. 6. Eliminating variable width from the road network, as determined using VHR images, is crucial. The task can be effectively tackled by extracting road centerlines. The centerlines acquired in this study exhibit completeness, smoothness, and a thickness of one pixel, as visually depicted. This particular attribute remains consistent even in road sections that exhibit curvature. The proposed methodology in this research paper results in a centerline that displays discontinuity caused by occlusion resulting from many factors, as demonstrated by the shaded rectangular area in extracted figures. The method has proven effective in repairing ruptures in road centerlines. The tensor voting technique effectively mitigates the presence of extraneous spurs in the extracted road centerline, resulting in a more streamlined and refined road network.

The efficacy of the suggested methodology is additionally confirmed through the analysis of images acquired from the ultimate road plan derived from the VHR image. Nevertheless, the impact of this phenomenon is observable as a result of blockage induced by several contributing elements. Huang’s approach relies on morphological thinning and CCA techniques, producing spurs that contribute to smoothness reduction [6]. Simultaneously, Miao’s approach relies on regression analysis to produce smooth road centerlines [27]. However, it fails to accurately capture the connectivity of the road network at intersections, resulting in fragmented representations. The extraction of the road network from the VHR aerial image is achieved by a multi-layered filtering approach that incorporates the utilization of the Gabor filter and form feature filtering techniques. The Euclidean distance transformation technique acquires the road centerlines, followed by tensor voting, to rectify any fractures in the road network. The challenges about spurs, completeness, and accuracy are significant considerations in road extraction methodologies. Nevertheless, the current method is unsuitable for photos with poor spatial resolution, and there is a need for enhancement in accurately determining the centerline at junctions.

Training and validation

In TinyML and road network detection, the training and validation phases using a neural network are crucial for achieving an optimized process in Tensorflow. During the training phase, the neural network is presented with a set of labeled images of road networks, and it learns to extract relevant features from the images and associate them with the correct label. It is done by adjusting the weights and biases of the network using backpropagation and gradient descent algorithms. The aim is to minimize the difference of 426 from actual labels and predicted, as shown in Fig. 7.

Fig. 7

Dataset-based random image training and validation.

Then, the test images are passed to the trained neural network to its performance in the validation phase. It would help protect against the neural network overfitting to just this data and remembering it. The following metrics have been considered in computing the performance: accuracy, precision, recall, and F1-Score. There are many ways to optimize the TensorFlow process - such as data augmentation, transfer learning, and tuning a few hyperparameters. Data augmentation—Creating a new training dataset through random changes on images. Thus, the transfer learning paradigm is chosen, wherein a pre-trained neural network model acts as the base that will be retrained to detect road networks. Hyperparameter tuning—Choosing the “right” values of different hyper-parameters, such as learning rate, batch size, etc, for neural network. These means can likely accelerate the recognition method of road networks and function successfully based on visualizing the classification results. In particular, we also used the task-specific loss and Intersection over Union (IoU) metric shown in Fig. 8 to validate how good our classification model is on images of datasets.

Fig. 8

Training vs validation IoU and loss of random imagery from the dataset.

Similarly to TinyML and road network detection, the training and validation following a data set of neural networking architectures can also prove vital in ensuring accuracy. During training, we’ll feed the neural network a large dataset of labeled images of road networks (the labels, in this case, would refer to whether roads are present in an image or not). What is observed in this type of network is that the network automatically starts learning to find features from images (which would help it identify if a place represents a road or not) as we keep feeding data (images). Usually, this is accomplished by adjusting the weights of the network’s connections through a process called backpropagation, where we propagate error backward towards minimizing some loss function, which indicates how far our predictions are from the accurate labels. It is necessary to validate this, which means it is on a separate set of images not used in training to confirm that the trained neural network works well. Given that the network has learned to predict road locations, it would typically make sense to evaluate images of a new location, and this is an important metric when considering the generalization power of a model as its ability to detect roads in images from places not seen before accurately. All these metrics, such as accuracy, precision, recall, F1-Score, etc., are calculated in this process. These are meaningful numbers because they give an idea of what the network is good and bad at doing.

The limited processing power available on tiny devices can be problematic for the training and validation of a new model. For example, training a large-scale neural network on a smartphone or microcontroller can be time-consuming and memory-demanding. In this way, the problem can be addressed using different methodologies like quantization, pruning, and distillation, which help reduce the cost & complexity of the network while keeping up reasonable precision. Additionally, using transfer learning features makes use of other pre-trained models on more significant devices. The developed models are adapted and further fine-tuned for road network detection. Remote sensing technology and small machine learning systems can now capture dense point clouds to extract more detailed information about road surfaces. It involves finding cracks, road markings, and other features. A greater level of detail makes it easier to find and analyze imperfections in a road surface. New technologies can harvest data when vehicles are driving up to highway speed, reducing the need for manual inspections and alleviating traffic disruptions. It improves operational efficiency and cuts back data capture and road maintenance costs. Table 2 displays the distribution of the Dubai dataset and DeepGlobe imagery according to the type of roads used to detect anomalies.

Table 2 Descriptive statistics of road network analysis.

Remote data collection enables staff to remain off the road, ultimately decreasing the risks of accidents and injuries. It is beneficial for data collection in hazardous or high-traffic areas. It can also be used for various applications without the need to go back into the field. For example, it makes the data more useful and effective by enabling its use in road-related scenarios other than crack extraction only. Models built through remote sensing and TinyML can be the key to better road maintenance management decisions by providing intricate, granular, accurate data. This feature assists governments in focusing on the need to maintain their existing investment and use money, and they can plan infrastructure upgrades. Fig. 9 shows the confusion matrix, which tells us everything about TinyML having or no detections in road network irregularity.

Fig. 9

Confusion of random imagery from the dataset.

Fig. 9 Confusion matrix illustrating model performance across anomaly classes. High misclassification rates explicitly highlighted between visually ambiguous classes, guiding future refinement areas. Analyzing the confusion matrix revealed specific classes where misclassifications were notably higher. In particular, the model faced challenges distinguishing between classes characterized by visually ambiguous features or overlapping textures, such as cracks versus road discoloration and minor surface wear versus intact road surfaces. These ambiguities primarily arose due to insufficiently distinct training examples or inherent similarities in remote-sensing images. The misclassifications may also stem from the model’s constrained representational capacity due to its optimized size for edge deployment, potentially limiting its ability to learn subtle feature differences thoroughly. Future research should address these challenges by incorporating more balanced and comprehensive datasets with more apparent distinctions among visually similar classes. Additionally, applying enhanced data augmentation techniques or integrating attention mechanisms could further improve the model’s ability to differentiate among closely related anomaly classes. Recent studies explicitly highlight the effectiveness of fuzzy logic and hybrid machine learning approaches in addressing misclassification issues arising from visually ambiguous anomalies in remote sensing images, providing valuable directions for future model refinements⁴³. These insights explicitly guide ongoing improvements and inform future datasets and training strategies.

Fuzzy logic and remote sensing data have been used in this evaluation. It demonstrates how original labels, measuring road issues such as intricacies, similarities, ambiguity, cracky situations, and wearing down, can be related to TinyML model-generated labels from the remote sensing data analysis. The methodology is based on applying fuzzy logic to the uncertainties and incompleteness associated with remote sensing data. These discrepancies can be potentially due to lighting, weather, and sensor readings. The proposed method surpasses binary classifications of images. It offers a more advanced way to evaluate complex and vague anomalous pathological appearances in the image. Most examinations in model efficacy are done using the matrix. Cells with diagonal orientation, dark shading, and high numerical values (i.e., 97 for category 6 or even higher, like in the case of class 5) mean that they perform well when detecting certain road conditions or anomalies. Nonetheless, off-diagonal cells are associated with non-zero values that indicate misclassifications. They were particularly problematic for fuzzy logic rules or the classifier in general to distinguish cases. The matrix also reveals which model sees anomalies differently. This observation is critical for identifying the model’s strengths and shortcomings, especially since road safety and maintenance require exact anomaly diagnosis. A confusion matrix is essential for assessing a model’s deployment in TinyML, where resources are scarce, and accuracy and computational efficiency are vital given the challenges of updating models on edge devices, assessing if the model needs more optimization before deployment is crucial. The study’s confusion matrix shows the model’s ability to classify road anomalies. This matrix also reveals fuzzy logic’s practicality. The above discoveries contribute to the continued development and refining of the TinyML model and its fuzzy logic parts, laying the groundwork for the effective deployment of this technology in actual applications.

Evaluation indicators of the proposed framework with other models

Quantization reduces model dimensions and neural network computational complexity while retaining accuracy. Deploying the model on low-power devices like microcontrollers with limited memory and processing is crucial. Weight quantization encodes neural network weights with fewer bits and is widely used. Instead of 32-bit floating-point data, 8-bit integers could represent weights. The decrease in model size and memory may be significant. Quantization includes activation quantization, which encodes neural network activations with fewer bits. Less complex procedures implemented on low-power devices may reduce neural network computational complexity. After quantization, MSE, IoU, and F1-Score are used to evaluate model accuracy. These metrics assess the model’s performance on test data and its accuracy for the intended application.

$$\begin{aligned} \text{Precision} &= \frac{TP}{TP + FP},\\ \text{Recall} &= \frac{TP}{TP + FN},\\ \text{F1-Score} &= \frac{2\,\text{Precision}\times\text{Recall}} {\text{Precision} + \text{Recall}},\\ \text{F1-Score}_{\mathrm{macro}} &= \frac{1}{n}\sum_{i=1}^{n}\text{F1-Score}_{i},\\ \text{F1-Score}_{\mathrm{micro}} &= \frac{2\,\text{Precision}\times\text{Recall}} {\text{Precision} + \text{Recall}},\\ \text{Accuracy} &= \frac{TP + TN}{TP + FP + FN + TN}\,.\end{aligned}$$

Quantization is essential for low-power machine learning model deployment. This phase is critical for reducing the model’s memory and computing needs while maintaining accuracy. Quantization approaches may help optimize the model for road network recognition from remotely sensed images, making it suitable for drones or other low-power remote sensing devices. TinyML has improved remote-sensed picture road network detection. For general and specialized analysis, pixel-based classification and object-based image analysis are used in road detection and network mapping. CNNs, especially tiny ones for edge devices, have enabled real-time road detection on devices with low computational resources. However, Lidar-based systems create precise 3D road maps. The Hough transform is essential for recognizing linear road segments, and fuzzy logic-based approaches excel in uncertain road detection circumstances. MobileNet, EfficientNet, and YOLO, designed for edge devices, have improved the TinyML ecosystem. Edge AI accelerators and TinyML libraries expedite model deployment on edge devices, improving performance. Transfer learning, quantization, pruning, support vector machine, and decision tree work well on resource-constrained devices. Remote servers and hybrid models in edge processing strive to balance edge and cloud processing, optimizing road detection.

Fuzzy logic increases road feature extraction using fuzzy C-means clustering and fuzzy morphological operations. Ambiguity-tolerant road detection uses fuzzy rule-based algorithms, texture analysis, and edge detection. Fuzzy inference algorithms and segmentation improve road classification and decision-making. Integrating fuzzy logic, neural networks, Markov random fields, and evolutionary algorithms optimize road detection parameters. Real-time road network detection using fuzzy decision trees and hybrid models with edge and cloud processing is complete and uncertainty-aware. With several benefits, TinyML can extract road network abnormalities from remotely sensed images. It allows instantaneous processing and analysis of large amounts of visual data, enabling road network problem identification and resolution. Road damage or other faults can endanger drivers or users, making them worth fixing. Deep learning methods like the fuzzy inference system can improve roads’ network anomaly detection accuracy. These algorithms can learn from large datasets and uncover subtle anomalies humans may unrecognize. TinyML’s extraction of road network abnormalities could improve transportation network safety and efficiency while reducing manual inspection and maintenance costs. This theoretical study uses Table 3 to evaluate remote sensing’s ability to detect road network anomalies using multiple measures:

Table 3 Comparison of different methods.

The Table 3 analyzes the various machine learning methods that can be utilized to identify anomalies in road networks using data obtained from remote sensing. To overcome the inherent uncertainties that are linked with this kind of data, it is likely that fuzzy logic is used. The information in this table has been arranged to illustrate the relative efficacy of these strategies relative to several different performance indicators. The only accuracy measure provided is TinyML-U-Net-FL, and it has been shown to have an accuracy of 73.91%. Consequently, it is impossible to compare it directly with any other methodologies. Despite this, the approach has a recall rate of 92.4%, which, when compared to the other methods listed, indicates that it is exceptionally effective in precisely finding true anomalies. Detecting anomalies is essential in practical contexts since failing to identify such abnormalities could result in significant repercussions, such as accidents caused by unreported road damage. The failure to identify such irregularities could also fail to detect anomalies. The suggested method now has a precision of 78.2%, placing it fairly competitive when measured against other approaches. However, it is not the highest, which suggests that it does not incorrectly label normal segments as abnormal nearly as frequently as other approaches do. The suggested method obtains the highest possible F1-Score of 84.7%, demonstrating its capacity to reach a greater balance between precision and recall. It suggests that the approach successfully recognizes many genuine abnormalities (recall) while ensuring these identifications’ accuracy (precision). When one considers the various ways in which one’s precision and memory can be compromised, the significance of this statistic becomes more apparent. The IoU measure is not a part of the proposed method; however, it is frequently used to quantify the degree to which the expected and actual anomaly zones overlap.

A higher IoU score indicates that the approach correctly detects anomalies and demonstrates precision in precisely localizing their places. It can be shown by the fact that the method accurately localizes their positions. We cannot evaluate the efficacy of the suggested methodology in terms of its geographical precision for anomaly detection because it does not include an IoU metric. The methods proposed in this study, the TinyML-U-Net-FL, capture a successful representation of road network anomaly detection by integrating remote sensing & fuzzy logic. We demonstrate this by the method presented here. A particularly remarkable feature of it is its precision in recognizing abnormalities. Also, the F1-Score is a good overall measure of performance. It is observed that the high recall comes in handy where missing an abnormality could mean great harm, which some argue is better to err on the side of caution. That suggests it might be a valuable tool for supporting road safety, as the connotation implies. With this evident absence of IoU data, one can only infer that it would be challenging to assess the same suggested method in accurate localization comprehensively, hence a window for further development and research. Recent research explicitly integrating Type-2 fuzzy logic controllers with neural networks and digital twin technology presents valuable insights into advanced control and detection systems applicable explicitly to future anomaly detection frameworks⁴⁰. Additionally, recent developments in hyperspectral aerial remote sensing combined with embedded machine learning solutions suggest promising approaches for enhancing data acquisition and anomaly detection robustness explicitly for road and environmental monitoring applications⁶⁶. The practical implications of our integrated TinyML and fuzzy logic framework are substantial, explicitly enabling real-time road anomaly detection critical for smart cities and automated driving systems. Our approach explicitly facilitates proactive road maintenance, significantly improving public safety infrastructure longevity and reducing municipal maintenance costs. Additionally, the proposed framework is particularly beneficial explicitly for resource-constrained environments, such as rural or developing areas, where conventional high-computational solutions are infeasible.

Conclusion

This study showcases significant advancement in road network anomaly identification by establishing a novel methodology that integrates TinyML, remote sensing imaging, and fuzzy logic inference techniques. Some research emphasizes proficiency in detecting road network anomalies by adopting TinyML-U-Net-FL. This method can achieve a recall of 92.4%, a precision of 78.2%, and provides an F1-Score of about 84.7%. Most importantly, it significantly outperforms modern equivalents in terms of effectiveness. The novel integration presented by TinyML-U-Net-FL explicitly addresses existing gaps by substantially reducing computational complexity and enabling efficient real-time anomaly detection on resource-constrained edge devices. The type of unpredictable and varied data inputs that can only be effectively managed using different techniques has been helpful (in terms of energy efficiency and carrying capacity) for smart city infrastructure and autonomous navigation. The potency of our approach focuses on U-Architecture neural networks used to build smaller and faster models implemented on microcontrollers, which are made robust by model compression, quantization, and pruning techniques. These presentations target methods that maximize the advantages of FIS specifically crafted for TinyML devices. This integration improves accuracy and speed compared to standalone approaches, effectively handling ambiguity inherent in remote sensing data (aesthetic imperfection and incompleteness). It represents a significant step forward. To further prove the effectiveness of our provided aid, a confusion matrix is utilized to illustrate how well the model accurately groups certain deviation types as road condition anomalies. Analysis of misclassification cases also highlighted opportunities to improve anomaly detection accuracy, particularly by addressing challenges posed by visually ambiguous road conditions through refined training approaches. However, limitations remain, such as computational resources or the availability of training data. Although TinyML frameworks are oriented towards power consumption efficiency, processing capacity still presents challenges, suggesting scope for further optimization.

Recent advances in model optimization techniques such as binary neural networks, fuzzy logic-enhanced control systems, and advanced pruning strategies provide valuable insights and practical methodologies for enhancing the future capabilities of TinyML-based anomaly detection frameworks. One should explore innovative methods to effectively address data complexities and improve accuracy in road anomaly detection for safety applications. Furthermore, a more comprehensive exploration of unsupervised and supervised learning scenarios in TinyML contexts could uncover novel research opportunities. Enhancing the model’s capacity for generalization and precision requires augmenting the dataset and optimizing the training procedure to encompass broader road conditions and irregularities. Engaging in collaborative efforts with urban planners and traffic management authorities can yield significant insights and empirical data, enhancing model applicability across diverse urban and rural settings. In summary, our study substantially contributes to cognitive infrastructure management and the development of automated navigation systems, highlighting the promising potential of integrating TinyML and fuzzy logic to advance ITS significantly. The practical impacts of this research extend significantly beyond theoretical advancements, clearly offering tangible benefits for road infrastructure management, traffic safety improvements, and enhanced responsiveness in ITS across diverse urban and rural contexts.