Introduction

Agriculture remains the backbone of many economies, playing a role in ensuring food security and promoting sustainable development. Among various staple crops, potatoes rank as the fourth most important food crop globally, following maize, rice, and wheat. Potatoes are cultivated in over 100 countries and consumed by more than a billion people worldwide1. However, their cultivation is frequently threatened by several biotic stress factors, especially leaf diseases such as early blight, late blight, bacterial wilt, and leaf spot diseases. These diseases not only deteriorate the quality of the produce but also result in substantial yield losses, impacting both farmers and supply chains.

Conventional disease screening is characterized by hand check-ups conducted by farmers or agricultural professionals, which are not only time-consuming but also prone to a considerable number of mistakes due to environmental changes, human exhaustion, and a lack of proficiency in far-flung areas. Additionally, such manual systems are not scalable; therefore, they cannot be applied in large-scale agricultural monitoring. The interest in implementing technology-based solutions has been growing due to the need to automate disease detection processes, particularly through the use of artificial intelligence (AI) and deep learning, which can achieve this with greater precision and consistency.

Deep learning, a type of machine learning, has transformed the computer vision domain, as it enables systems to learn hierarchically directly from feature representations on raw image data. CNNs have demonstrated impressive performance in image recognition and object recognition applications, making them suitable for plant disease identification based on leaf image recognition. In particular, MobileNetV3 is a lightweight, high-performance CNN architecture explicitly built to support mobile and embedded vision tasks. This provides a good trade-off between accuracy and computational efficiency, making it suitable for real-time agricultural deployments in resource-constrained areas, such as remote farms or edge devices.

High-performance notwithstanding, deep learning models typically require years of computational resources and vast amounts of labeled data to train. With precision agriculture generating vast amounts of sensor and image data, it becomes increasingly challenging to make those data sets manageable and processable. To resolve the scalability issue, the proposed research incorporates the MapReduce paradigm, a distributed computing model pioneered by Google, to facilitate parallel and fault-tolerant workload execution on distributed computing nodes, supporting large-scale dataset analysis.

The MapReduce mechanism works by breaking down advanced data processing into two major phases: the Map stage and the Reduce stage. Filtering and sorting tasks are carried out on the Map stage, while the summarization job requires the aggregation of intermediate results on the Reduce stage. This parallel processing capacity has led MapReduce to remain very well-suited for running big data in real-time applications. In collaboration with MobileNetV3, MapReduce can support the distribution of training and inference tasks, leading to reduced processing latency and a more responsive disease detection system.

In this study, we present a deep learning-based MapReduce framework for detecting and classifying potato leaf diseases. The system will be trained with a significant number of labeled potato leaf images, including those of different diseases and healthy leaves. The deep learning model, which is composed of MobileNetV3, is trained with the help of specifically managed distributed computing resources using the MapReduce framework. Such an arrangement not only expedites the model training procedure but also ensures that the system can scale to meet the needs of larger datasets in the future.

The integration of MobileNetV3 and MapReduce presents a novel approach that leverages the strengths of lightweight deep learning and distributed computing. Our framework is designed to be scalable, cost-effective, and suitable for real-time deployment in field conditions. Experimental evaluations demonstrate that the proposed system achieves high accuracy in disease classification, with reduced training time and improved performance on large datasets. This work contributes to the growing field of smart agriculture by providing a practical and scalable solution for disease detection, which is essential for timely intervention and effective crop management.

Literature review

The integration of advanced technologies in agriculture, especially for disease detection in potato crops, has become a prominent area of research. This literature review focuses on the critical components relevant to the proposed study, the use of the MapReduce framework for large-scale data processing, and the application of deep learning for transfer learning. It explores recent studies, processes, and findings that form the foundation of the proposed approach for effective potato leaf disease detection.

Potato leaf diseases, such as early and late blight, cause significant crop losses globally, and traditional manual diagnosis is slow, error-prone, and unsuitable for large-scale farming operations. Iqbal and Talukder developed a Random Forest classifier using PlantVillage images, achieving 97% accuracy in distinguishing diseased from healthy leaves. However, their approach relied heavily on handcrafted features derived from preprocessed images, limiting its robustness and scalability in varied real-world conditions due to its dependence on manually tuned filters and thresholds2.

These deep learning-based models can be more computationally expensive, but can be more effective at feature abstraction, making it impractical to run them on mobile or edge devices constrained to the field. Arshad et al. proposed PLDPNet, which merges the features of both VGG19 and Inception-V3, utilizing vision transformers to achieve an accuracy of 98.66% and an F1-score of 96.33%. Yet the model’s enormous architecture demands extensive GPU resources and suffers from latency issues, making it less suitable for real-time inference on resource-limited platforms3. Lightweight CNNs tailored for edge deployment have gained attention, but they often compromise accuracy. An MDPI study proposed RegNetY-400MF (0.4 GFLOPs, ~ 16.8 MB) with transfer learning for classifying seven potato leaf diseases. The system reported 90.68% accuracy, but the relatively low accuracy compared to heavier models suggests that lightweight networks may struggle to detect subtler disease symptoms and lack generalizability under variable lighting and background noise4.

DenseNet architectures have been explored to strike a balance between depth and efficiency. A study optimizing potato leaf disease detection using DenseNet-121 achieved 98.37% accuracy, illustrating strong representation power despite a moderate model size. However, the research did not address deployment challenges or tractability in distributed systems, and the lack of real-time performance benchmarks leaves its practical adoption uncertain5. Focusing on custom-designed CNNs, a study proposed a lightweight five-layer architecture compared against a four-layer CNN and the original MobileNet. The custom model achieved 97.16% accuracy, outperforming both alternatives (78–73%). Notwithstanding, this model was evaluated on a limited dataset and lacked stress-testing against unseen disease variations or conditions. Scalability issues remain unexplored when transitioning to larger datasets/ MapReduce-style distributed frameworks6.

Potato plant disease detection remains impeded by the variability of field conditions, including lighting and background clutter. Sinamenye et al.7 proposed a hybrid EfficientNetV2B3 + ViT model trained on a realistic “Potato Leaf Disease Dataset,” achieving 85.06% accuracy—a notable 11.43% improvement over traditional CNNs. However, the model’s relatively modest accuracy highlights the difficulty in distinguishing subtle disease symptoms in diverse environments7. Lightweight networks are appealing for edge deployment, yet often sacrifice precision. Authors introduced RegNetY-400MF (0.4 GFLOPs, 16.8 MB), which scored 90.68% accuracy across seven disease classes. The computational savings are evident, but accuracy lags behind that of heavier architectures, limiting their applicability for critical fields4.

Stacked ensembles have been tested to boost performance. Jha et al.8‚9 combined ResNet, MobileNet, and Inception via transfer-learning stacking, reporting 98.86% accuracy. This hybrid strategy improved robustness, but the ensemble’s complexity and dependency on large datasets challenge real-time edge deployment8. These studies proposed to enhance the plant disease classification using bidirectional long short-term memory9,31. Disease severity estimation remains underexplored. A recent study applied CNNs for the classification and severity assessment of potato leaves, achieving 93% accuracy, but relied on manual labeling and a small dataset, thus limiting the robustness of the severity model10. An improved YOLOv5 with ShuffleNetV2 addresses detection in complex field scenarios and coordinates attention, halving the number of parameters (from 7.02 to 3.87 M) while improving average precision by 3.22% and speeding up detection by 16%. However, manual annotation and reliance on YOLO limit scalability and generalisation11.

Taylor & Francis Online published an optimised potato leaf disease model with just 204 K parameters, yielding 98.6% accuracy. Though efficient, the model’s reduced complexity may underperform on novel disease variants or mixed-field conditions12. CBSNet-based contrastive features achieve 92.04% accuracy, making them adept at detecting subtle and blurry disease regions. However, its reliance on high-quality imagery constrains its performance, as lower-resolution field captures are not supported13.

Reis and Turk (PMC, 2024) introduced MDSCIRNet, which integrates DenseNet and SVM, yielding an accuracy of 99.33%. Despite strong performance, the ensemble complexity and need for SVM tuning impede deployment scalability14. A comprehensive study comparing ResNet50 and other models on potato leaf tasks found that it achieved an accuracy of 97%. However, all models underperformed on healthy-leaf cases, highlighting challenges in non-diseased classification15.

Tambe et al. presented a basic CNN that achieved 99.1% accuracy across early blight, late blight, and healthy classes. Though impressive, the research lacks evaluation on diverse environmental factors and MapReduce scalability16. Finally, a GAT-GCN hybrid model with superpixel segmentation achieved a 0.974 F1-score for potato leaf detection, showcasing its adaptability to structured features. However, graph-based models struggle with large datasets and require significant computation, hindering their integration with distributed models like MapReduce17.

Deep learning is a powerful tool for crop disease detection, particularly in potato farming, where leaf diseases significantly reduce yield. Traditional inspection methods are inefficient and prone to error. Recent studies have applied CNNs to detect diseases with high accuracy, yet challenges remain in scalability, real-world deployment, and handling large datasets. This review examines existing approaches and highlights the need for distributed solutions, such as MapReduce.

Methodology

The proposed framework for potato leaf disease detection integrates deep learning with distributed data processing to enable accurate, scalable, and efficient classification of leaf conditions. The methodology comprises several key stages: data acquisition, image preprocessing, deep learning-based feature extraction using MobileNetV3 during the mapping stage, and distributed classification utilising MapReduce’s shuffling and reducing operations.

The image dataset used for this study consists of 2152 labelled images categorised into three distinct classes: early blight, late blight, and healthy. These images are collected from a publicly available dataset. All collected images are stored in a centralised database for preprocessing and training. For the present dataset (2152 images), model training is performed on a single GPU. We introduce MapReduce not to accelerate training at this scale, but to support operational scalability of the data pipeline—namely, image ingest, preprocessing (resize/normalize/augment), and distributed batch inference across commodity nodes when image volume spikes (e.g., multi-farm, multi-season ingestion). In our pipeline, mappers perform deterministic preprocessing and forward passes to produce logits; reducers aggregate results and write per-batch outputs. This design preserves model parity while improving throughput as data volume grows.

In the preprocessing stage, raw images undergo various transformations to enhance feature clarity and ensure uniformity. Data augmentation strategies, such as rotation, flipping, brightness variation, and zoom, are applied to improve model generalization and robustness. Each image is resized to 224 × 224 × 3, the input dimensions expected by MobileNetV3. The preprocessed dataset is then split into training (80%), validation (10%), and testing (10%) sets as shown in Fig. 1.

Fig. 1
Fig. 1
Full size image

Workflow of the proposed MapReduce-based MobileNetV3 framework for potato leaf disease detection.

Once preprocessing is completed, the system proceeds to the MapReduce-based distributed learning pipeline. During the mapping phase of the MapReduce framework, the number of mappers is dynamically assigned based on the number of image batches, as shown in further detail in Fig. 2. Each mapper applies the MobileNetV3 model to its subset of images, extracting high-level features in the form of logits and class probabilities. The output of this phase is structured as key-value pairs, where the key is the predicted class label, and the value is the corresponding feature prediction. The mapping process runs in parallel, leveraging batch-wise parallelism across nodes.

Fig. 2
Fig. 2
Full size image

MapReduce-based distributed deep learning process for potato leaf disease detection.

MobileNetV3-Large is selected as the base model due to its efficiency and low computational cost. It utilises depth wise separable convolutions to reduce the parameter count, along with Hard-Swish activation in intermediate layers and SoftMax in the output layer. The network architecture includes a dropout rate of 0.4 to prevent overfitting, a Global Average Pooling layer for dimensionality reduction, and fully connected layers mapping from 1280 feature dimensions to the three target classes: healthy, early blight, and late blight. The model is initialised with pretrained ImageNet weights to enable effective transfer learning, as shown in Fig. 3.

Fig. 3
Fig. 3
Full size image

Deep learning-based classification pipeline using MobileNetV3 for potato leaf disease detection.

During training, the system uses the Adam optimiser, with an initial learning rate of 0.001. The model is trained for 52 epochs and a batch size of 32, and the learning rate is dynamically adjusted using the ReduceLROnPlateau scheduler, which monitors the validation loss.

In the shuffling phase, feature vectors and predictions from the mapping stage are grouped by class labels. This enables class-wise redistribution of intermediate outputs, supporting more efficient reduction and classification. Buffered I/O and memory caching enhance the shuffling process, and fault tolerance is ensured by backing up intermediate results to disk. The reducing phase aggregates the shuffled outputs using a SoftMax-based merging function, resulting in final predictions. The reduced outputs are stored as predicted class labels, handled by a set of compute reducers.

Let the input dataset be:

$$D = \{ (x_{i} ,y_{i} )\}_{i = 1}^{N}$$
(1)

where \(x_{i} \in {\mathbb{R}}^{224x224}\) is the \(i^{th}\) Image. \(y_{i} \in \{ 0,1,2\}\) represent the class label (Healthy = 0, Early Blight = 1, Late Blight = 2) and N = 2152 on the Eq. (1).

Data Augmentation is represented as a transformation function shown in Eq. (2)

$$x_{i}{\prime} = \Gamma (x_{i} )$$
(2)

where \(\Gamma\) Includes random operations like rotation, flipping, zooming, and brightness adjustment.

Each pre-processed image \(x_{i}{\prime}\) is passed through the MobileNetV3 model to extract deep features in Eq. (3)

$$f_{i} = GAP(\phi (xi\prime )) \in R^{1280}$$
(3)

In Eq. (3): \(\phi (x)\) is the convolutional output of MobileNetV3. GAP is a Global Average Pooling layer that compresses feature maps.

Then, logits and class probabilities are computed in Eq. (4):

$$z_{i} = Wf_{i} + b\left( {logits} \right)$$
(4)
$$p_{i} = Softmax(z_{i} ) = \frac{{e^{Zic} }}{{\sum\nolimits_{j = 1}^{3} {e^{zij} } }}$$
(5)

where Eqs. (4, 5) \(W \in {\mathbb{R}}^{3x1280} ,b \in {\mathbb{R}}^{3}\), \(p_{i} = [p_{i1} ,p_{i2} ,p_{i3} ]\) gives the probability for each class.

Mapper Output in Eq. (6)

$$Map Output: (k_{i} ,v_{i} ) = (argmax(p_{i} ),p_{i} )$$
(6)

In Eq. (7), predictions are grouped by predicted class label:

$$G_{c} = \{ p_{i} |argmax(p_{i} ) = c\} ,for c \in \{ 0,1,2\}$$
(7)

The shuffle phase is performed in Eq. (8):

$$Partition(p_{i} ) = Hash(k_{i} )modR$$
(8)

where \(k_{i}\) is the predicted class label. \(R\) is the number of reducer nodes.

Each reducer aggregates the predictions for a class. \(c\) using averaging in Eq. (9):

$$\rho_{c} = \frac{1}{{|G_{c} |}}\sum\limits_{{p_{i} \in G_{c} }} {p_{i} }$$
(9)

Final prediction for each class group in Eq. (10)

$$\Psi c = argmax(\rho_{i} )$$
(10)

The model is trained by minimizing the categorical cross-entropy loss in Eq. (11):

$$L = \sum\limits_{i = 1}^{N} {\sum\limits_{c = 1}^{3} {y_{ic} } \log (p_{ic} } )$$
(11)

where \(y_{ic} = 1\) if true label is c, else 0, \(p_{ic}\) is the predicted probability for class c.

$$\theta \leftarrow \theta - \eta \cdot \nabla_{\theta } L$$
(12)

In Eq. (12) \(\theta\) are the model weights. \(\eta\) is the learning rate (initially set to 0.001, adjusted by ReduceLROnPlateau).

For evaluation, the model’s performance is assessed using classification metrics, including accuracy, sensitivity, specificity, and confusion matrices. This configuration, which combines lightweight MobileNetV3 with distributed MapReduce processing, strikes an effective balance between accuracy, making the system well-suited for large-scale monitoring of agricultural diseases.

Results

The performance evaluation of the proposed MapReduce-based MobileNetV3 framework for detecting potato leaf disease was conducted through comprehensive training, validation, and testing experiments. A deeper analysis of the model’s classification performance was conducted using confusion matrices for the training, validation, and testing phases, as illustrated in Figs. 4, 5, and 6, respectively.

Fig. 4
Fig. 4
Full size image

Training confusion matrix of the proposed model.

Fig. 5
Fig. 5
Full size image

Validation confusion matrix of the proposed model.

Fig. 6
Fig. 6
Full size image

Testing the confusion matrix of the proposed model.

The training confusion matrix, as shown in Fig. 4, indicates that the model achieved high true positive rates across all three classes, with minimal false positives and false negatives. This suggests that the model accurately learned the distinguishing features during the training process. There are three classes of datasets.

The validation confusion matrix, as shown in Fig. 5, similarly reflects high classification performance. However, it reveals slightly higher misclassifications compared to the training phase, which is expected when evaluating on unseen data. Class 0 has 69 true positives, 5 false negatives, 1 false positive, and 140 true negatives.

Notably, the testing confusion matrix, as shown in Fig. 6, confirms the model’s strong generalization capability, as it maintains high true positive rates with minimal errors on the independent test set. The low false positive and false negative counts across all classes highlight the model’s robustness in distinguishing even visually similar disease classes.

Metrics such as sensitivity, specificity, detection accuracy, miss classification rate, fallout, likelihood positive ratio, likelihood negative ratio, positive predictive values, and negative predictive values are calculated. The proposed model calculated the output using multiple statistical measures, as outlined in Eqs. (13)–(21), in comparison with the counterpart for potato disease.

$$Sensitivity = TPR = \frac{\sum TP}{{\sum TP + FN}}$$
(13)
$$Specificity = TTNR = \frac{\sum TN}{{\sum TN + FP}}$$
(14)
$$Detection\;Accuracy = \frac{{\sum TP + { }\sum TN}}{\sum P + N}$$
(15)
$$Miss{\text{ Classification }}Rate = 1 - Detetion\;Accuracy$$
(16)
$$Fallout = FPR = \frac{\sum FP}{{\sum FP + TN}}$$
(17)
$$LR + = \frac{\sum TPR}{{\sum FPR}}$$
(18)
$$LR - = \frac{\sum FPR}{{\sum TPR}}$$
(19)
$$PPV = \frac{\sum TP}{{\sum {\text{ T}}P + FP}}$$
(20)
$$NPV = \frac{\sum TN}{{\sum {\text{ TN}} + FN}}$$
(21)

Figure 7 shows the training accuracy and loss graphs. The lines represent the model’s accuracy and loss on the training and validation datasets throughout the epochs. Training accuracy increases rapidly in the initial epoch, indicating that the model is learning and fitting the training data. The validation accuracy also increases, but at different rates. After a certain number of epochs, accuracy continues to improve. In the proposed model’s first epoch, the training accuracy starts at 0.60, and the validation accuracy is approximately 0.75. However, at the 10th epoch, the proposed model’s training accuracy approaches 0.95, and validation accuracy is approximately 0.92. Finally, at the 52th epoch, the proposed model converges with its best accuracy, 0.980 and 0.958, during training and validation, respectively. Similarly, in Fig. 8, the training loss decreases rapidly in the initial epoch, indicating that the model is learning and reducing errors in the training data. The smooth convergence of both losses towards lower values implies stable training with an appropriate learning rate.

Fig. 7
Fig. 7
Full size image

Training and validation accuracy curves of the proposed MobileNetV3-based model.

Fig. 8
Fig. 8
Full size image

Training and validation loss curves of the proposed MobileNetV3-based model.

Table 1 presents the performance metrics of the proposed model’s statistical measures for training, validation, and testing. The overall result of the training model of detection accuracy is 0.980, sensitivity is 0.970, specificity is 0.985, miss classification rate is 0.020, the fallout is 0.015, LR+ is 65.861, LR− is 0.015, PPV is 0.970, and NPV is 0.985. The result of the validation model of detection accuracy is 0.969, sensitivity is 0.954, specificity is 0.977, miss classification rate is 0.031, the fallout is 0.023, LR+ is 67.084, LR− is 0.024, PPV is 0.954, and NPV is 0.977 shown in Table 2. Similarly, Table 3 shows the testing statistics. The testing model of detection accuracy is 0.969, sensitivity is 0.953, specificity is 0.977, miss classification rate is 0.031, the fallout is 0.023, LR+ is 41.866, LR− is 0.024, PPV is 0.953, and NPV is 0.977.

Table 1 Statistical performance metrics of the proposed MobileNetV3-based model.
Full size table
Table 2 Statistical performance metrics of the proposed MobileNetV3 validation model.
Full size table
Table 3 Statistical performance metrics of the proposed MobileNetV3 testing model.
Full size table

Table 4 shows that multi-GPU (PyTorch DistributedDataParallel (DDP) TensorFlow (TF) shortens training time by ~ 1.6–2.3 × with the same accuracy, while MapReduce targets data-pipeline throughput rather than training speed. Model training was conducted using TensorFlow 2.14 and PyTorch 2.2 with CUDA 11.8 and cuDNN 8.6 support. Image preprocessing utilized OpenCV 4.8, NumPy 1.26, and Pandas 2.1. Visualization was performed using Matplotlib 3.8. The distributed data-processing pipeline was implemented using Apache Hadoop MapReduce 3.3.5.

Table 4 Training baselines: single-GPU vs. DDP/TF.
Full size table

Table 5 shows MobileNetV3 leads overall (96.8% accuracy, 96.4 F1) with the best balance high sensitivity (95.3) and top specificity (97.7). EfficientNet-B3 and DenseNet-121 are well-balanced contenders (~ 91–92% accuracy), while the ResNet variants trail on both recall and F1.

Table 5 Comparative performance of the proposed MobileNetV3.
Full size table

Discussion

The proposed MobileNetV3 classifier attains strong accuracy on a 3-class potato leaf task (test = 96.8%) with stable training dynamics. The MapReduce pipeline reliably parallelizes preprocessing and batch inference, delivering throughput gains when image volume or ingest concurrency increases. On the current small dataset, MR does not improve model accuracy and, due to job and shuffle overhead, can be slower than a well-tuned single-node pipeline.

MapReduce vs. distributed training. MapReduce and distributed training target different bottlenecks. PyTorch DDP/TF-Mirrored shorten training wall-clock on small data by ~ 1.6–2.3 × with accuracy parity (Table 4), while MapReduce improves data-pipeline throughput (ingest → augment → forward → write) sources scale. In practice they are complementary: use DDP/TF when gradient-synchronized compute is the limiter; use MapReduce when I/O and preprocessing dominate.

Limitations

The dataset is modest and may not capture field variability (lighting, occlusion, mixed symptoms). Results may overestimate performance in the wild.

Future work

  • Larger, heterogeneous field datasets (multi-site, multi-season) to validate robustness under domain shift.

  • Hybrid pipelines: MR for ETL + DDP/TF for training and model-parallel batched inference on GPU-enabled mappers; explore Ray/Spark Structured Streaming alternatives.

  • Severity estimation & multi-task learning (type + severity + quality flags).

Table 6 presents a comparison of the proposed model with existing studies18,19,20,21,22,23,24,25,26,27,28,29,30, highlighting that the proposed MapReduce-based MobileNetV3 framework achieves the highest classification accuracy of 96.9%, outperforming previous methods that lacked distributed processing capabilities. This demonstrates the effectiveness of combining remote sensing, deep learning, and scalable computing for detecting potato diseases.

Table 6 Comparative analysis of the proposed MobileNetV3 with MapReduce framework against existing models.
Full size table

Conclusion

This study introduces an innovative and scalable deep learning framework to accurately identify and distinguish potato leaf disease. The proposed system, utilizing the lightweight MobileNetV3 architecture with the distributed computing architecture of MapReduce, is beneficial in addressing the dual challenge of efficiency and scalability in large-scale agricultural disease monitoring. The structure also leverages high-performance preprocessing, data augmentation, and transfer learning to enhance the model’s generalization capability. On the publicly available dataset with 2152 labeled images, the model has been demonstrated to be highly accurate during training, validation, and testing procedures (accuracy of 96.8 percent overall in classification). The model was highly sensitive, specific, and had good predictive values, as confirmed in the statistical evaluation, indicating that the model is robust in distinguishing between healthy leaves and those affected by early and late blight. Moreover, the proposed framework outperformed state-of-the-art techniques, demonstrating the applicability of both deep learning and distributed computing in precision agriculture.

The study can advance the nascent field of smart agriculture by introducing an inexpensive, effective, and scalable system for early disease identification. The system can support large datasets and process them quickly and accurately, making it useful for real-time deployment in the field, especially in resource-limited environments. Future research directions will include generalizing this framework to identify additional types of crop diseases, integrating temporal disease evolution analysis, and examining how the framework can be deployed on edge computing devices to enhance accessibility and inference speed.