Development, performance evaluation and prediction of optimal operational conditions for a double-row sugarcane harvester using deep learning

Abstract

Sugarcane is a vital global crop, serving as a primary source of sugar, biofuel, and renewable energy. Advancements in harvesting are critical to meeting rising demand, enhancing profitability, and supporting eco-friendly agricultural practices in the sugarcane sector. Based on the current challenges of sugarcane harvesting in developed countries, the current study aimed to develop a semiautomatic whole-stalk sugarcane harvester (SWSH) for harvesting two rows of sugarcane stalks at a time and to be front-mounted on a classic four-wheel agricultural tractor. Then performance evaluation and prediction of optimal operational conditions for a double-row sugarcane harvester using Feedforward Neural Network (FNN) and Deep Neural Network (DNN) at different levels of forward speeds (3, 3.5, 4.5, and 5 km/h), row spacing (71, 78.89, and 88.75 cm), cutting heights (0, 2, and 4 cm), and numbers of knives (2 and 4) of the cutting systems. The obtained results showed that the cutting efficiency of the developed SWSH reached 100%. Where the higher cutting efficiency was observed at a cutting height equal to zero (ground level), forward speed of 3 km/hand row spacing of 71 cm using both 2 and 4 knives. The minimum total operating cost of the developed SWSH was about 4.42 USD/ha, and it was detected when using a forward speed of 4.5 km/h, row spacing of 88.75 cm, a cutting height of 4 cm, and two knives only on the cutting disk. Furthermore, at a row spacing of 88.75 cm, the maximum field capacity of the developed SWSH was 0.554 ha/h, observed at a forward speed of 4.5 km/h.

Introduction

Sugarcane is one of the world’s most vital crops, playing a crucial role in agriculture, industry, and the global economy. As the primary source of sugar, it meets over 70% of the world’s sweetener demand, making it essential for food and beverage production. Beyond sugar, sugarcane is used to produce biofuels like ethanol, offering a renewable energy alternative that reduces reliance on fossil fuels and lowers carbon emissions^1,2,3,4. Economically, sugarcane supports millions of farmers and workers, particularly in tropical and subtropical regions such as Brazil, India, and Thailand. Its byproducts, including bagasse and molasses, are used in animal feed, paper production, and electricity generation, enhancing sustainability. Additionally, sugarcane cultivation improves soil health through crop rotation and contributes to rural development^5,6,7. Environmentally, sugarcane helps sequester carbon dioxide, mitigating climate change. Its versatility and renewable nature make it a key crop for food security, energy sustainability, and economic growth, highlighting its global significance^{8,9,10,11,12,13}. Sugarcane is a major cash crop globally, in 2023, cultivated on over 26 million hectares worldwide, with leading producers like Brazil, India, and Thailand dominating production. Based on production volume in million metric tons. In that year, Brazil produced some 782.59 million metric tons of sugarcane¹⁴. In Egypt, sugarcane holds significant agricultural importance, primarily grown in the southern regions, particularly in Upper Egypt (Aswan, Qena, Luxor, and Minya). The total cultivated area in Egypt is approximately 143,808 ha, contributing to an annual production of 15–16 million tons in season 2020–2021^4,15,16.

The manual cutting is very labor-intensive, the workers usually become fatigued after a few hours, and they need frequent pauses for rest¹⁷. Due to the high levels of sun exposure, Precautions need to be taken to limit or protect the workers because it can result in various types of skin cancer conditions¹⁸. Due to wages increasing and the unavailability of labor to cut sugar cane by hand, The South African Industry considered various options for mechanical sugar cane harvesting¹⁹. Hence, mechanization of sugarcane harvesting is essential not only for reducing the production cost but also for reducing hard work involved in manual harvesting operations and also to ensure quality produce^5,20. Since sugarcane production is entirely mechanized, Hawaii, Australia, Southern USA, and Japan have pioneered mechanical harvesters. Brazil, India, Cuba, South Africa, and China have large sugar cane plantations that utilize complete mechanization, medium-sized farms that employ semi-mechanization, and small farms that harvest manually.

No sugarcane harvester has been successful developed in Egypt. Several trials have been conducted to evaluate the performance of imported sugarcane harvesters under local conditions. However, local farmers have largely rejected these machines due to their limited benefits in terms of cost savings, labor reduction, and time efficiency. As a result, despite their proven technical performance, sugarcane harvesters have not been widely adopted. Additionally, previous research efforts by graduate students to design and test locally developed sugarcane harvesters have also been unsuccessful²¹. According to the local manufacturing of sugarcane harvester, three distinct sugarcane harvester prototypes have been developed through graduate research programs at Egyptian universities. These machines were designed to address key challenges in manual harvesting, such as labor fatigue, harvesting losses, and operational costs^21,22,23. Where three sugarcane harvester prototypes were developed between 2002 and 2014 represent incremental progress in mechanization yet highlight persistent challenges. While each iteration showed improvements in power (from 6hp to tractor PTO to 14hp) and field capacity (0.0294–0.05 ha/h), all models struggled with lodged cane, labor dependency, and inconsistent performance. The most promising aspects were the 2011 model’s increased capacity and 2014 model’s better maneuverability, but neither achieved sufficient labor savings nor cost efficiency to justify widespread adoption. Future development must prioritize self-propelled operation, intelligent crop handling systems, and rigorous economic analysis to create truly viable solutions for Egypt’s sugarcane farmers. These prototypes provide valuable foundations, but substantial innovation remains needed for effective mechanization^24,25.

Key challenges for sugarcane harvesters in Egypt include, most Egyptian sugarcane farms are small (0.5–1 hectare), with narrow row spacing and irregular field shapes, making it difficult for large, imported harvesters—designed for bigger, more uniform fields—to operate efficiently or maneuver effectively^26,27; Imported harvesters are often too expensive for local farmers, and their operation is not economically viable given the scale and profit margins of Egyptian sugarcane production ; Many imported machines are designed for different agricultural systems, not accounting for the specific characteristics of Egyptian sugarcane (e.g., stalk size, density, and trash content), leading to poor performance in cutting, cleaning, and handling²⁸; Existing harvesters have shown low efficiency in Egyptian conditions, particularly in de-trashing and base-cutting, due to the high force required for cutting and removing leaves from local cane varieties²⁶; and Loader efficiency is also reduced by the need to move between multiple small storage sites, causing time losses and delivery delays²⁷. So, the imported sugarcane harvesters have generally failed in Egyptian conditions due to incompatibility with small farm sizes, high costs, and poor adaptation to local agronomic practices. Locally designed and fabricated harvesters, tailored to the specific needs of Egyptian fields and cane varieties, have demonstrated better efficiency, cost-effectiveness, and suitability for the local context^26,27.

The integration of Machine Learning (ML) and Deep Learning (DL) represents a paradigm shift in the optimization of agricultural machinery, moving from traditional, often trial-and-error methods towards data-driven, predictive approaches²⁹. For harvesting operations, key applications include real-time control and operational parameter prediction. For instance, deep neural networks have been employed to analyze multi-sensor data to dynamically predict and regulate harvester ground speed, balancing throughput and losses³⁰. Similarly, intelligent systems using header power shaft torque and other parameters have been developed to estimate harvester feed rate, a critical variable for efficient operation³¹. Beyond real-time control, predictive maintenance models using ML forecast machine failures, enabling proactive scheduling to reduce downtime and costs^32,33,34.

The choice of ML architecture is critical and is dictated by the nature of the data and the prediction task. While complex models like Convolutional Neural Networks (CNNs) excel with image-based data³⁵, and Recurrent Neural Networks (RNNs) are suited for time-series analysis³⁶, the prediction of optimal operational parameters from a set of numerical inputs is a classic regression problem. In this domain, Fully Connected Neural Networks (FNNs) and Deep Neural Networks (DNNs) have proven highly effective. Their architecture is inherently designed to learn complex, non-linear relationships between multiple continuous input and output variables. FNNs offer a balance between model complexity and computational efficiency, making them suitable for applications with well-defined numerical relationships³⁷. In contrast, DNNs, with their deeper layers, can capture more abstract and intricate interactions within the data, potentially leading to higher prediction accuracy for complex systems³⁸. This justifies their selection for the current task of modeling the multifaceted performance of a sugarcane harvester as a function of its operational parameters. ³⁷The automation and intelligent applications of sugarcane harvester have the potential to reduce the labor intensity, solve labor shortage issues, and provide more efficient quality in harvesting operations. Accordingly, research has been done using machine learning and especially advanced deep learning algorithms. Control harvester’s speed is one of these applications, where it proved the operational performance enhancement without any source of manual intervention. For example, Chen et al. ²⁹ used deep neural networks to analyze multi-sensor data to predict and regulate the harvester speed. Achieving the balanced of both higher power density and lower resonant frequency is also one of the main problems that could be solve by deep learning. Chimeh et al. ³⁰ provided a piezoelectric MEMS energy harvester with the ability to work at a low resonant frequency.

Monitoring and predicting harvester feed rate is a clear significance for guiding harvesting work and improving harvesting efficiency. Sun et al.³¹ proposed an intelligent system considering header power shaft torque, header height, and grains moisture content to estimate harvester feed rate. Harvester maintenance prediction is an interesting direction, where machine learning forecasts machine failures, and provides proactive maintenance scheduling to reduce downtime and costs^32,33,34.

As well, it is possible to predict optimal maintenance intervals considering operating conditions, usage patterns, and historical performance to increase energy harvester lifespan and efficiency.

³⁷Addressing the challenges of mechanizing sugarcane harvesting in Egypt requires machinery that is not only locally adapted but also intelligently optimized. While previous efforts have focused on mechanical design, a critical gap remains in the predictive modeling of harvester performance. Specifically, there is a lack of robust models capable of accurately predicting key performance indicators (e.g., cutting efficiency, operational cost, field capacity) under the complex interplay of multiple operational parameters—forward speed, row spacing, cutting height, and knife configuration—in the context of small-scale Egyptian farms. This study, therefore, aims to bridge this gap by: (1) developing a semiautomatic whole-stalk sugarcane harvester (SWSH) front-mounted for use with a classic agricultural tractor, and (2) employing and comparing Feedforward Neural Network (FNN) and Deep Neural Network (DNN) models to evaluate performance and predict the optimal operational conditions for this double-row harvester. This integrated engineering-and-AI approach provides a comprehensive solution for enhancing the efficiency and economic viability of sugarcane harvesting in Egypt.

Materials and methods

Experimental setup

All field experiments were conducted using the G.T.54 − 9 sugarcane variety under uniform agronomic conditions. A total of 216 field trials were performed, representing 72 treatment combinations (four forward speeds × three row spacings × three cutting heights × two knife numbers), each replicated three times. To minimize field variability, a stratified randomization protocol was applied in which treatment sequences were randomly assigned within each replicate block, and test order was alternated across different field sections to mitigate positional and time-of-day effects. Prior to each trial, key agronomic parameters influencing cutting performance were documented, including the average stalk diameter for each row spacing (2.63 cm at 71 cm, 3.12 cm at 78.89 cm, and 3.76 cm at 88.75 cm), crop density measured using five randomly selected quadrats. These measurements ensure accurate interpretation of the machine’s performance and strengthen the reproducibility of the experimental methodology.

Description of the developed SWSH

During the current study, a whole-stalk sugarcane harvester (SWSH) was developed using available materials in local markets in Egypt and recycling some materials to decrease the manufacturing costs. There are many points taken into consideration during the development, such as the developed SWSH being applicable with small and medium holdings, working at different levels of cutting height, low manufacturing and operating costs, low operating labor, and light weight. The developed SWSH was constructed and evaluated during summer session 2017, at the research farm, Faculty of Agriculture, Minia University, Egypt. Figure 1 presents the detailed methodology of the study and it will be explained in detail in the next few subsections.

Fig. 1

Detailed methodology of the study.

Figure 2 shows the main components of the developed SWSH. Where it consists of many parts, including (1) machine frame, (2) cutting system, and (3) power transmission system.

Fig. 2

Main components of the developed tractor operator double-row sugarcane whole-stalk harvester (SWSH). Whereas (A) crown gearbox, (B) installing fixed frame on the front side of the tractor, (C) hydraulic system hoses, (D) connecting the power transmission shaft with the tractor front wheels operating shaft; (1) movable frame, (2) sliding shoes, (3) cutting system, (4) hydraulic pistons, (5) connecting rods, (6) fixed frame, (7) sugarcane stalk push arm, and (8) Power cultch.

Machine frame

As shown in Fig. 2[part no. 1], the machine frame of the developed SWSH consists of three main parts. (1) Movable frame: it was installed on the front side of the developed SWSH. It supports the cutting system and power transmission system. It connected from the back side with the fixed frame by four connecting arms. On the front side, it slid over the earth’s surface using three sliding shoes. These sliding shoes were used for controlling the cutting height by adjusting the vertical height between the cutting knives and the sliding shoes. The movable frame has a dimension of 140 cm in height, 70 cm in width, and 167 cm in length. The movable was fabricated from channel bars (100 mm in length and 50 mm in width) and angle bars (51 in length, 51 in width, and 3.2 mm in thickness), then covered with galvanized steel sheets 3 mm in thickness. and the total weight of the movable frame was about 240 kg. Additionally, as shown in Fig. 2[part no. 7], a sugarcane stalk push arm was installed on the front side of the movable frame above 45 cm from the soil surface. Where it was responsible for pushing the sugarcane stalks by about 75° to facilitate the cutting process. The sugarcane stalk push arm was constructed using a steel pipe that has a dimension of 101.6 mm in diameter and 3.2 mm in thickness. (2) Fixed frame: it was installed on the front side of the four-wheel tractor (Fig. 2[part no. 6]). Where it was fixed with the tractor chassis by eight bolts (M16) at each side. The fixed frame was constructed out of channel bars with a dimension of 100 mm in width and 50 mm in edge and angle bars with a dimension of 51 mm in length, 51 mm in width, and 3.2 mm in thickness. The fixed frame has a dimension of 83 cm height and 120 cm length, and the width was manufactured to be adjustable with a wide range of tractor models ranging from 50 to 80 cm in width. The overall weight of the fixed frame was about 59.5 kg. (3) The connecting rods: they were used to connect the removable frame with the fixed frame. The connecting rods consist of four steel arms made of channel bars that have a dimension of 100 mm in length and 50 mm in width. The total length of the connecting rods was about 40 cm, and the weight of each one was about 4.5 kg. And it has one hole 3 cm in diameter at each side. As shown in Fig. 2 [part No. 4], two hydraulic pistons were used for lifting the movable frame; each of them has a 40 cm closed length, a 70 cm open length, and a 500 kg lifting weight. And they were connected to the tractor-front hydraulic valves with two hydraulic hoses (Fig. 2C).

Cutting system and power transmission systems

The cutting system was used for clean cutting of sugarcane stalks and reducing losses. The cutting system consists of two main parts of cutting disks as shown in Fig. 3. It was used for cutting sugarcane stalks from two adjacent rows. The cutter system was installed on the movable frame (Figs. 1 and 2A). As shown in Fig. 3, the cutting system consists of many parts including (1) Cutting knives: they were constructed of tempered steel and have a dimension of 20 cm in length, 10 cm in width and 1.5 cm in thickness and were shaped with 30° on one side. (2) Knives holder disk: the knives were temporarily fixed with the knives holder disk with two bolts (M16). The knives holder disk was manufactured from meld steel, has a dimension of 46 cm in diameter, and 1.5 cm in thickness. it was adjusted to hold four knives in his circumference (Fig. 3A and D). (3) Deflectors: On the upper part of the knives holder disk, four deflectors were welded and symmetrically distributed on its circumference. It was used to deflect the sugarcane stalks from the cutter head after the cutting process and deliver them through an outlet below the harvester and the tractor.

Fig. 3

Main components of cutting system and power transmission system. Whereas (A, and B), cutting system, (C and D), power transmission system, (1) upper pivot shaft, (2) coupling, (3) power transmission shaft, (4) power delivery shaft (small sprockets), (5) large sprocket, (6) crown gear, (7) coupling, (8) lower pivot shaft of the cutting disk, (9) cutting knives, (10) deflector, 11. knives holder disk.

The cutting knives used in sugarcane harvesters are subjected to continuous mechanical stress and abrasive contact with fibrous stalks, soil, and other field debris. Therefore, the selection of appropriate material and mechanical properties is critical to ensure durability, operational efficiency, and safety. To meet the rigorous demands of sugarcane harvesting, cutting knives are typically manufactured from high-carbon alloy steels or hardened tool steels that possess high tensile strength and wear resistance. during the current study, spring Steel (EN47) was used for its high fatigue strength and flexibility after cutting the cutting knives with the desired dimensions. The cutting knives were heat-treated to achieve a balance between toughness and hardness. Proper quenching and tempering procedures were applied to minimize brittleness while maintaining sharpness and strength. The tensile strength, fatigue limit and wear rate of cutting knives used in developed sugarcane harvester was about 900 MPa, 390 MPa, 5 × 10⁻⁵ mm³/N·m, respectively, to ensure operational reliability and resistance to mechanical failure. A comprehensive heat-treatment protocol was applied to enhance the mechanical performance of the cutting knives. The EN47 steel was first austenitized by heating to 820–840 °C, followed by oil quenching to achieve high hardness while minimizing thermal shock. Subsequently, the quenched blades were tempered at 480–520 °C for 1.5–2 h to refine toughness and stabilize the microstructure. This thermal sequence yielded mechanical properties suitable for demanding field conditions, including a tensile strength of approximately 900 MPa, a fatigue limit of about 390 MPa, and a wear rate of 5 × 10⁻⁵ mm³/N·m, collectively ensuring reliable cutting performance, minimal deformation under load, and strong resistance to mechanical failure during harvesting operations.

Power transmission system

During the current study, the developed SWSH was front-mounted and operated by an agricultural tractor (model: Belarus MTZ-82) with a rated power of 80 hp [59.7 kW] (Fig. 2B, C, and D). The power transmission system of the developed SWSH consisted of four main components: universal joints, shafts, differential gear, and chains and sprockets. It was designed to achieve a wide range of cutting speeds, depending on the forward speed, since the machine draws power from the tractor’s front axles (Fig. 2D), providing a constant speed ratio of 5.59 at any forward speed. Additionally, cutting velocity can change the speed ratio by adjusting the reduction ratio between the chains and sprockets, which is approximately 3.5:1 within a speed range of 1 to 5 km/h. For saving the power transmission system from over cutting load, a power cultch (Fig. 1, part no. 8) was installed, and it was operated when the cutting power excessed 50 kW. For comparative analysis, prior studies have reported velocity ratios within a range of 13 to 22^39,40,41, with a commonly assumed reference value of 20 in similar agricultural machinery applications. The front wheels of the Belarus MTZ-82 tractor were selected with dimensions of 8.3–20 inches, ensuring compatibility with standard agricultural implements. These dimensions influence the machine’s ground contact, traction, and overall performance. Visual references for the wheel configuration and power extraction mechanism are provided in), illustrating the integration of the experimental setup with the tractor’s front axle.

Performance evaluation of the developed SWSH

The main factors affecting the performance of the developed SWSH are field capacity, throughput capacity, power requirements, cutting efficiency, and total operating costs of the machine. Table 1 shows the experimental factors and their levels of the developed SWSH.

Table 1 Experimental factors and their levels.

All experiments were carried out using the C 9 sugarcane variety. Each test was repeated a minimum of three times, and the average of the recorded values was used for subsequent calculations. Each experimental plot measured 30 m in length, with the width corresponding to the effective working width of the machine. This setup was used to ensure consistent evaluation of machine performance under uniform field conditions. All experiments were conducted under field conditions to assess the impact of controlled factors on the performance of this prototype harvesting unit. During the field evaluation, several precision instruments were utilized, including a digital stopwatch with an accuracy of ± 0.01 to ± 0.1 s, a 30-meter measuring tape with an accuracy of ± 0.5 cm per 30 m, and a graduated cylinder (0–5 l) for fuel measurement with an accuracy of ± 2 mL.

Theoretical field capacity (FC)

The theoretical field capacity (ha/h) of a machine refers to the actual area covered, or material processed per unit time during field operations, accounting for operational delays. It is a key performance metric in agricultural machinery, determined by factors such as working speed, working width, field efficiency, and downtime for adjustments or maintenance. Unlike theoretical capacity, effective field capacity considers real-world conditions like terrain, crop density, and operator skill. Optimizing this capacity ensures higher productivity, reduced operational costs, and efficient resource use. For sugarcane harvesters, improving effective field capacity involves minimizing idle time, proper machine calibration, and adapting to field conditions to maximize harvesting efficiency. The effective field capacity of the developed SWSH was calculated according to Eq. (1).

$$\:Fc=\:\frac{A}{T}$$

(1)

where, A is the harvested area, ha, and T = time consumed, h.

Throughput capacity

The throughput capacity of a sugarcane harvester is a critical performance indicator that measures the amount of cane harvested per unit of time, typically expressed in tons per hour. It depends on factors such as harvester speed, cutting mechanism efficiency, crop density, and field conditions. High throughput capacity ensures timely harvesting, reduces operational costs, and maximizes productivity. However, challenges like crop variability, machine maintenance, and operator skill can affect performance. Optimizing throughput involves proper machine calibration, minimizing downtime, and adapting to field variations. The throughput capacity of the developed SWSH was calculated according to Eq. (2).

$$\:Mc=\:Fc\:\times\:Y$$

(2)

where, Y is the sugarcane yield, t/ha.

Cutting efficiency

The cutting efficiency ($\:{\eta\:}_{c}$) of the developed SWSH was determined by selecting a random harvested area equal to 3 m², then counting the total number of sugarcane stalks, number of clean-cut stalks and partially and no cut stalks, then the cutting efficiency of the developed SWSH was determined according to Eq. (3).

$$\:{\eta\:}_{c}=\:\frac{Nc}{Nt}\:\times\:100$$

(3)

where, $\:Nc$ is the number of sugarcane stalks completely cut, $\:Nt$ is the total number of sugarcane stalks.

Fuel consumption

Fuel consumption was recorded as the total fuel usage of the MTZ-82 tractor while operating the SWSH attachment, as measured using a 0–5 l graduated cylinder with ± 2 mL precision. Because the cutting system draws mechanical power directly from the tractor’s front axle and no independent auxiliary fuel circuit exists, it is not technically feasible to isolate the fuel consumption attributable solely to the harvester attachment. Therefore, all reported fuel consumption values represent the combined energy demand of both the tractor engine and the front-mounted SWSH unit. This approach reflects the actual operational fuel requirement experienced by farmers and ensures that the economic analysis (USD/h and USD/ha) accurately captures real-world harvesting costs. A volume of fuel consumed ($\:{F}_{c}$) was measured during each test by estimating the fuel consumption and time consumed for each trial. The fuel consumption was calculated using Eq. (4).

$$\:{F}_{c}=\frac{v\times\:3.6}{t}$$

(4)

where, $\:{F}_{c}$ is the fuel consumption, l/h, $\:v$ is the volume of fuel consumed, cm³, t is the time of the test, s.

Power requirements

The power requirements for harvesting sugarcane stalks were calculated according to Eq. (5).

$$\:{P}_{b}=\frac{{F}_{c}\times\:{C}_{v}\times\:{\eta\:}_{th}}{3600}$$

(5)

where, $\:{P}_{b}$ is the brake horsepower, kW, $\:{C}_{v}$ is the calorific value (44800 kJ/kg for diesel), $\:{\eta\:}_{th}$ is the thermal efficiency, %.

Artificial neural network modeling for performance prediction

To predict the optimal operational conditions for the developed sugarcane harvester, two artificial neural network architectures were developed and compared: a Deep Neural Network (DNN) and a Feedforward Neural Network (FNN). The dataset (n = 180), comprising all combinations of the experimental factors (forward speed, row spacing, cutting height, and number of knives), was used for model development. The input variables (X, 1 × 8 vector) were: cutting height, number of knives, row spacing, average stalk diameter, forward speed, knife rotational speed, knife linear speed, and speed ratio. The corresponding output targets (Y, 1 × 9 vector) were the machine performance metrics: machine performance (a qualitative score), number of uncut stalks, cutting efficiency (%), field capacity (ha/h), throughput capacity (t/h), fuel consumption (l/h), total operating cost (USD/h), total operating cost (USD/ha), and power requirements (kW).

Deep neural network (DNN)

The DNN model is illustrated in Fig. 4. It represents the DNN structure, including the input/ output parameters. It consists of five main layers as follows:

• The input layer: X size 1 × 8 presents the system’s inputs of cutting height, No. of knives, Row spacing, Av. Stalk diameter, Forward speed, Knife rotational speed, Knife linear speed, and Speed ratio.

• The first FC1 connected layer: Implements Eq. 1 to input vector X and produces the hidden neurons HA of size 1 × 400.

$$\:{H}_{a}=\sum\limits_{p=1}^{8}{X}_{p\:}{w}_{pA}+\:bias$$

(6)

$$\:{H}_{A}=\:[{H}_{1}\:{H}_{2}..........\:{H}_{a}]\:\:\:\:\:\:for\:a=\:\text{1,2},3......400$$

(7)

where, H_A is the concatenated hidden neurons of size 1 × 400; X is the system input vector of size 1 × 8; W_PA is a weight matrix of size 8 × 400; bias and W_PM weight matrix are randomly generated. Next, a batch normalization operation and a hyperbolic tangent transfer function are applied to $\:{H}_{A}$ hidden neurons.

The general form of hyperbolic tangent transfer function is shown in Eq. (8) ⁴²:

$$\:{H}_{A}{\prime\:}=\frac{1\:-\:{e}^{-2\left({H}_{A}\right)}}{1\:+\:{e}^{-2\left({H}_{A}\right)}}$$

(8)

The following FC2 connected layer: It operates with HA’ hidden neurons to produce HB hidden neurons of size 1 × 200.

$$\:{H}_{B}=\sum\limits_{a=1}^{400}{H}_{a\:}^{\prime\:}\:{w}_{AB}+\:bias$$

(9)

$$\:{H}_{B}=\:[{H}_{1}\:{H}_{2}..........\:{H}_{b}]\:\:\:\:\:\:for\:b=\:\text{1,2},3......200$$

(10)

where H_B is the concatenated hidden neurons of size 1 × 200; H_A‘ is the hidden neurons vector of size 1 × 400 from the previous fully connected layer; W_AB is a weight matrix of size 400 × 200; bias and W_MA weight matrix are randomly generated. Next, FC2 is followed by batch normalization operation and a hyperbolic tangent transfer function to generate H_B’.

• The following FC3 fully connected layer: It operates with HB’ hidden neurons to produce HC hidden neurons of size 1 × 100.

$$\:{H}_{C}=\sum\limits_{b=1}^{200}{H}_{b\:}^{\prime\:}\:{w}_{BC}+\:bias$$

(11)

$$\:{H}_{C}=\:[{H}_{1}\:{H}_{2}..........\:{H}_{c}]\:\:\:\:\:\:for\:c=\:\text{1,2},3......100$$

(12)

Next, It is followed by batch normalization operation and a hyperbolic tangent transfer function to generate H_C’.

• The last FC4 fully connected layer: It operates with H_C’ hidden neurons and produces the final output vector Y_D of size 1 × 9.

$$\:{Y}_{d}=\sum\limits_{c=1}^{100}{H}_{c\:}^{\prime\:}\:{w}_{CD}+\:bias$$

(13)

$$\:{Y}_{D}=\:[{Y}_{1}\:{Y}_{2}..........\:{Y}_{d}]\:\:\:\:\:\:for\:q=\:\text{1,2},.....9$$

(14)

where Y_D is the concatenated output neurons of size 1 × 9 which represents machine performance, no. of uncutting stalks, cutter head efficiency, field capacity, throughput capacity, fuel consumption, total operating costs USD/h, total operating costs USD/ha, and power requirements; H_C‘ is the hidden neurons vector of size 1 × 100; W_CD is a weight matrix of size100 × 9; bias and W_CD weight matrix are randomly generated.

Fig. 4

The used construction of DNN.

The dataset (n = 180) was partitioned using with 4-fold cross-validation technique, and stratified random sampling to ensure proportional representation of all input parameters (forward speed, row spacing, cutting height, knife count). First, the data were grouped into 72 unique experimental combinations (4 speeds × 3 spacings × 3 heights × 2 knife configs). From each group, 70% of replicates were randomly assigned to training and 30% to testing, preserving the original distribution. The training performance was evaluated using Root Mean Square Error (RMSE), and the Mean Absolute Error (MAE)⁴³ :

$$\:RMSE\:=\:\sqrt{\sum\limits_{i=1}^{n}\frac{{({y}_{i}^{{\prime\:}}\:-\:{y}_{i})}^{2}}{n}}$$

(15)

where, RMSE is the root mean square error; yi’ is the predicted value; yi is the target value; and n is the total size of the database.

$$\:MAE=\frac{1}{n}\sum\limits_{n=1}^{n}\left|({\acute{y}}_{i}-{y}_{i})\right|$$

(16)

Where MAE is the mean absolute error, $\:{\acute{y}}_{i}\:$is the predicted value; $\:{y}_{i}\:$is the original value, and n is the number of data samples.

Feedforward neural network (FNN)

Figure 5 illustrates FNN design:

Fig. 5

The used construction of FNN⁴³.

The constructed FNN includes three hidden neurons, where each one has 100 neurons.

• The input layer.

• The three hidden layers ‘equations:

$$\:H1\:=\:f\left(\:\sum\limits_{p=1}^{8}{X}_{p}{W}_{8x100}\:+\:hb1\:\right)$$

(17)

$$\:H2\:=\:f\left(\:\sum\limits_{n=1}^{100}H1\:{W}_{100x100}\:+\:hb2\:\right)$$

(18)

$$\:H3\:=\:f\left(\:\sum\limits_{n=1}^{100}H2\:{W}_{100x100}\:+\:hb3\:\right)$$

(19)

where H1, H2, and H3 are the generated three hidden layers’ neurons of size 1 × 100; X_P represents the system input of size 1 × 8; hb1, hb2, and hb3 are the bias values; W is a random generated weight matrix; and f is the Levenberg-Marquardt transfer function.

• The output layer:

$$\:O=\:f\left(\:\sum\limits_{q=1}^{100}{H3}_{q}{\:W}_{100x9}\:+\:ob\right)$$

(20)

where $\:O$ represents the system output of size 1 × 9; $\:ob$ is the bias value; and W is a random generated weight matrix; and $\:f$ is the Levenberg-Marquardt transfer function.

Safety clutch calibration

The clutch was calibrated to disengage at 50 kW, determined by

$$\:P=\frac{2\pi\:NT}{60}$$

Where, P = power (kW), N = rotational speed (rpm), T = torque (N·m).

The threshold torque was set experimentally by incrementally increasing load until disengagement occurred consistently without inducing shaft vibration or slippage. This ensured safe operation in the event of stalled cane, lodged stalks, or excessive biomass density.

Cost analysis of the developed SWSH

The cost analysis of the developed SWSH was performed during 2017. The American Society of Agricultural and Biological Engineers (ASABE) provides standardized formulas to calculate machinery costs, encompassing both fixed (ownership) and variable (operating) expenses. These formulas consider factors such as purchase price, salvage value, economic life, annual usage hours, fuel consumption, maintenance, and labor. Table 2 shows the price list of different parts of the developed SWSH. Where the total purchase price of machine parts was about 363 USD.

Table 2 Price list of different parts of the developed SWSH.

Results and discussions

The effect of study parameters on the machine’s cutting efficiency

The cutting efficiency of the developed SWSH was evaluated at four forward speeds (3, 3.5, 4.5, and 5 km/h), three row spacings (71 cm, 78.89 cm, and 88.75 cm), three cutting heights (0 cm (ground level), 2 cm, and 4 cm), and two groups of knife numbers (2 and 4), and the obtained data was plotted in Fig. 5. The plotted data showed that, at a cutting height equal to zero, row spacing of 71 and 78.89 cm, and using either two or four knives, the cutting efficiency was optimum and equal to 100% at different forward speeds. While when the forward speed was about 5 km/h at a row spacing of 78.89 cm, the cutting power exceeded the safe limit, and the power clutch cut off the power transmission, and the cutting knives stopped to protect the power transmission system from failure. At this speed the cutting efficiency was equal to zero because the cutting was stopped. On the other hand, increasing the cutting height to 2 cm led to a decrease in the cutting efficiency at different variables. Where the cutting efficiency ranged between 95 and 98%, 96–99%, and 98–100% at raw spacing of 71 cm, 78.89 cm, and 88.75 cm, respectively, when the cutting system operated with two knives, while it ranged between 97 and 100%, 97–100%, and 98–100% at raw spacing of 71 cm, 78.89 cm, and 88.75 cm, respectively, when the cutting system operated with four knives. Furthermore, increasing the cutting height up to 40 cm above the soil surface led to a significant decrease in the cutting efficiency. Where the cutting efficiency ranged between 93 and 97%, 0–97%, and 0–98% at raw spacing of 71 cm, 78.89 cm, and 88.75 cm, respectively, when the cutting system operated with two knives, while it ranged between 95 and 98%, 0–99%, and 0-100% at raw spacing of 71 cm, 78.89 cm, and 88.75 cm, respectively, when the cutting system operated with four knives. So, the cutting efficiency increased with increasing the row spacing; the average stalk diameters at different row spacings were 2.63 cm, 3.12 cm, and 3.76 cm, respectively.

The cutting efficiency of sugarcane harvesters typically increases with larger stalk diameters due to several mechanical and operational factors: (1) improved structural stability during cutting: thicker stalks are more rigid and less prone to bending or lodging before being cut. this ensures better alignment with the harvester’s base cutter blades, leading to cleaner cuts. While thin stalks may flex or slip away from the blades, requiring re-cutting or resulting in incomplete cuts^44,45; (2) optimized blade engagement: thicker stalks provide a larger cutting surface area, allowing the blades to engage more effectively without excessive blade slippage. While thin stalks may require higher blade speed or force per unit area, increasing energy waste and wear^46,47; (3) lower cutting frequency for the same cane volume: a field with thicker stalks has fewer individual stalks per unit area compared to thin stalks for the same biomass. This reduces the number of cuts required, allowing the harvester to process more sugarcane with fewer cutting cycles, improving efficiency^48,49; (4) reduced energy loss due to vibration & deflection: thicker stalks absorb and transfer cutting force more efficiently, minimizing energy loss from stalk deflection or blade bounce. While thin stalks may vibrate or bend under cutting force, leading to incomplete cuts and higher energy consumption^50,51,52.

According to the forward speed, the presented data in Fig. 6 shows that the optimum cutting efficiency of the developed SWSH was observed when the cutting height was zero at different forward speeds using both cutting systems. But at different cutting heights and raw spacings the cutting efficiency was decreased with increasing forward speed, where the highest cutting efficiency was observed at slow travel speed (3 km/h), and the lowest cutting efficiency was observed at slow travel speed (5 km/h). This drawback may be due to several interrelated mechanical and operational factors as stated in previous studies, (1) Reduced cutting time per stalk: At higher speeds, the base cutter spends less time in contact with each stalk, leading to incomplete or uneven cuts. This can result in ragged cuts, stalk splitting, or missed stalks^53,54; (2) Increased load on knives: Faster forward motion raises the instantaneous cutting force required, potentially overwhelming the blade’s capacity. This can cause slippage, excessive wear, or even blade stalling, especially in dense cane fields^54,55; (3) Inertia and Vibration Effects: Higher speeds amplify vibrations in the cutting mechanism, reducing precision. The base cutter may struggle to maintain optimal alignment with the stalks, leading to inefficient cuts^55,56; (4) Power Limitations: Harvesters have finite engine power. Increasing speed diverts more power to propulsion, leaving insufficient power for the cutting mechanism to maintain optimal rotational speed under heavier loads^57,58; (5) Field Conditions Interaction: At higher speeds, terrain irregularities (e.g., uneven ground, stool damage) exacerbate machine instability, further reducing cutting accuracy. Additionally, taller or lodged cane may escape proper cutting^59,60; (6) Material Flow Congestion: Faster operation can overwhelm the internal conveyor systems, causing cane accumulation at the base cutter. This “choking” effect forces the cutter to process multiple stalks simultaneously, reducing efficiency^54,60.

Fig. 6

Machine’s cutting efficiency at different levels of cutting heights, number knives, and cutting forward speeds.

As shown in Fig. 6, the cutting efficiency was not affected by the number of knives when the cutting height equaled zero at different forward speeds. While increasing the number of knives from 2 to 4 knives led to an increase in the cutting efficiency of 2 cm and 4 cm cutting knives at different forward speeds. Where the cutting efficiency of a sugarcane harvester increases with the number of knives on the cutting disk due to several key mechanical and operational advantages, as stated in previous studies: (1) Higher cutting frequency (more cuts per rotation): Each additional knife increases the number of cuts per disk revolution. This reduces the time between successive cuts, allowing the harvester to process more stalks per second, especially at high forward speeds^61,62; (2) Reduced load per knife (better force distribution): With more knives, the cutting force is distributed across multiple blades rather than concentrated on a few. This minimizes blade stress, wear, and power spikes, leading to smoother operation and longer blade life^63,64; (3) Lower stalk slippage & incomplete cuts: Fewer knives mean longer intervals between cuts, increasing the chance that stalks slip away before being fully severed. More knives ensure continuous engagement, reducing partial cuts and recuts^54,65; (4) Improved cutting continuity (reduced vibration & jerking): A higher knife count provides a more continuous cutting action, reducing torque fluctuations and mechanical vibrations. This results in cleaner cuts and less damage to the stool (root system), improving ratoon regrowth^55,66. So, increasing the number of knives on a sugarcane harvester’s cutting disk improves efficiency by distributing cutting forces, reducing slippage, and enabling smoother, faster stalk severance. However, there is an upper limit where additional blades may cause diminishing returns due to increased drag or power requirements.

The effect of study parameters on the total operational costs

The total operating costs of the developed SWSH were evaluated at four forward speeds (3, 3.5, 4.5, and 5 km/h), three row spacings (71 cm, 78.89 cm, and 88.75 cm), three cutting heights (0 cm (ground level), 2 cm, and 4 cm), and two groups of knife numbers (2 and 4), and the obtained data was plotted in Fig. 7. The presented results in the same figure showed that:

Increasing forward speed (3 → 5 km/h) → higher costs; for example, the total operation cost of the developed SWSH decreased from 5.71 USD/ha to 4.53 USD/ha when the forward speed increased from 3 to 5 km/h at a raw spacing of 71 cm, a cutting height of zero, and using two cutting knives. Additionally, maximum operation cost was observed at a forward speed of 5 km/h for all experiments. Because higher forward speeds require more engine power and more fuel consumption for both propulsion and cutting, increasing diesel usage. The highest fuel consumption rate was 15.61 l per hour, and it was consumed when using four cutting knives, a travel speed of 5 km/h, a row spacing of 88.75 cm, and a cutting height of 4 cm. Additionally, because the forward speed of the tractor and the cutting system were integrated with each other. Faster operation accelerates wear on knives, gearboxes, and hydraulic systems, raising maintenance costs. Furthermore, higher speeds amplify mechanical vibrations, increasing fatigue on structural components such as cutting knives, which can lead to the fracturing of knives and increase repair and maintenance costs.

Narrower row spacing (88.75 cm → 71 cm) → higher costs; for example, the total operation cost of the developed SWSH increased from 7.9 to 5.71 USD/ha when the row spacing decreased from 88.75 to 71 cm at a forward speed of 5 km/h, a cutting height of zero, and using two cutting knives. Where increased crop density (tighter spacing) means more stalks per ha, requiring more cuts and higher energy input. Also, there is blockage & clogging risk, where narrow rows can cause stalk jamming in the feed mechanism, leading to downtime and repairs. Additionally, tight spacing may force slower turns or more passes, increasing operational time and fuel use.

Lower cutting height (4 cm → 0 cm (ground level)) → Higher costs, for example, the total operation cost of the developed SWSH increased from 5.71 to 5.5 USD/ha when the cutting height decreased from 4 to 0 cm at a forward speed of 3 km/h, a row spacing of 78.89 cm, and using two cutting knives. Where cutting closer to the ground increases blade contact with dirt and rocks, accelerating wear. Also, cutting through tougher, lower stalk sections (near the base) requires more torque and fuel. Furthermore, harvesting near ground level picks up more soil and trash, raising post-harvest cleaning costs.

Higher knife number (2 → 4 knives) → mixed impact; for example, the total operation cost of the developed SWSH increased from 5.71 USD/ha to 5.8 USD/ha when the knife number increased from 2 to 4 knives at a forward speed of 3 km/h, a cutting height of zero, and a row spacing of 71 cm. Increasing the number of knives can lead to an increase in capital and replacement expenses. Furthermore, if the harvester operates at high speeds, additional knives may cause excessive drag, raising fuel consumption. More knives also mean more cycles of sharpening and replacement, though distributed wear may make them last longer. In some cases, 4 knives reduce peak load per cut, improving fuel efficiency at moderate speeds.

Comprehensive field data were gathered from sugarcane farms across four key locations in Upper Egypt: Aswan, Luxor, and Minya. These regions are pivotal to Egypt’s sugarcane production, collectively contributing a significant portion of the national yield. In traditional manual harvesting systems, sugarcane fields with an average yield of approximately 45 tons per feddan (about 0.42 hectares) necessitate the deployment of around 16 laborers for the base cutting of stalks. This labor-intensive process is critical for ensuring the quality and efficiency of the harvest. Labor wages exhibit variability influenced by factors such as geographic location and the specific timing within the harvesting season. For instance, during the peak harvest months from December to May, labor demand surges, potentially impacting wage rates. In Qena, a prominent sugarcane-producing governorate, farm workers typically engage in harvesting activities from early morning until noon, earning approximately 2 USD per day for an eight-hour shift. This means that the cost of manual harvesting of sugarcane was about 76.19 USD per ha. The manual harvesting process is not only labor-intensive but also physically demanding, often conducted under high-temperature conditions. Workers are exposed to risks such as dehydration and heat stress, underscoring the importance of adequate hydration and rest periods.

Fig. 7

The total operating costs at different levels of cutting heights, number knives, and cutting forward speeds.

The effect of study parameters on the field capacity

The field capacity of the developed SWSH was tested at four speeds (3, 3.5, 4.5, and 5 km/h), three distances between rows (71 cm, 78.89 cm, and 88.75 cm), three heights for cutting (0 cm (ground level), 2 cm, and 4 cm), and two different numbers of knives (2 and 4), and the results were shown in Fig. 8.

The field capacity (ha/h) of the developed SWSH is influenced by forward speed, row spacing, cutting height, and knife number, as these factors determine how much area the machine can cover per hour. The presented results in the same figure showed that Increasing forward speed (3 → 5 km/h) → increases field capacity. Where the obtained results showed a direct relationship between the forward speed and field capacity, as shown in Eq. (1), the highest field capacities were observed at a forward speed of 5 km/h and covered more area than at 3 km/h. Additionally, increasing row spacing (71 cm → 88.75 cm) → mixed impact: Wider row spacing leads to achieving higher field capacity. The highest field capacities were observed at a raw spacing of 88.75 cm, but, as shown above, when the forward speed was about 5 km/h and the raw spacing was about 88.75 cm (stalk diameter was about 3.76), the cutting power exceeded the safe limit, and the power clutch cut off the power transmission, and the cutting knives stopped to protect the power transmission system from failure. At this speed the cutting efficiency was equal to zero because the cutting was stopped. So, at a raw spacing of 71 cm, the maximum field capacity of the developed SWSH was 0.487 ha/h, observed at a forward speed of 5 km/h. Additionally, at a row spacing of 78.89 cm, the maximum field capacity of the developed SWSH was 0.495 ha/h, observed at a forward speed of 4.5 km/h. Furthermore, at a row spacing of 88.75 cm, the maximum field capacity of the developed SWSH was 0.554 ha/h, observed at a forward speed of 4.5 km/h. On the other hand, the plotted results in Fig. 7 showed a slight impact of both knife number and cutting height on the field capacity.

Fig. 8

Field capacity at different levels of cutting heights, number knives, and cutting forward speeds.

Training and validation loss across epochs with early stopping in the deep neural network model

Figures 9 and 10 illustrate the training progression of the Deep Neural Network (DNN) model, plotting the Mean Squared Error (MSE) across epochs for training (blue curve), validation (red curve), and test datasets. The best validation performance (MSE = 7.9448) occurred at epoch 3, signalling rapid initial convergence. Beyond this point, the validation loss plateaued after approximately 600 epochs (marked by a green dashed line), prompting early stopping to prevent overfitting, while the training loss continued its downward trend. The logarithmic-scale MSE curves highlight stable optimization, with training and validation errors aligning closely after epoch 3, despite the extended training duration. This divergence between persistently decreasing training loss and stagnating validation loss underscores the necessity of early stopping to balance model generalization and memorization of training noise, ultimately demonstrating the DNN’s efficacy in minimizing prediction errors.

Fig. 9

Dual-panel analysis of ML training dynamics: rapid MSE convergence and validation-guided early stopping.

Fig. 10

Dual-panel analysis of ML training dynamics: rapid MSE convergence to mitigate overfitting.

Permutation importance

For sensitivity analysis, the permutation importance technique was applied to observe the effect of each input feature on the predictive model. Permutation importance is a technique used in machine learning to assess the importance of an individual feature in a predictive model. It provides a way to understand which features have the most impact on the model’s performance.

For input data X, the first feature is permuted randomly to break the association between this feature and the true outcome. Then, the loss function (error) is calculated. The process of permutation and loss function calculation is repeated individually for each input feature. At the end, the median error is calculated, where the feature of maximum error refers to the most significant impact on the predicted model.

Figure 11 below presents the permutation importance result for both DNN and FNN predicted models.

Fig. 11

The permutation importance result for both DNN and FNN predicted models.

For DNN model, feature number 3 which is Row Spacing, has the most impact effect, while for FNN model feature number 7, which is Knife Linear Speed, has the most impact effect.

Model performance evaluation

Figure 12 demonstrates the predictive accuracy of the Deep Neural Network (DNN) across training, testing, and validation datasets, quantified by the coefficient of determination (R). The model achieves perfect training accuracy (R = 1), indicating robust learning of input-output relationships. For testing and validation phases, R-values of 0.99142 and 0.99799, respectively, confirm strong generalization to unseen data, with minimal overfitting. The combined dataset (All) yields an R = 0.99829, underscoring the model’s consistency and reliability. The alignment between predicted (Y) and target (T) values across all phases validates the DNN’s efficacy in optimizing sugarcane harvester parameters, making it a trustworthy tool for operational predictions.

Fig. 12

Predictive accuracy of the deep neural network (DNN): coefficient of determination (R) across training, testing, and validation datasets.

Comparative performance evaluation of deep neural network (DNN) and feedforward neural network (FNN) models

Table 3 compares the performance metrics of the two machine learning models, DNN and FNN, in predicting various operational parameters of sugarcane cultivation machinery. The DNN consistently outperforms the FNN across all metrics, achieving higher accuracy percentages and tighter tolerances. For machine performance, the DNN achieves 88% accuracy compared to the FNN’s 68%. Similarly, for parameters like cutter head efficiency, field capacity, throughput capacity, fuel consumption, and power requirements, the DNN demonstrates superior predictive accuracy (ranging from 96% to 100%) with minimal tolerances. In contrast, the FNN exhibits lower accuracy (62% to 76%) and comparable tolerances. Notably, the DNN achieves 100% accuracy in critical metrics, such as fuel consumption and total operating costs, highlighting its robustness in optimizing agricultural machinery operations. These results underscore the DNN’s capability to provide precise predictions for enhancing the mechanized cultivation efficiency of sugarcane.

Table 3 Comparative accuracy metrics of DNN and FNN models for sugarcane harvester operational parameters.

Validation of training-testing data partition

To confirm the randomness and representativeness of the training-testing split, we conducted frequency distribution analysis and Kolmogorov-Smirnov (K-S) tests. Figure 13 compares histograms of all input parameters across the whole dataset, training set (n = 130), and testing set (n = 50). The distributions align closely, with proportional representation of experimental conditions.

Fig. 13

Frequency distribution analysis of input parameters: (A) Whole dataset, (B) training set, (C) testing set. Histograms confirm proportional representation across groups.

Residual analysis for model validation

Comprehensive residual diagnostics (Fig. 14) confirm the statistical unbiasedness of both DNN and FNN models. The residuals exhibit.

Random scatter around zero in predicted vs. residual plots (Fig. 14A, D), indicating homoscedasticity and absence of systematic bias, Near-linear alignment in Q-Q plots (Fig. 14B, E), validating normality of errors, and Symmetric, unimodal distributions in histograms (Fig. 14C, F). The DNN shows tighter residual clustering (± 0.25 SD) compared to FNN (± 0.41 SD), consistent with its superior predictive accuracy in Table 2.

Fig. 14

Residual analysis of DNN and FNN models: (A–C) DNN diagnostics; (D–F) FNN diagnostics. (A, D) Residuals vs. predicted values, (B, E) Q-Q plots, (C, F) Residual histograms with kernel density estimates.

Comparison of neural network performance with literature

The neural network models developed in this study demonstrated exceptional performance in predicting optimal operational conditions for the double-row sugarcane harvester. The FNN and DNN achieved high prediction accuracy, which aligns with and extends findings from similar agricultural machinery optimization studies.

The FNN model’s performance is consistent with the work of Ali et al. ³⁶, who achieved high prediction reliability (R² = 0.99974) using feedforward neural networks for optimizing a compact date seed milling unit. Their study demonstrated the effectiveness of FNN in predicting optimal operating conditions for agricultural processing equipment, including parameters such as cylinder rotational speed, feeding rate, and screen hole diameter. Similarly, this study achieved comparable accuracy in predicting optimal forward speeds, planting distances, cutting heights, and the number of knives for sugarcane harvesting operations.

The application of deep neural networks in agricultural machinery optimization has shown promising results across various studies. Wang et al. ³⁴ achieved high prediction accuracy using geometry-integrated neural networks for metal processing applications, demonstrating the potential of advanced ML architectures in machinery optimization. While their focus was on industrial applications, the methodology and performance metrics provide valuable benchmarks for agricultural machinery applications. The current DNN model’s performance in predicting harvester operational parameters shows similar trends in accuracy and reliability.

The computational efficiency and parameter optimization achieved in this study align with findings from Issa et al. ³⁸, who emphasized that FNN’s simpler architecture provides efficient parameter optimization with lower computational requirements compared to more complex architectures. This is particularly relevant for agricultural machinery applications where real-time optimization and field implementation are crucial considerations.

Recent advances in machine learning applications for manufacturing and processing systems have consistently shown the superiority of neural network approaches over traditional optimization methods. Ma et al. ³⁵ implemented multi-objective approaches for optimizing cooling systems, highlighting the importance of integrated modeling solutions similar to our approach in simultaneously optimizing multiple harvester parameters. Their work demonstrates that neural networks can effectively handle multi-parameter optimization problems, which is directly applicable to the sugarcane harvester optimization challenge.

Furthermore, the application of machine learning in agricultural equipment optimization has shown consistent benefits in terms of operational efficiency and cost reduction. Our findings regarding the optimal operational conditions (forward speed of 4.5 km/h, planting distances of 88.75 cm, cutting height of 4 cm, and two knives) align with the general trend in literature that emphasizes the importance of balanced parameter selection for achieving optimal performance while minimizing operational costs.

Definition of optimal operational conditions

To ensure clarity in evaluating the performance of the developed semiautomatic whole-stalk sugarcane harvester (SWSH), the term “optimal operational conditions” in this study refers to the specific combination of forward speed, row spacing, cutting height, and knife number that simultaneously maximized harvesting efficiency while minimizing operational costs and mechanical loading. Optimality was quantitatively determined using five primary performance indicators: (i) cutting efficiency (%), (ii) total operating cost (USD/ha), (iii) field capacity (ha/h), (iv) fuel consumption (L/h), and (v) power requirements (kW). Conditions were considered optimal when they achieved: (1) maximum cutting efficiency (≥ 98–100%) with no uncut stalks; (2) minimum total operational cost, reflecting the combined effects of fuel consumption, labor, depreciation, and maintenance based on ASABE standards; (3) maximum achievable field capacity that did not trigger clutch disengagement or cutting system overload; (4) moderate fuel consumption that avoided excessive load on the tractor engine; and (5) power demand below the safe operating limit of 50 kW, ensuring stable mechanical operation. These criteria were applied to all 180 experimental combinations, and the same multi-objective framework was used for both the Feedforward Neural Network (FNN) and Deep Neural Network (DNN) models by normalizing each performance metric and identifying the parameter combination that globally optimized the integrated performance score. The resulting experimentally validated and DNN-predicted optimal conditions corresponded to a forward speed of 4.5 km/h, row spacing of 88.75 cm, cutting height of 4 cm, and two knives on the cutting disk.Practical impact and validation of the deep learning optimization.

The integration of Deep Learning models—particularly the DNN architecture—provided a substantial practical improvement in identifying and refining optimal operating conditions for the developed SWSH compared with traditional empirical tuning. Conventional field-based optimization typically relies on trial-and-error adjustments of forward speed, cutting height, knife configuration, and row spacing, often leading to suboptimal combinations that either increase losses or raise operational cost. In contrast, the DNN model evaluated all 180 experimental combinations simultaneously and quantified their multi-objective performance, enabling rapid identification of the globally optimal settings without repeated field trials.

From a practical standpoint, the improvements achieved through Deep Learning are significant. When comparing the DNN-optimized conditions to the best performing combinations selected empirically during field experiments, the DNN achieved:

• I. Reduction in harvest losses:

  • The DNN model recommended operating conditions that maintained a cutting efficiency of 100%, eliminating the 2–7% losses observed under non-optimized empirical settings, especially at suboptimal cutting heights or forward speeds.

  • This corresponds to a reduction of 2–7% in total stalk losses, which is highly meaningful given that each 1% loss in Egyptian sugarcane represents thousands of tons annually at the national scale.

• II. Increase in field capacity:

  • Traditional empirical selection often recommended lower forward speeds (3–3.5 km/h) to avoid cutter overload, resulting in field capacities of 0.40–0.48 ha/h.

  • The DNN model identified that 4.5 km/h can be safely used at 88.75 cm row spacing, increasing field capacity to 0.554 ha/h, representing an improvement of 15–22% in effective harvesting rate.

• III. Reduction in total operating costs:

  • Empirical tuning resulted in operating costs ranging between 5.3 and 7.9 USD/ha depending on spacing and cutting height.

  • The DNN-optimized parameters produced the lowest recorded cost of 4.42 USD/ha, representing a cost reduction of 17–44% compared with commonly used empirical settings.

• IV. Fuel savings and lower power load:

  • At high forward speeds, empirical trials frequently triggered power clutch disengagement or excessive fuel consumption.

  • The DNN model identified the safest high-efficiency zone where power remained below the critical 50 kW threshold, reducing fuel use by 0.8–1.6 l/h relative to non-optimized combinations.

• V. Elimination of unsafe or unstable operating modes:

  • Empirical trials at 5 km/h and narrow spacing sometimes resulted in complete cutting system shutdown (0% cutting efficiency) due to overload.

  • The DNN model successfully avoided all unsafe combinations, ensuring stable operation and reducing downtime.

Overall, the Deep Learning-based optimization not only improved numerical performance but also enhanced machine reliability, operational safety, and economic viability. The quantified benefits—up to 7% reduction in losses, 22% increase in field capacity, and 44% reduction in operating cost—demonstrate that intelligent predictive modeling represents a transformative tool for mechanized sugarcane harvesting, particularly under the constraints of smallholder Egyptian farming systems. This confirms that the DNN approach offers practical, field-relevant advantages over conventional empirical methods and can serve as a decision-support system for farmers, manufacturers, and machinery operators.

Comparison with commercial harvesters and local prototypes, and practical implications

In light of the challenges outlined in the Introduction, the performance of the developed semiautomatic whole-stalk sugarcane harvester (SWSH) demonstrates notable advantages when compared with both commercial harvesters and previously developed Egyptian prototypes. Imported full-scale harvesters—designed for large, uniform fields in countries such as Brazil, Australia, and South Africa—have consistently failed under Egyptian conditions due to their high cost, poor maneuverability, and inability to operate efficiently within the narrow row spacing and irregular field geometry typical of smallholder farms. Their complex de-trashing and base-cutting mechanisms also underperform on Egyptian cane varieties, which require higher cutting forces and produce substantial trash loads, contributing to low efficiency and frequent clogging. Earlier locally fabricated prototypes developed between 2002 and 2014 achieved incremental improvements in power, maneuverability, and field capacity, yet still struggled with lodged cane, high labor demand, low reliability, and limited cost-effectiveness—factors that prevented wide adoption despite technical promise.

The developed SWSH addresses many of these limitations through a simpler mechanical architecture, reduced power requirements, and a front-mounted configuration that enhances stability, visibility, and maneuverability in small, fragmented fields. When evaluated against commercial benchmarks, the SWSH offers a more appropriate match for Egyptian conditions by enabling double-row harvesting, minimizing operational losses, and reducing the physical strain associated with manual cutting. Additionally, the integration of FNN and DNN predictive models introduces a unique capability not found in previous prototypes or commercial units: data-driven optimization of operating parameters such as speed, cutting height, and knife configuration. This allows operators to achieve optimal performance despite field variability—a critical need in Egypt’s heterogeneous farming environment.

From a practical standpoint, the simplified design significantly lowers maintenance demands, allowing farmers to service the machine using commonly available parts and basic workshop tools. The modularity of key subsystems supports scalability, enabling future increases in cutting width or upgrades to semi-automation without requiring a complete redesign. Furthermore, adopting this design within local manufacturing facilities can overcome earlier challenges related to machining precision and heat-treatment capacity by focusing on fewer critical components with standardized fabrication requirements. These improvements enhance the feasibility of widespread adoption and reinforce the potential of the SWSH as a scalable, cost-effective, and locally appropriate solution for modernizing sugarcane harvesting in Egypt.

Conclusion and future work

This study developed and evaluated a semiautomatic whole-stalk sugarcane harvester (SWSH) designed specifically for the constraints of small- and medium-scale Egyptian farms. The integration of engineering design and data-driven analysis enabled the identification of a clear optimal operational window, defined by a forward speed of 4.5 km/h, row spacing of 88.75 cm, cutting height of 4 cm, and two cutting knives. This combination simultaneously achieved the dual objectives of minimizing operational cost—with a recorded minimum of 4.42 USD/ha—and maximizing field capacity, reaching 0.554 ha/h, while maintaining a cutting efficiency of up to 100% under appropriate spacing. These findings demonstrate that the developed SWSH offers a highly efficient and economically viable alternative to manual harvesting or incompatible imported machinery. The adoption of Deep Learning tools proved essential for optimizing the machine’s performance. The DNN model demonstrated superior predictive accuracy, enabling reliable multi-output estimation of cutting efficiency, throughput capacity, fuel consumption, and operating cost across all tested conditions. This confirms the value of AI-driven modeling as a powerful decision-support tool for operational tuning of agricultural machinery. These findings reflect the performance of the prototype under the specific field conditions, crop characteristics, and machine configuration tested in Egypt. Therefore, the recommended operating settings are most applicable to similar local agro-mechanical environments and should not be generalized beyond these conditions without further validation on additional cane varieties, soil types, and machine sizes.

Looking forward, future work should advance the system toward greater intelligence and autonomy. Integrating the trained DNN model into a real-time, sensor-based adaptive control system could allow dynamic adjustment of operational parameters under variable field conditions. Developing predictive maintenance frameworks using vibration, acoustic, or thermal sensors would enhance reliability and reduce downtime. Additionally, expanding model validation across diverse cane varieties, soil types, and topographies will strengthen generalizability. Collectively, these developments could pave the way toward next-generation, smart, and fully adaptive sugarcane harvesting systems.