Journal of Agricultural Machinery

Abstract This research focused on creating an IoT-enabled color-sorting machine for Robusta coffee cherries, utilizing image processing as an effective alternative to manual sorting. The system tackles a significant issue with the strip-picking harvesting method, which gathers cherries at different ripeness levels, negatively affecting coffee quality. The machine sorts cherries by ripeness—red for ripe, green for unripe, and black for overripe—using a detection model trained through image processing and implemented on a Raspberry Pi 4 Model B. The performance was assessed based on sorting speed and classification accuracy. The detection model successfully identified 277 out of 300 cherries, resulting in an overall classification accuracy of 92.33% and a mean precision of 92.55%. In practical tests with 100 cherries over 10 trials, the machine achieved an average sorting accuracy of 86.83% and a mean sorting time of 21 minutes and 33 seconds. When compared to a previously developed coffee bean sorter, the new device showed improved accuracy and faster processing speed.

Abstract Introduction
Paying attention to the technical aspects of production plays a crucial role in increasing yield and ensuring sustainable agriculture. Organic fertilizers, such as poultry manure, contribute to plant growth by providing essential nutrients and improving soil quality. However, they alone cannot fully meet the nutritional needs of plants. The combination of organic and chemical fertilizers is an effective approach to enhancing soil fertility and boosting crop performance, ultimately leading to sustainable agricultural development. Integrated nutrient management also helps reduce the use of chemical fertilizers while minimizing their harmful effects on the environment. Potassium is an essential element in plant nutrition, playing a key role in processes such as photosynthesis, growth, chlorophyll production, and transpiration regulation. Additionally, under stress conditions, potassium enhances water uptake and regulates osmotic pressure, helping to maintain plant health. Potassium fertilizers are classified into two categories: chloride-based and chloride-free. Potassium sulfate, due to its lack of chloride, is a suitable option for chloride-sensitive crops such as tea, potatoes, and sugar beets. Meanwhile, hyperspectral imaging has emerged as an innovative technique with broad applications in detecting chemical parameters, assessing quality, and analyzing the purity of agricultural and food products. This study utilizes hyperspectral image processing technology to determine the pH level of potassium sulfate.
Materials and Methods
The present study was conducted in the Image Processing Laboratory at the Ilam University, Iran. To determine the pH level of potassium sulfate, four different levels of 2.5, 2.6, 2.8, and 2.9 were considered. The pH measurement was performed in the laboratory using a flame photometer. The required images were obtained through hyperspectral imaging using the line-scan method. For each pH level, three samples were obtained and six hyperspectral images were captured for each sample, resulting in 18 images per pH level and a total of 72 hyperspectral images for each pH level. MATLAB software was used for the analysis and processing of these images. The image processing stage included wavelength selection, feature extraction, and feature selection. Finally, the selected features were classified using an artificial neural network.
Results and Discussion
Principal Component Analysis performed on the hyperspectral image channels of potassium sulfate revealed significant variations in the principal component values across different pH levels. This finding indicates that pH conditions exert a considerable influence on the spectral response of the samples. Based on the prominent peaks obtained from the analysis, the most relevant channels were identified, and their corresponding wavelengths were determined as the optimal spectral bands. The selected channels for the four pH levels were 65, 327, 334, 482, 510, 607, and 644, with their corresponding effective wavelengths being 453.32, 669.95, 675.74, 798.11, 821.26, 901.47, and 932.06 nm, respectively. To extract discriminative spectral information, six features were computed from each of the selected wavelengths. Consequently, a total of 42 features were obtained, which were subsequently employed in the classification process of different pH levels. The confusion matrices of the classification model based on the artificial neural network were obtained to evaluate the model's accuracy. The classification accuracy for detecting the pH level of potassium sulfate was 98.6% with effective features and 97.2% without them.
Conclusion
The results of this study demonstrated the high potential of hyperspectral imaging technology combined with the artificial neural network classification method, using strategies with and without effective feature selection, in detecting the pH level of potassium sulfate. The proposed method offers several advantages over laboratory-based approaches, such as being non-destructive, having high speed, and being cost-effective. It is suggested to explore other methods for classifying hyperspectral images for determining the pH level of potassium sulfate. The proposed method in this study could also be applied in the future to identify various chemical elements in potassium sulfate.

Abstract Introduction
Traditional methods for evaluating fruit quality, such as pH measurement, are often destructive, time-consuming, and costly, leading to product loss and reduced efficiency in the supply chain. The growing need for rapid, accurate, and non-destructive methods makes the use of technologies like Hyperspectral Imaging (HSI) essential. HSI combines two-dimensional imaging with spectroscopy to simultaneously acquire spatial and spectral information from an object. Numerous studies have shown that this method is capable of accurately estimating internal fruit parameters in a non-destructive manner. The objective of this research was to develop a fast and reliable method for the non-destructive estimation of pH in two plum cultivars using HSI and machine learning algorithms such as Partial Least Squares Regression (PLSR) and Artificial Neural Networks (ANN). This study aims to overcome the limitations of conventional methods by leveraging the power of advanced imaging and computational techniques, providing a sustainable and efficient solution for the fruit industry.
Materials and Methods
In this study, 80 samples from each of the Khormaei and Khoni plum cultivars were used, which were purchased from local orchards. The samples were uniform in size, shape, and colour and were free from any physical damage. Hyperspectral images of the samples were acquired using a rotating hyperspectral imaging system in the range of 418 to 1072 nm. The pH of each fruit juice sample was measured using a digital pH meter. In the analysis of spectral data, the initial part of the spectrum was first removed due to high noise, and then the remaining data were processed with preprocessing methods such as a Gaussian filter and Multiplicative Scatter Correction (MSC). To select effective wavelengths (EWs), a hybrid approach using a Decision Tree (DT) and five metaheuristic algorithms was employed, with the Particle Swarm Optimisation (PSO) algorithm showing the best performance. Finally, pH modelling was performed on the selected wavelengths using PLSR and ANN. This comprehensive methodology ensures that the models are trained on high-quality data and are optimised for maximum accuracy.
Results and Discussion
Spectral analysis showed that the reflectance spectra of the Khoni and Khormaei plums had a high degree of variation, which is related to the differences in their chemical composition and structure. Descriptive statistics indicated that the average pH of Khormaei plum (3.909) was higher than that of Khoni plum (3.7375), and the pH range of Khoni plum (3.15 to 4.44) was wider than that of Khormaei plum (3.6 to 4.2). The results showed that modelling with ANN on the wavelengths selected by PSO, especially for Khoni plum, significantly increased prediction accuracy. The best ANN model for Khoni plum achieved an R²​ of 0.9834 and an RPD of 8.01, which indicates the outstanding accuracy of this method. For the Khormaei plum, the best ANN model also reached an R²​ of approximately 0.76 and a Ratio of Performance to Deviation (RPD) of 2.12, showing a considerable improvement over the PLSR model. The superior performance of the ANN models can be attributed to their ability to capture complex, non-linear relationships between spectral data and pH values, which linear models like PLSR may miss.
Conclusion
This research successfully demonstrated that hyperspectral imaging, in combination with machine learning algorithms, particularly ANN and PSO, can be an accurate and reliable method for the non-destructive prediction of pH in different plum cultivars. The hybrid approach used in this study, which combined DT for initial feature selection with PSO for optimal wavelength selection, enabled the models to predict pH values with very high accuracy, especially for the Khoni plum cultivar. This method can be used as an efficient tool in post-harvest quality control processes, helping to reduce waste and improve efficiency in the fruit supply chain. This work paves the way for the development of smart grading and sorting systems that can quickly and accurately assess fruit quality, benefiting both producers and consumers.

Abstract This study assesses the capability of Sentinel-2A imagery combined with machine-learning classifiers on the Google Earth Engine platform for accurate orchard mapping in the environmentally stressed Sharviran Plain along the southern margin of Lake Urmia, Iran. Random Forest, Support Vector Machine, and Classification and Regression Trees were applied to multispectral bands and selected spectral and vegetation indices, using 836 systematically collected samples from high-resolution imagery and field surveys for training and validation. The study specifically focuses on orchard mapping, with particular emphasis on reliably distinguishing orchards from other agricultural and non-agricultural land-use classes. Results show that orchards cover approximately 11% of the study area, and that the Random Forest classifier achieved the highest performance (Kappa = 0.84) and the best orchard F1-score, demonstrating superior class-level discrimination. Validation against 2023 ground-based orchard statistics showed a low error margin of 2.2%, confirming the reliability of the approach for operational orchard mapping. These findings confirm that integrating Sentinel-2A imagery with machine-learning classifiers on cloud-based platforms offers a robust, reproducible, and cost-efficient framework for orchard mapping in data-scarce and drought-affected regions. The resulting high-resolution orchard maps can effectively support agricultural planning, irrigation management, and evidence-based policy-making in the Lake Urmia basin. Future work may further improve performance by incorporating multi-temporal imagery, additional ecological and socio-economic variables, and advanced models such as deep learning.

Abstract The red palm weevil (Rhynchophorus ferrugineus) is a destructive pest of date palms whose concealed larval feeding makes early detection extremely difficult. To capture weak chewing signals under natural orchard conditions, a portable bioacoustic sensing unit was developed using a MAX9814 electret microphone integrated with an STM32F103C8T6 12-bit data acquisition module operating at 16 kHz. The compact design, isolated power, and vibration-damping mount enabled reliable, non-invasive recording in field conditions. Acoustic emissions from infested palms were analysed in both time and frequency domains using six key spectral features: mean and median frequency, band power, occupied and power bandwidth, and peak location, across five frame durations. Statistical analysis revealed how segmentation scale affects signal stability, impulsiveness, and redundancy among features. A structured spectral atlas was then established, consolidating scattered acoustic descriptors into a unified representation. Dominant larval activity occurred below 7 kHz, with mean and median frequencies between 2.1 and 3.8 kHz (r > 0.85), and Band power values ranged from approximately 1.7×10⁻⁵ to 2.6×10⁻⁵ (relative linear scale) and increased during periods of intense chewing. The occupied bandwidth (1.5 to 2.8 kHz) narrowed during intense chewing, confirming spectral consistency and diagnostic value. Unlike prior studies that primarily report classification performance, this work introduces a structured spectral atlas that quantitatively characterises the stability, variability, and interrelationships of key acoustic features across multiple time scales. This descriptive baseline provides a foundation for informed feature selection, sensor bandwidth design, and the development of future embedded, non-invasive acoustic systems for early RPW infestation monitoring.

Abstract Accurate pre-harvest yield estimation is essential in precision horticulture, enabling informed decisions in labour allocation, logistics, and market forecasting. This study presents an integrated computer vision framework based on the YOLOv12 architecture for accurate detection and quantification of persimmon (Diospyros kaki) fruits under real orchard conditions. Multi-scale feature extraction was combined with lightweight regression modelling to predict individual fruit weights directly from RGB imagery. A dedicated dataset of 2,118 annotated images was compiled to capture variations in canopy density and illumination, ensuring robust performance under natural occlusions. After fine-tuning five YOLOv12 variants, detection precision exceeded 0.92 mAP@0.5, with the YOLOv12x configuration achieving the highest accuracy of 0.945. However, the YOLOv12n variant provided the best trade-off between accuracy (0.925 mAP@0.5) and efficiency (5.7 ms inference, >100 FPS), making it optimal for IoT-based real-time deployment in smart orchard environments. Yield estimation was derived by mapping ellipsoidal volume approximations to actual fruit weights using linear regression, achieving a mean absolute error of 6.2 g per fruit and an average tree-level deviation below 5%. The framework maintained real-time inference speeds on edge devices, confirming its suitability for practical field applications. The results demonstrate its reliability for data-driven yield forecasting and highlight its potential integration into IoT-enabled harvesting systems to minimise post-harvest losses through predictive resource allocation.

Abstract Reliable cow detection in barn environments remains challenging due to visual occlusion, overlapping animals, and cluttered backgrounds. To improve accuracy and real-time performance under such conditions, this study presents YOLO11-DML, an enhanced detection framework built upon YOLO11. To address the decline in recognition accuracy caused by cow overlaps, occlusions, and background interference, this paper has implemented optimisations in three aspects: feature extraction, attention mechanism, and lightweight detection head design: The C3K2-DIMB module is introduced into the Backbone and Neck, enhancing multi-scale feature modelling capabilities through dynamic convolution kernel weights; The mixed local channel attention (MLCA) hybrid local-channel attention mechanism is embedded at the end of the Backbone as an auxiliary feature enhancement module to refine hierarchical feature representations; A lightweight shared convolutional detection head (LSCDH) is designed for the detection head, effectively reducing parameter count and computational overhead while maintaining detection accuracy. Experiments conducted on the CBVD-5 and Dairy Cow datasets show that YOLO11-DML achieves a precision (P) of 92.57%, an F1-score of 89.07%, an mAP@0.5 of 93.11%, an mAP@0.5:0.95 of 59.37%, an inference speed of 105.79 FPS, a parameter count of 2.16M, 5.1 GFLOPs of floating-point operations, and a model size of only 4.5 MB. The research demonstrates that YOLO11-DML achieves high precision while maintaining a lightweight design and real-time performance, providing a feasible solution for multi-object dairy cow detection and behaviour monitoring in smart farms.

Abstract Accurate and timely detection of maize (Zea mays L.) foliar diseases is essential for improving crop productivity and ensuring food security. This study evaluates the performance of state-of-the-art deep learning models for automated maize leaf disease classification using a real-field dataset collected from agricultural farms in Kandahar Province, Afghanistan. The dataset was acquired under natural field conditions, preserving complex backgrounds, variable illumination, and diverse leaf orientations to reflect practical farming environments. Multiple advanced transformer-based deep learning architectures, including Vision Transformer (ViT), and hybrid Multi-axis Vision Transformer (MaxViT-tiny, MaxViT-small, and MaxViT-base) were investigated and compared with EfficientNet (B5, B6, and B7) models. Experimental results demonstrate that modern CNN-based and transformer-based architectures achieve up to 97% accuracy, despite challenging real-world conditions. Among the evaluated models, EfficientNet-B7, ViT, and MaxViT-Tiny attained the highest accuracy of 97%, with MaxViT-Tiny recording the highest precision (0.9741), indicating exceptional class discrimination, particularly for challenging Spot and Blight classes under real-field variability. Minor misclassifications were primarily observed between visually similar disease classes, while overall precision and recall remained consistently high. The findings confirm the effectiveness of hybrid and attention-based architectures for real-world maize disease detection and demonstrate their feasibility for real-world agricultural applications. By leveraging authentic field data, this study provides practical and ecologically valid insights that support early disease diagnosis, reduced manual inspection, and improved sustainable crop management.

Abstract Improving wheat productivity is essential for maintaining food security in the coming years. A major challenge to wheat production is the presence of weeds, which can greatly reduce its yield. In this context, precision weed management based on weed-crop classification can be a key factor for achieving sustainable wheat production. To accomplish this goal, in the present study, we utilised colour images and deep learning techniques to distinguish winter wheat from four prevalent weeds, creating five distinct classes. Four deep learning networks–Xception, EfficientNetB0, VGG19, and InceptionResNetV2–were evaluated to serve as backbone networks to fulfil the classification requirements. These networks were pre-trained on ImageNet using transfer learning, then enhanced with additional layers to improve performance on our dataset. The improved InceptionResNetV2 model demonstrated the highest performance among the four models, achieving an accuracy of 98.17% and a loss of 3.19% on the test data. Nevertheless, all models demonstrated excellent performance in distinguishing plant classes, achieving weighted average F1-scores of 97%, 86%, 94%, and 98% for the improved models based on Xception, EfficientNetB0, VGG19, and InceptionResNetV2, respectively. Additionally, we analysed fifteen scenarios of weed presence in winter wheat fields, focusing on various weed types studied, to propose effective weed management strategies utilising the four models. The research findings provide a foundation for developing mobile apps or robotic weeders that facilitate site-specific weed management. This approach aims to reduce herbicide use and environmental impact while also improving wheat yield and quality.

Abstract The Dubas bug (Ommatissus lybicus) poses a significant threat to agriculture in the Middle East by weakening palm trees and reducing fruit production. Effective pest control depends on accurate and timely localization of the infestation. However, regular field inspections are difficult and time-consuming, especially for large areas. This research investigates the potential of Sentinel-2 satellite imagery for detecting Dubas bug infestations. The aim is to improve monitoring capabilities, accelerate intervention strategies, and mitigate the associated economic impact. The field trial to assess the infestation occurred in May 2023, coinciding with the peak of the pest outbreak. The severity of the infestation was assessed through pest counts conducted in date palm groves within the urban area of Bam, Iran. Sentinel-2 multispectral images of a specific area were acquired and processed for correction, raw data preparation, and information extraction. The Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) method was used for the atmospheric correction of the acquired images. The Nearest Neighbor Interpolation method was used to resample satellite images, standardizing all bands to a uniform 10-meter resolution. Following the pre-processing phase, the KD-tree-based K-Nearest Neighbor classifier model was selected to develop a model specifically designed for identifying areas infested by the Dubas bug. For training, 70% of the measured field data were used, including uninfested areas and areas with three levels of infestation from light to heavy, as well as other land features such as buildings, roads, etc. The remaining 30% of the data was utilized to evaluate the trained model, using the correct prediction rate as the assessment criterion. The trained classifier, validated against the ground truth data, achieved an accuracy of approximately 83% on the test dataset. This accuracy highlights the ability of Sentinel-2 multispectral imagery and machine learning to detect Dubas bug infestations in date palm groves and can facilitate targeted and sustainable pest management strategies.

Abstract In this research, the amount of vitamin C, aromatic compounds, and color change of orange powder was measured using chemical methods, an olfactory machine, and a scanner in four dryers at 45℃. These dryer apparatuses included normal atmospheric vacuum, atmospheric control vacuum, convective, and convective-infrared. The highest response of sensors to aromatic compounds in convective and lowest response in control vacuum and normal vacuum dryers was observed. The two main components of principal component analysis (PCA) explained 88% of the data variance. The structure of the artificial neural network (ANN) was 8-5-4. Further, based on loading diagrams of partial least squares (PLS) and principal component regression (PCR) models, the MQ3 and MQ6 sensors were the best to predict the amount of vitamin C and the color change of orange powder. MQ135 sensor can also be removed from the set of electronic nose sensors due to their low accuracy and cost reduction. The multiple linear regression (MLR), compared to PCR and PLS models, proved to be more accurate (i.e., R²= 0.83 and RMSE= 0.144 for vitamin C prediction and R²= 0.94 and RMSE= 0.68 for predicting color change). The highest and lowest values of measured color change was observed in convective dryer and atmospheric control vacuum dryer, respectively. Also, the highest and lowest measured vitamin C was observed in convective-infrared dryer and atmospheric control vacuum dryer, respectively. The best dryer to maintain the quality of the orange powder is the convective-infrared dryer. The results of this article showed that the data obtained from the olfactory machine is able to predict the color change and vitamin C of orange powder. Also, the olfactory machine can be used to identify and classify the type of dryer used to prepare orange powders with the least time and cost, without distorting the sample, and to determine the best dryer for preparing orange powder.

Abstract This study proposes a novel method for identifying grape leaf diseases through RGB image analysis combined with weighted group decision-making. The investigation focused on five disease types, Black Measles, Black Rot, Leaf Blight, Powdery Mildew, and Downy Mildew, along with healthy leaves. Three machine learning classifiers, namely support vector machine (SVM), random forest (RF), and k-nearest neighbor (k-NN), were employed individually and in a weighted ensemble. Each classifier was assigned a weight based on its accuracy, and the final disease classification was determined using a majority voting strategy. To determine the most discriminative features related to texture, color, and shape, the Relief feature selection algorithm was applied, which identified the top five effective features in diagnosing grape leaf diseases. Experimental results indicated that the classification accuracies of SVM, RF, and k-NN were 88.33%, 80.08%, and 75%, respectively. Furthermore, the proposed weighted group decision-making approach improved the overall classification performance, achieving an accuracy of 91.67%.

Abstract Introduction
Remote sensing is considered a key management tool in precision agriculture, particularly for monitoring and identifying plant coverage. Grapes are among the most valuable horticultural crops, with Hamedan province accounting for approximately 7.3 % of Iran's total vineyard area. This study evaluates the accuracy of vineyard identification in Hamedan province using machine learning algorithms including support vector machine (SVM), minimum distance (MD), and random forest (RF) models along with normalized difference vegetation index (NDVI) and normalized difference water index (NDWI) estimated from combined optical and radar images of Sentinel (Sentinel-1 and Sentinel-2). Based on the most accurate vineyard identification map, the maximum time series of NDVI and NDWI of the MODIS satellite in vineyards was estimated between 2007 and 2020, and their correlation with the actual yield of the grape crop was examined.
Materials and Methods
In this research, we first extracted images of pivotal remote sensing vegetation indicators, including NDVI and NDWI, from Sentinel-2 images in Hamedan province in 2020. The approach of addressing speckle noise through median pixels allowed for the acquisition of median radar images from Sentinel-1 over the designated study area. To create high-accurate images, spectral composition was used to combine these images with the NDVI and NDWI from Sentinel-2 images. Using these images, vineyard identification maps were generated through classification algorithms, including support vector machine, random forest, and minimum distance models. Training samples were used to train these algorithms. Samples from six land coverage classes involving vineyards, were collected using a combination of field observations and Google Earth imagery. Of these, 70% were used for training and 30% for testing the classification models. In order to assess the accuracy of the vineyard identification maps, indicators including overall accuracy and kappa coefficient were examined. Subsequently, the vineyard map with the highest assessment indicator was selected. Finally, using this accurate vineyard identification map, the maximum monthly NDVI and NDWI indices estimated from MODIS sensor images in the vineyards were calculated from 2007 to 2020, and their correlation with yields of the grape crop was computed using Pearson correlation.
Results and Discussion
Based on the comparison of different classification algorithms for distinguishing vineyards, random forest model along with NDVI and NDWI indices outperformed support vector machine and minimum distance models. With regard to accuracy, however, the random forest along with the NDWI has the best overall accuracy (95%) and kappa coefficient (0.95). The superior performance of NDWI is attributed to the high moisture levels in vineyards resulting from irrigation, as NDWI is particularly sensitive to variations in vegetation water content. The lower accuracy of vineyard identification using SVM and MD models can be linked to shadow effects caused by the canopy structure of grapevines, as well as imbalanced training data used for the support vector machine model. Correlation analysis of real grape yields with NDVI and NDWI of MODIS extracted from the highest accuracy vineyards map indicates NDVI (correlation coefficient 0.81) has a stronger linear relationship with yield than NDWI (correlation coefficient 0.75). This can be explained by NDVI's sensitivity to leaf chlorophyll changes, which results in a strong correlation with yield.
Conclusion
Vineyards can be accurately identified using machine learning algorithms and remote sensing vegetation indices derived from combined radar and optical satellite images. Furthermore, the strong correlation between NDVI and grape yield enable reliable yield prediction based on NDVI time series analysis. The outcomes of this study facilitate the identification of grape cultivation areas, improved water resourse management, the development of optimized irrigation strategies, pre-harvest yield estimation, and the exploration of export options.

Abstract Introduction
The increasing demand for automation in agriculture, particularly for repetitive and labor-intensive tasks, has driven the development of robotic harvesting systems. Recent advances in computer vision, deep learning, and the availability of large image datasets have made it possible to create robust object detection models for agricultural applications. Traditional harvesting methods, such as bulk harvesting, often lead to fruit damage and loss owing to non-selective picking. Selective harvesting, particularly with the use of robotic systems, offers a promising alternative by combining the precision of human labor with the efficiency of automation. This study presents a deep learning-based model for detecting cucumber fruits on plants in a real greenhouse environment, which is an essential step towards developing autonomous harvesting robots that selectively pick ripe cucumbers.
Materials and Methods
A dedicated image dataset was curated in a commercial greenhouse, comprising 300 images of cucumber plants captured under various lighting conditions (morning, noon, and evening), to ensure robustness against real-world variability. Images were manually labeled to identify the cucumber fruits and their pedicels. To enhance the model training and prevent overfitting, data augmentation techniques were applied to the training set. Several architectures of the YOLO (You Only Look Once) object detection algorithm were evaluated, including the nano-scale versions YOLOv5n and YOLOv8n, and the small-scale YOLOv8s, in addition to the RT-DETR model.
The YOLOv8 algorithm is known as one of the state-of-the-art algorithms in computer vision because of its high speed, detection accuracy, and adaptability. The YOLOv8 architecture consists of three main parts: backbone, neck, and head, which are responsible for extracting image features, combining and enriching features, and predicting bounding boxes and object classes, respectively.
These models were trained, and their performances were compared based on the detection accuracy and inference time metrics. Training and evaluation were conducted using a suitable computational platform.
Results and Discussion
The performances of different YOLO models and RT-DETR were rigorously evaluated. The results demonstrated that the YOLOv8n model achieved the highest detection accuracy of 87.5%, surpassing the performances of the other tested models. Importantly, the YOLOv8n model also exhibited a favorable balance between the accuracy and inference time, making it suitable for real-time applications. The analysis considered the trade-off between the number of parameters and detection speed, highlighting the efficiency of YOLOv8n.
The YOLOv8n model demonstrated superior performance in terms of pedicel detection accuracy compared to YOLOv5n, achieving a fitness score of 91.08% (calculated as a weighted average of mAP@50 and mAP@50-95). While exhibiting strong performance in fruit and pedicel detection (Figure 6), the sensitivity of the model for pedicel detection (88.0%) was comparatively lower than that for fruit detection (96.1%). The highest F1 score (0.89) was observed at a confidence level of 39.5%, indicating the effectiveness of the model in balancing the precision and recall for pedicel detection. Overall, YOLOv8n outperformed the other tested models in identifying the class and location of the fruit pedicel. The superior performance of YOLOv8n can be attributed to its architectural advancements and optimized training processes.
Conclusion
This study successfully developed a deep learning-based model for accurate and efficient cucumber fruit detection in a greenhouse environment. The YOLOv8n model demonstrated superior performance compared with the other evaluated architectures, achieving a detection accuracy of 87.5% while maintaining a good processing speed. These findings suggest that the YOLOv8n model has significant potential for integration into autonomous vegetable harvesting robots, contributing to the automation of agricultural processes and increased efficiency in greenhouse operations. Future works should explore further optimization and testing under diverse environmental conditions.

Abstract Correct and timely diagnosis of plant diseases is crucial for improving crop performance. Therefore, developing a precise and reliable intelligent system for managing leaf diseases in trees is very important for farmers. This study aims to develop an artificial intelligence-based solution for detecting leaf diseases in quince trees using a state-of-the-art single-stage object detection model, YOLO (You Only Look Once). Images of diverse leaf diseases affecting this tree were collected from multiple sources, including agricultural research centers in Isfahan Province, Iran, relevant websites, and researchers. In this study, a transfer learning approach was employed to evaluate three well-known YOLO models (YOLOv5m, YOLOv7, and YOLOv8m) based on their detection and identification performance. Statistical metrics, including precision, recall, F1-score, and accuracy, were used to evaluate and compare the performance of the investigated models. The results indicate that the accuracy of the YOLOv5m, YOLOv7, and YOLOv8m models were 78%, 83%, and 87%, respectively. Experimental results revealed that YOLOv8m, trained from scratch on the dataset, demonstrates substantial capability in identifying leaf diseases in quince trees. In addition, a comparison showed that this model outperformed other investigated models with scores of 0.87, 0.66, 0.69, and 0.67 for accuracy, precision, recall, and F1-score, respectively. Based on the overall results of this research, the YOLOv8m model trained in this study can be introduced as a specialized tool for this particular crop. Therefore, the developed model in this study, specifically tailored to quince leaf diseases, can be integrated into diagnostic software for tree leaf diseases. Such software can assist farmers in accurately diagnosing diseases, ultimately reducing economic losses.

Abstract The Soil and Plant Analysis Development (SPAD) value is a significant parameter indicating chlorophyll content, particularly in the green parts of plants. Conventional SPAD meters determine this value by measuring the transmission and absorption of red and infrared radiation at a single point (2×3 mm² sensor size). However, obtaining a comprehensive value for an entire leaf requires multiple measurements, increasing processing time. In this study, a non-destructive method for predicting SPAD values was developed using image processing techniques to determine dominant wavelength values from leaf photographs. A custom-designed photo box with controlled 6000 lux white LED lighting was used to capture images at a fixed distance of 15 cm. Images were processed using Color Picker (2024) software, where green components of the leaf were analyzed to extract dominant wavelength values. The results demonstrated that SPAD values could be accurately predicted using dominant wavelength data, with a 98.33% accuracy for the linear model (RMSE: 1.308) and 98.43% for the polynomial model (RMSE: 5.467). The findings indicate that a linear model provides a more precise correlation. This novel approach enhances the efficiency of SPAD measurement and offers a rapid, non-destructive alternative to conventional methods.

Abstract Influence of a single atmospheric component or meteorological variable on the host, pathogen, or their interaction in controlled environments has accounted for the majority of climate change’s impact on plant pests and diseases. Climate change can lead to alterations in the stages and rates of growth of pests and diseases, host resistance, and the physiology of host-pathogen or host-pest interactions, which can cause substantial harm and reduce tomato crop yields. Different approaches have been ineffective in the accuracy of pest and disease forewarning in past years. The remarkable progress in Deep Convolutional Neural Networks (DCNNs) is revolutionizing the early detection of pests and diseases in crops. By analysing vast amounts of present and historical climate data, alongside their expertise in object identification and image categorization, these AI models can predict outbreaks with impressive accuracy. However, understanding the specific microclimate suitable for each pest and disease is crucial for truly effective intervention. Combining these two elements creates a powerful, targeted approach to preserving crops. A forewarning system can help to reduce the use of pesticides, thereby reducing the cost of production and environmental pollution. Proper cloud servers and IoT-based sensor networks should be used for a better forewarning of pests and diseases in future circumstances.

Abstract Introduction
Soil surface roughness is an important factor in determining the intensity and quality of tillage operations, and obtaining accurate information essential for precision tillage. Using an inappropriate technique due to the lack of precise discrepancy detection can lead to increased time spent on analysis and potential damage. Generally, there are two methods for measuring soil surface roughness: contact and non-contact. Contact methods are less accurate for measuring the roughness of soft soil because they involve physical contact, which can partially disturb the soil. Most non-contact measurement methods are also performed in stop-and-go conditions, which increases measurement time and related analysis. The aim of this study is to measure soil surface roughness in real-time using optical sensors in the field. The accuracy and precision of two non-contact measurement methods will be compared to determine the best approach for precision tillage operations.
Materials and Methods
In the current research, a real-time soil surface roughness measurement system consisting of mechanical and electrical modules, data collection, and processing was built. The system performance was evaluated at different forward speeds and roughness categories, with two types of infrared and laser sensors. To assess the sensors’ accuracy, the collected data was compared against the pin gauge method, which served as the reference standard. The method exhibiting the least variation from this reference is considered to provide the most reliable data. Also, to further examine the accuracy of the sensors, the roughness data obtained from the sensor at various frequencies was compared against the roughness data obtained from the pin measuring device at the same level, resulting in a suitable curve plot. The interpretation of the obtained mathematical relationship indicates the precision of the sensor data.
Results and Discussion
The results obtained from the optical sensors were compared to the pin meter, used as the reference method, in both stationary and moving conditions. It was demonstrated that the optical sensors detect distance in the static state similarly to the reference pin meter. The calibration curve interpretation factor was 0.99 for the infrared sensor and 1 for the laser sensor, indicating a strong correlation between the sensor signals and their distance from the soil surface. The random roughness index was significant for different roughness classes at the 1% probability level, showing that this index effectively distinguishes between the resulting roughness classes. Analysis of variance results revealed that the measurement method had a significant effect at the 1% level. The method with the smallest difference from the reference method is considered the most appropriate measurement technique. The effect of forward speed was also significant at the 1% level; the speed at which the sensor’s performance did not significantly differ from the reference method was identified as the optimal speed for the system. Additionally, the effect of roughness class was significant at the 1% level, confirming that the created roughness classes had meaningful differences. The results of the sensor accuracy evaluation showed that the data obtained from the laser sensor at speeds of 1 and 2.6 km h^-1 had no significant difference with the reference method. Therefore, it is appropriate to use the laser sensor at speeds of 1 and 2.6 km h^-1. At speeds higher than 3.5 km h^-1, the laser sensor successfully detected smooth surfaces, but did not correctly distinguish uneven surfaces. In general, the laser sensor was able to detect all categories of roughness at a speed of 2.6 km h^-1. One reason the laser sensor did not perform well at speeds above 2.6 km h^-1 was its low data acquisition rate. By using laser sensors with a higher data collection rate, the soil height profile can be plotted similarly to a pin scale. The infrared sensor was successful only in detecting smooth surfaces but failed to detect other types of surfaces.
Conclusion
Due to limited accuracy and the risk of damaging or altering the surface roughness, the contact method is not recommended for use on soft soil surfaces. Among non-contact methods, the most suitable technique is the one that provides the highest accuracy and precision while minimizing cost and time for data collection and analysis. In this study, two types of sensors including laser and infrared ranging were selected based on their reasonable price, ease of operation, compatibility with a mobile system, and ability to deliver real-time roughness measurements in the shortest possible time. The results demonstrated that real-time measurement of soil surface roughness can effectively replace traditional, tedious, and time-consuming methods.

Abstract Manually picking the flowers of the Damask rose is significantly challenging due to the numerous thorns on its stems. Consequently, the accurate detection of bloomed Damask roses in open fields is crucial for designing a robot capable of automating the harvesting process. Considering the high speed and precise capabilities of deep convolutional neural networks (DCNN), the objective of this study is to investigate the effectiveness of the optimized YOLOv8s model in detecting bloomed Damask roses. To assess the impact of the YOLO model size on network performance, the precision and detection speed of other YOLO network versions, including v5s and v6s, were also examined. Images of Damask roses were taken under two lighting conditions: normal light conditions (from civil twilight to sunrise) and intense light conditions (from sunrise to 10 AM). The outcomes demonstrated that YOLOv8s exhibited the highest performance, with a mean average precision (mAP50) of 98% and a detection speed of 243.9 fps. This outperformed the mAP50 and detection speed of YOLOv5s and YOLOv6s networks by margins of 0.3%, 6.1%, 169.3 fps and 198.6 fps, respectively. Experimental results show that YOLOv8s performs better on images taken in normal lighting than on those taken in intense lighting. A decline of 5.2% in mAP50 and 2.4% in detection speed signifies the adverse influence of intense ambient light on the model's effectiveness. This research indicates that the real-time detector YOLOv8s provides a feasible solution for the identification of Damask rose and provides guidance for the detection of other similar plants.

Abstract In the context of plant diseases, the selection of appropriate preventive measures, such as correct pesticide application, is only possible when plant diseases have been diagnosed quickly and accurately. In this study, a transfer learning model based on the pre-trained EfficientNet model was implemented to detect and classify some diseases in tomato crops, using an augmented training dataset of 2340 images of tomato plants. The study's findings indicate that during the model's validation phase, the rate of image categorization was roughly 5 fps (frames per second), which makes sense for a deep learning model operating on a laptop computer equipped with a standard CPU. Furthermore, the model was learned well because increasing the number of epochs no longer improved its accuracy. After all, the curves of the train and test accuracies, as well as the losses versus epoch numbers, remained largely horizontal for epoch numbers greater than 20. Notably, the highest coefficient of variation across these four cases was only 7%. Furthermore, the cells of the primary diagonal of the confusion matrix were filled with larger numbers in comparison with the values of the other cells; precisely, 88.8%, 7.7%, and 3.3% of the remaining cells of the matrix (cells of the primary diagonal excluded) were filled with 0, 1, and 2, respectively. The model's performance metrics are: sensitivity 85%, specificity 98%, precision 86%, F1-score 84%, and accuracy 85%.

Abstract Robotic harvesting in agriculture is an effective method for producing healthy fruit, reducing costs, and increasing productivity. Detecting and harvesting sweet peppers, however, remains a challenging task. This study aims to develop an unsupervised machine vision algorithm to recognize colored sweet peppers using a combination of geometric features (Fast Point Feature Histogram- FPFH) and color features (H, S, and V). Depth images were captured using a Kinect v2 sensor, and a 3D model was reconstructed. After extracting the geometric and color features, data preprocessing involved undersampling to ensure balance and applying the Z-score criterion to eliminate outliers. Principal component analysis (PCA) was used to reduce the feature dimensions, and the K-means clustering model was implemented to categorize the data using six geometric features and three color features. The silhouette coefficient was employed to evaluate clustering quality, and human evaluation demonsterated that the algorithm achieved a detection accuracy of 95.10% for sweet peppers.

Abstract Drones have emerged as a promising technology in precision agriculture, supporting Sustainable Development Goals (SDGs) by enhancing sustainable farming practices, improving food security, and reducing environmental impact. This review article is intended to meticulously analyze the multiple applications of drone technology in agriculture, such as crop health monitoring, pesticide and fertilizer spraying, weed control, and data-driven decision-making for farm optimization. It emphasizes the role of drones in precision spraying, promoting targeted interventions, and minimizing environmental impact compared to conventional methods. Drones play a vital role in weed management and crop health assessment. The paper focuses on the importance of data collected by drones to acquire the necessary information for decision-making concerning irrigation, fertilization, and overall farm management. However, using Unmanned Aerial Vehicles (UAVs) in agriculture faces challenges caused by batteries and their life, flight time, and connectivity issues, particularly in remote areas. There are legal challenges whereby regulatory frameworks and restrictions are present in different regions that affect the operation of drones. With the help of continuous research and development initiatives, the challenges depicted above could be solved, and the fullest potential of drones can be tapped for achieving Sustainable Agriculture.

Abstract Introduction
Subsoiling is a critical tillage operation for many crops, particularly sugarcane, due to the impact of agricultural machinery traffic and its significance in managing heavy-textured and compacted soils. Given the extensive size of sugarcane fields and the time-intensive nature of subsoiling operations, the application of intelligent control techniques for monitoring and managing these processes is of considerable importance. Currently, subsoiling operations are monitored using manual gauges. This approach involves collecting a limited number of samples per hectare, typically after the operation is completed, which makes it nearly impossible to implement real-time corrections. To address this limitation, the development and implementation of a depth measurement system offer a promising solution. Such a system enables real-time observation of working depth by both the operator, via an on-screen display, and by a remote observer through an online platform. This capability allows for immediate adjustments during the operation, ensuring greater precision and efficiency. Furthermore, by integrating recorded depth data with geospatial information, it becomes possible to generate detailed maps illustrating depth variations across the field. These maps can serve as valuable tools for further evaluations, such as performance monitoring in areas where subsoiling depth deviates from the desired range, either being too shallow or excessively deep. This technological advancement has the potential to significantly enhance the accuracy and effectiveness of subsoiling operations in modern agricultural practices.
Materials and Methods
This study focused on the design, development, and evaluation of a depth measurement system for a subsoiler attached to a track-type tractor, specifically tailored for sugarcane fields. The system not only provided real-time depth display but also recorded the location and transmitted it online. The research employed three distinct depth measurement techniques and was conducted using a randomized complete block design with split plots. The main plots are the three depth measurement techniques: based on the angles of the driving profiles of the subsoiler shanks (T₁), the laser distance measurement method (T₂), and the ultrasonic distance measurement method (T₃), and sub-plots are depth ranges at three levels: 0-30 cm (R₁: surface range), 30-60 cm (R₂: mid-range), and 60-90 cm (R₃: deep range). Initially, we calculated the absolute difference between the depths recorded by the system and those measured manually with a rod at each location. Following this, we analyzed key statistical indicators, including the average, standard deviation, and the minimum and maximum of errors, for comparison.
Results and Discussion
The results showed that the depth measurement error was significantly influenced by the technique employed. The angle technique yielded the lowest average error of 1.91 cm, while the ultrasonic technique resulted in the highest average error of 3.83 cm. Across all depth ranges, statistical indicators for depth error were significant. Specifically, within these ranges, the deep range exhibited an average depth error of 2.33 cm, and the surface range had an average error of 3.65 cm. Statistical analysis revealed that only indices related to minimum and maximum errors for interactions between factors were significant. The lowest minimum error value (0.05 cm) was observed with the angle technique at deeper depths, whereas the highest minimum error (0.34 cm) occurred with ultrasonic measurements at shallower depths on surfaces. Similarly, maximum errors followed this trend: The lowest maximum error (3.21 cm) was associated with angle measurements at deeper depths, while ultrasonic measurements on surfaces yielded a higher maximum error (8.63 cm). Both laser and ultrasonic techniques consistently demonstrated greater errors across all three depth ranges compared to angle-based methods. This discrepancy may be attributed to inaccuracies inherent in rangefinders when their beams encounter obstacles like clods or pits during field operations. Notably, as working depths increased across all measurement techniques, errors in depth measurement decreased significantly due to reduced vibrations from subsoiler devices at greater depths, thereby minimizing vibration-related inaccuracies.
Conclusion
The results indicate that the depth measurement technique based on the angles of the driving profiles of subsoiler shanks exhibits superior accuracy in determining the working depth of subsoilers mounted on tractors, particularly during sugarcane field operations. The laser distance meter technique ranked second in terms of accuracy, while the ultrasonic distance meter method demonstrated the least precision. Notably, as working depths increased, reduced vibrations during operation were observed, leading to enhanced accuracy in depth calculations across all techniques. This improvement is attributed to decreased mechanical disturbances at greater depths. Overall, measurements within deeper ranges achieved higher levels of accuracy compared to those at shallower surface ranges. This trend suggests that operational conditions and device stability play significant roles in optimizing measurement accuracy.

Abstract Introduction
The formation of ruts induced by vehicle traffic poses a significant challenge for agricultural soils due to soil compaction both at the surface and deeper layers. This phenomenon compromises vehicle performance increases energy consumption, and leads to long-term environmental degradation, such as soil erosion and fertility reduction. To enhance vehicle performance and reduce soil damage, it is crucial to accurately predict how factors such as vehicle speed, vertical load, and the number of passes impact rut depth. The findings of this study hold significant practical implications, facilitating the development for the creation of more efficient agricultural practices, while simultaneously minimizing environmental impact. The complexity of these interactions necessitates using machine learning models, especially artificial neural networks (ANNs), to predict rut depth based on input parameters. In this study, two machine learning models, namely the multilayer perceptron (MLP) and the radial basis function (RBF) networks, were employed to predict rut depth.
Materials and Methods
Experiments were conducted using a soil bin that allows for precise control of independent parameters, measuring 24 meters in length, 2 meters in width, and 0.8 meters in depth. The soil used was agricultural soil, comprising 35% sand, 22% silt, and 43% clay, with a moisture content of 8%. The tests included three independent parameters: vertical load (2, 3, and 4 kN), forward speed (1, 2, and 3 km h^-1), and number of wheel passes (up to 15). Two types of traction devices, including a rubber wheel and a track wheel, were tested. A caliper was used to measure the rut depth after each pass with an accuracy of 0.02 mm. The data collected from soil bin tests were used to train neural network models in MATLAB 2021-b software. The MLP model had a topology with two hidden layers and included three inputs and one output. In the RBF model, the network topology had a single hidden layer. The trial-and-error method was used to adjust the hyperparameters of the neural networks, including the number of neurons in the hidden layers, the learning rate, and momentum for the MLP network, as well as the spread rate and regularization rate for the RBF network.
Results and Discussion
Experimental data confirmed that increasing the vertical load and the number of passes resulted in deeper ruts. Conversely, an increase in speed led to a reduction in rut depth, particularly during the initial pass. Both artificial neural network (ANN) models accurately predicted rut depth, with the multilayer perceptron (MLP) neural network outperforming the radial basis function (RBF) neural network. Specifically, the root mean square error (RMSE) for the optimal MLP model, which utilized a learning rate of 0.001 and a momentum of 0.67, was 0.10. In contrast, the optimal RBF model, with an expansion rate of 0.23456, yielded an RMSE of 0.12. The findings indicate that the MLP artificial neural network model surpasses the RBF neural network model in terms of accuracy and overall performance. However, the RBF neural network exhibits a faster response time, making it particularly suitable for real-time applications.
Conclusion
This study demonstrates the efficacy of machine learning techniques, particularly artificial neural networks (ANNs), in predicting rut depth caused by off-road vehicle traffic. Both multilayer perceptron (MLP) and radial basis function (RBF) neural networks exhibited robust predictive capabilities, with the MLP model providing slightly superior accuracy and the RBF model offering better computational efficiency. These findings highlight the potential of machine learning in modeling complex interactions between soil and vehicles, which can enhance vehicle performance, mitigate soil erosion, and guide the design of off-road vehicles. Future research directions could include investigating additional soil parameters, various vehicle configurations, and the real-world implementation of autonomous off-road vehicles to promote more environmentally sustainable operations.

Abstract The IoT monitoring system for stingless bee colonies aims to provide real-time information about temperature, humidity, and hive weight in response to the issue of colony collapse disorder (CCD) caused by human intervention in beekeeping. It also aims to improve the current monitoring methods for the bees more effectively and efficiently. The monitoring system features a water-cooling control system to maintain an optimal temperature for Tetragonula Biroi (Stingless Bees). The system also includes a user dashboard for remote monitoring and alerts the beekeeper when it is time to harvest. It’s primarily built around the ESP8266-MOD microcontroller, with an Arduino Mega 2560 R3 for the water valve control system. Data were collected from a DHT22 sensor for temperature and humidity, and load cells connected to an HX711 amplifier for hive weight. The system was tested by comparing samples from the system and actual measuring instruments using MAPE for two months, and it demonstrated 98.74% and 97.89% accuracy for surrounding temperature and humidity, respectively. An accuracy of 95.92% for the weight scale and 93% for the water valve control system was also obtained. Hives equipped with the IoT system gained 3.414% more weight than those without it, indicating that the project succeeded in achieving its objectives.