Pub Date : 2026-01-20DOI: 10.1016/j.compag.2026.111462
Alberto Carraro , Giulia Bugin , Francesco Marinello, Maddi Aguirrebengoa, Stefano Frattini, Andrea Pezzuolo
Accurate assessment of dairy cow cleanliness is essential for ensuring animal welfare, maintaining udder health, and optimising milk production. Traditional visual inspections are subjective and often fail to distinguish dirt from natural coat patterns, especially in spotted breeds. This research investigates the applicability of a two-stage approach for automated cleanliness evaluation, consisting of (i) semantic segmentation of dirt areas on cow coats and (ii) regression from the resulting masks to numerical cleanliness scores. The first stage was implemented using the U-Net and DeepLabV3 architectures, which were trained on either RGB-only or RGB-Thermal (RGB-T) images. Incorporating thermal information significantly improved segmentation accuracy: U-Net achieved a mean Intersection over Union (mIoU) of 0.5244 on RGB-T images, compared to 0.3537 on RGB images, while DeepLabV3 on RGB-T images reached an mIoU of 0.5049. The second stage compared two regression strategies: multiple linear regression (MLR) on the number of pixels classified as dirt, and convolutional neural networks (CNNs) trained directly on the masks. CNN-based regression consistently outperformed MLR, with the best performance obtained by combining RGB-T segmentation and CNN regression (DeepLabV3 + CNN: MAPE 23.05 %; U-Net + CNN: MAPE 25.24 %). These findings support the feasibility of a two-stage RGB-T-based approach for objective cleanliness evaluation, highlighting the benefits of thermal information for segmentation and CNNs for score prediction.
{"title":"AI-driven analysis of animal cleanliness: A data-fusion model using RGB and thermal imaging","authors":"Alberto Carraro , Giulia Bugin , Francesco Marinello, Maddi Aguirrebengoa, Stefano Frattini, Andrea Pezzuolo","doi":"10.1016/j.compag.2026.111462","DOIUrl":"10.1016/j.compag.2026.111462","url":null,"abstract":"<div><div>Accurate assessment of dairy cow cleanliness is essential for ensuring animal welfare, maintaining udder health, and optimising milk production. Traditional visual inspections are subjective and often fail to distinguish dirt from natural coat patterns, especially in spotted breeds. This research investigates the applicability of a two-stage approach for automated cleanliness evaluation, consisting of (i) semantic segmentation of dirt areas on cow coats and (ii) regression from the resulting masks to numerical cleanliness scores. The first stage was implemented using the U-Net and DeepLabV3 architectures, which were trained on either RGB-only or RGB-Thermal (RGB-T) images. Incorporating thermal information significantly improved segmentation accuracy: U-Net achieved a mean Intersection over Union (mIoU) of 0.5244 on RGB-T images, compared to 0.3537 on RGB images, while DeepLabV3 on RGB-T images reached an mIoU of 0.5049. The second stage compared two regression strategies: multiple linear regression (MLR) on the number of pixels classified as dirt, and convolutional neural networks (CNNs) trained directly on the masks. CNN-based regression consistently outperformed MLR, with the best performance obtained by combining RGB-T segmentation and CNN regression (DeepLabV3 + CNN: MAPE 23.05 %; U-Net + CNN: MAPE 25.24 %). These findings support the feasibility of a two-stage RGB-T-based approach for objective cleanliness evaluation, highlighting the benefits of thermal information for segmentation and CNNs for score prediction.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111462"},"PeriodicalIF":8.9,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025266","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-19DOI: 10.1016/j.compag.2025.111391
Preston Fairchild, Claudia Chen, Xiaobo Tan
Properly handling delicate produce with robotic manipulators is a major part of the future role of automation in agricultural harvesting and processing. Grasping with the correct amount of force is crucial in not only ensuring proper grip on the object, but also to avoid damaging or bruising the product. In this work, a flexible pressure sensor that is both low cost and easy to fabricate is integrated with robotic grippers for working with produce of varying shapes, sizes, and stiffnesses. The sensor is successfully integrated with both a rigid robotic gripper, as well as a pneumatically actuated soft finger. Furthermore, an algorithm is proposed for accelerated estimation of the steady-state value of the sensor output based on the transient response data, to enable real-time applications. The sensor is shown to be effective in incorporating feedback to correctly grasp objects of unknown sizes and stiffnesses. At the same time, the sensor provides estimates for these values which can be utilized for identification of qualities such as ripeness levels and bruising. It is also shown to be able to provide force feedback for objects of variable stiffnesses. This enables future use not only for produce identification, but also for tasks such as quality control and selective distribution based on ripeness levels.
{"title":"Efficient force and stiffness prediction in robotic produce handling with a piezoresistive pressure sensor","authors":"Preston Fairchild, Claudia Chen, Xiaobo Tan","doi":"10.1016/j.compag.2025.111391","DOIUrl":"10.1016/j.compag.2025.111391","url":null,"abstract":"<div><div>Properly handling delicate produce with robotic manipulators is a major part of the future role of automation in agricultural harvesting and processing. Grasping with the correct amount of force is crucial in not only ensuring proper grip on the object, but also to avoid damaging or bruising the product. In this work, a flexible pressure sensor that is both low cost and easy to fabricate is integrated with robotic grippers for working with produce of varying shapes, sizes, and stiffnesses. The sensor is successfully integrated with both a rigid robotic gripper, as well as a pneumatically actuated soft finger. Furthermore, an algorithm is proposed for accelerated estimation of the steady-state value of the sensor output based on the transient response data, to enable real-time applications. The sensor is shown to be effective in incorporating feedback to correctly grasp objects of unknown sizes and stiffnesses. At the same time, the sensor provides estimates for these values which can be utilized for identification of qualities such as ripeness levels and bruising. It is also shown to be able to provide force feedback for objects of variable stiffnesses. This enables future use not only for produce identification, but also for tasks such as quality control and selective distribution based on ripeness levels.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111391"},"PeriodicalIF":8.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025085","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-19DOI: 10.1016/j.compag.2026.111434
Mingji Wei , Fei Lyu , Shuai Lu , Weijie Liu , Zhaoxuan Fan , Ning Yang , Wenhao Hui
Tomato, a vital global crop, faces severe annual yield losses of 20–40% due to the pest stress, while conventional identification methods often fail to detect early infections before visible symptoms emerge. To overcome the limited sensitivity and generalizability of image-based approaches, this article proposed a tomato pest stress classification method based on a mechanism-guided hybrid network (MGHN) to provide the theoretical foundation for real-time early-warning systems. We first extracted three mechanism-driven features including transient change rates, pulse amplitude/frequency and baseline offset from the dynamics of jasmonic acid (JA) and salicylic acid (SA). These features were then fused with a 1D-CNN to classify the three pest types including piercing-sucking, chewing and crawling-feeding pests. Results show that our proposed MGHN method achieves 94.8% average accuracy on 135 independent test samples, with significantly optimized recognition for piercing-sucking pests (95.56% F1), chewing pests (94.50% F1), and crawling-feeding pests (94.37% F1). Comparative analysis against traditional 1D-CNN and a mechanism-only model demonstrates that MGHN outperformed 1D-CNN (91.1%) and the mechanism-only model (89.6%) by 3.7% and 5.2%, respectively. This research can establish a theoretical foundation for real-time early pest stress warning systems of crops using wearable sensors.
{"title":"Mechanism-guided deep learning for pest classification in tomato leaves","authors":"Mingji Wei , Fei Lyu , Shuai Lu , Weijie Liu , Zhaoxuan Fan , Ning Yang , Wenhao Hui","doi":"10.1016/j.compag.2026.111434","DOIUrl":"10.1016/j.compag.2026.111434","url":null,"abstract":"<div><div>Tomato, a vital global crop, faces severe annual yield losses of 20–40% due to the pest stress, while conventional identification methods often fail to detect early infections before visible symptoms emerge. To overcome the limited sensitivity and generalizability of image-based approaches, this article proposed a tomato pest stress classification method based on a mechanism-guided hybrid network (MGHN) to provide the theoretical foundation for real-time early-warning systems. We first extracted three mechanism-driven features including transient change rates, pulse amplitude/frequency and baseline offset from the dynamics of jasmonic acid (JA) and salicylic acid (SA). These features were then fused with a 1D-CNN to classify the three pest types including piercing-sucking, chewing and crawling-feeding pests. Results show that our proposed MGHN method achieves 94.8% average accuracy on 135 independent test samples, with significantly optimized recognition for piercing-sucking pests (95.56% <em>F</em>1), chewing pests (94.50% <em>F</em>1), and crawling-feeding pests (94.37% <em>F</em>1). Comparative analysis against traditional 1D-CNN and a mechanism-only model demonstrates that MGHN outperformed 1D-CNN (91.1%) and the mechanism-only model (89.6%) by 3.7% and 5.2%, respectively. This research can establish a theoretical foundation for real-time early pest stress warning systems of crops using wearable sensors.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111434"},"PeriodicalIF":8.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025088","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-19DOI: 10.1016/j.compag.2026.111438
Yuncheng Deng , Suling He , Jinliang Wang , Jianpeng Zhang , Bangjin Yi , Congtao Hu , Yanling Jiang , An Chen
Reforestation is crucial for ecological restoration in mining areas, and precise classification of tree species is an important prerequisite for evaluating the effectiveness of ecological restoration. Unmanned aerial vehicles (UAVs) hyperspectral imagery enables highly accurate classification of tree species, leveraging its superior spectral and spatial resolutions. However, the challenge of dimensionality escalation, caused by hundreds of spectral features across numerous bands, introduces new hurdles for conventional classification methods. Deep learning provides a new solution for automatic feature extraction and tree classification and mapping. However, there are problems such as single scale of extracted features, static model inference, and weak model generalization ability. Especially in the ecological restoration area of the mining area, where the terrain environment is complex and tree species of multiple forest age levels coexist, and the classification accuracy needs to be improved. Therefore, this study uses UAV-based hyperspectral imagery acquired from the Jianshan mining area in Kunming, Yunnan Province, as the primary data source, and proposes a novel three-dimensional convolutional neural network (SDTA-3DCNN) incorporating a separable depth transposed attention mechanism. The model enhances the extraction of high-dimensional sparse spectral features through a cascaded 3DCNN architecture, while effectively resolving complex spectral patterns by leveraging an attention mechanism to fuse local features with high-level global representations. The final tree classification accuracy reaches F1-score = 0.9814, OA = 0.9903 and Kappa = 0.9868. The proposed method is also well applied in other tree classification and crop classification scenarios, with OA and Kappa of 0.9871 and 0.9688 in other tree classification scenarios, respectively, and the highest classification accuracy in crop classification scenarios reaches OA = 0.9712, Kappa = 0.9635. The method can provide scientific basis and technical support for the fine classification of tree species in mining areas, the monitoring of ecological restoration status in mining areas, the evaluation of effects, and the decision-making.
{"title":"Joint cascaded 3DCNN and SDTA encoding for tree species classification using UAV-based hyperspectral image in mining areas","authors":"Yuncheng Deng , Suling He , Jinliang Wang , Jianpeng Zhang , Bangjin Yi , Congtao Hu , Yanling Jiang , An Chen","doi":"10.1016/j.compag.2026.111438","DOIUrl":"10.1016/j.compag.2026.111438","url":null,"abstract":"<div><div>Reforestation is crucial for ecological restoration in mining areas, and precise classification of tree species is an important prerequisite for evaluating the effectiveness of ecological restoration. Unmanned aerial vehicles (UAVs) hyperspectral imagery enables highly accurate classification of tree species, leveraging its superior spectral and spatial resolutions. However, the challenge of dimensionality escalation, caused by hundreds of spectral features across numerous bands, introduces new hurdles for conventional classification methods. Deep learning provides a new solution for automatic feature extraction and tree classification and mapping. However, there are problems such as single scale of extracted features, static model inference, and weak model generalization ability. Especially in the ecological restoration area of the mining area, where the terrain environment is complex and tree species of multiple forest age levels coexist, and the classification accuracy needs to be improved. Therefore, this study uses UAV-based hyperspectral imagery acquired from the Jianshan mining area in Kunming, Yunnan Province, as the primary data source, and proposes a novel three-dimensional convolutional neural network (SDTA-3DCNN) incorporating a separable depth transposed attention mechanism. The model enhances the extraction of high-dimensional sparse spectral features through a cascaded 3DCNN architecture, while effectively resolving complex spectral patterns by leveraging an attention mechanism to fuse local features with high-level global representations. The final tree classification accuracy reaches F1-score = 0.9814, OA = 0.9903 and Kappa = 0.9868. The proposed method is also well applied in other tree classification and crop classification scenarios, with OA and Kappa of 0.9871 and 0.9688 in other tree classification scenarios, respectively, and the highest classification accuracy in crop classification scenarios reaches OA = 0.9712, Kappa = 0.9635. The method can provide scientific basis and technical support for the fine classification of tree species in mining areas, the monitoring of ecological restoration status in mining areas, the evaluation of effects, and the decision-making.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111438"},"PeriodicalIF":8.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025089","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-19DOI: 10.1016/j.compag.2026.111425
Baocheng Zhou , Shaochun Ma , Wenzhi Li , Jinzhi Ma , Yansu Xie , Sha Yang
Real-time adjustment of extractor speed according to feed rate is essential to reduce impurity content and cane loss in mechanized sugarcane harvesting. An automatic control system for sugarcane harvester extractor was developed in this study aiming to achieve dynamic matching between speed and feed rate, thereby reducing impurity content and cane loss during harvesting. An optimal control strategy between feed rate and rotational speed was established using impurity content and cane loss as indicators. A variable universe fuzzy multi-parameter adaptive PID (VUFMA-PID) control method was proposed and modeled in Simulink. Compared with conventional PID and fuzzy PID, the VUFMA-PID achieved the shortest steady-state response time, 0.32 s and 0.26 s faster than PID and fuzzy PID, with both steady-state error and maximum overshoot reduced to zero. Field experiments were conducted under different feed rate fluctuation orders, with fixed extractor speed and manual adjustment speed based on operator experience used as control groups. The results indicated that, compared to manual and constant mode, the average power consumption of the automatic control mode was reduced by 17.44 % and 30.40 % respectively. The average impurity content was 4.00 %, which decreased by 23.58 % and 10.71 %. The average cane loss was 1.89 %, which decreased by 25.01 % and 28.52 %. The developed automatic control system effectively adapts to varying feed rates and significantly improves harvesting quality. It provides a feasible solution and theoretical support for intelligent control in mechanized sugarcane harvesting.
{"title":"Development and performance evaluation of an automatic control system for sugarcane harvester extractor","authors":"Baocheng Zhou , Shaochun Ma , Wenzhi Li , Jinzhi Ma , Yansu Xie , Sha Yang","doi":"10.1016/j.compag.2026.111425","DOIUrl":"10.1016/j.compag.2026.111425","url":null,"abstract":"<div><div>Real-time adjustment of extractor speed according to feed rate is essential to reduce impurity content and cane loss in mechanized sugarcane harvesting. An automatic control system for sugarcane harvester extractor was developed in this study aiming to achieve dynamic matching between speed and feed rate, thereby reducing impurity content and cane loss during harvesting. An optimal control strategy between feed rate and rotational speed was established using impurity content and cane loss as indicators. A variable universe fuzzy multi-parameter adaptive PID (VUFMA-PID) control method was proposed and modeled in Simulink. Compared with conventional PID and fuzzy PID, the VUFMA-PID achieved the shortest steady-state response time, 0.32 s and 0.26 s faster than PID and fuzzy PID, with both steady-state error and maximum overshoot reduced to zero. Field experiments were conducted under different feed rate fluctuation orders, with fixed extractor speed and manual adjustment speed based on operator experience used as control groups. The results indicated that, compared to manual and constant mode, the average power consumption of the automatic control mode was reduced by 17.44 % and 30.40 % respectively. The average impurity content was 4.00 %, which decreased by 23.58 % and 10.71 %. The average cane loss was 1.89 %, which decreased by 25.01 % and 28.52 %. The developed automatic control system effectively adapts to varying feed rates and significantly improves harvesting quality. It provides a feasible solution and theoretical support for intelligent control in mechanized sugarcane harvesting.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111425"},"PeriodicalIF":8.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025087","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-19DOI: 10.1016/j.compag.2026.111445
Philippe Vigneault , Samuel de la Sablonnière , Arianne Deshaies , Kosal Khun , Joel Lafond-Lapalme , Louis Longchamps , Étienne Lord
This study addresses the challenge of forecasting maize yield in southeastern Quebec by comparing weekly Unmanned Aerial Vehicle (UAV) and PlanetScope satellite imagery throughout the cropping season and across diverse growing conditions. Using five nitrogen treatments over three years with two sowing windows each year to generate variability within the dataset, eleven vegetation indices were evaluated to identify the best-performing indices and the optimal forecasting window. Indices were interpolated using curve fitting to enable evaluation at any stage of the growing season. Cross-validation simulated real-world application by excluding entire sowing events during model testing. Using a linear regression approach, results demonstrate that indices combining green and near-infrared bands (Green Normalized Difference Vegetation Index [GNDVI] and Chlorophyll Index Green [CIG]) exhibit superior forecasting potential compared to red-near-infrared (like NDVI) and RGB-based indices (like NGRDI). The optimal forecast window occurs during early grain filling (R2-R3 stages, around 2300 Crop Heat Units [CHU]), achieving Root Mean Square Coefficient of Variation (RMSCV) values of 12.51 % for UAVs and 15.28 % for PlanetScope. While PlanetScope maximum performance approached UAV capabilities, results showed a CHU range between 200 and 400 in the effective forecasting period (RMSCV < 20 %) compared to 850 to 1450 for UAV. For PlanetScope, adding multiple indices marginally improved precision, slightly reduced forecasting window and reduced model transferability. The analysis revealed weak correlations between indices and yield during early vegetative and senescence phases, indicating limited potential for enabling timely in-season management interventions. This study established UAV-based models as a reference point for assessing the limitations of satellite-derived forecasts.
{"title":"Yield forecasting in maize: Performance and limits of unmanned aerial vehicle and PlanetScope remote sensing across multiple growth cycles","authors":"Philippe Vigneault , Samuel de la Sablonnière , Arianne Deshaies , Kosal Khun , Joel Lafond-Lapalme , Louis Longchamps , Étienne Lord","doi":"10.1016/j.compag.2026.111445","DOIUrl":"10.1016/j.compag.2026.111445","url":null,"abstract":"<div><div>This study addresses the challenge of forecasting maize yield in southeastern Quebec by comparing weekly Unmanned Aerial Vehicle (UAV) and PlanetScope satellite imagery throughout the cropping season and across diverse growing conditions. Using five nitrogen treatments over three years with two sowing windows each year to generate variability within the dataset, eleven vegetation indices were evaluated to identify the best-performing indices and the optimal forecasting window. Indices were interpolated using curve fitting to enable evaluation at any stage of the growing season. Cross-validation simulated real-world application by excluding entire sowing events during model testing. Using a linear regression approach, results demonstrate that indices combining green and near-infrared bands (Green Normalized Difference Vegetation Index [GNDVI] and Chlorophyll Index Green [CIG]) exhibit superior forecasting potential compared to red-near-infrared (like NDVI) and RGB-based indices (like NGRDI). The optimal forecast window occurs during early grain filling (R2-R3 stages, around 2300 Crop Heat Units [CHU]), achieving Root Mean Square Coefficient of Variation (RMSCV) values of 12.51 % for UAVs and 15.28 % for PlanetScope. While PlanetScope maximum performance approached UAV capabilities, results showed a CHU range between 200 and 400 in the effective forecasting period (RMSCV < 20 %) compared to 850 to 1450 for UAV. For PlanetScope, adding multiple indices marginally improved precision, slightly reduced forecasting window and reduced model transferability. The analysis revealed weak correlations between indices and yield during early vegetative and senescence phases, indicating limited potential for enabling timely in-season management interventions. This study established UAV-based models as a reference point for assessing the limitations of satellite-derived forecasts.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111445"},"PeriodicalIF":8.9,"publicationDate":"2026-01-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025265","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-18DOI: 10.1016/j.compag.2026.111432
Florian Kitzler , Alexander Bauer , Viktoria Kruder-Motsch
The emergence of smart farming in recent years has substantially increased the importance of artificial vision systems in crop production. Data augmentation is essential for developing robust semantic segmentation models when dealing with small datasets, such as in selective weed control. Due to advances in multi-modal data fusion, RGB-D image datasets contribute substantially to improve model performance. However, most data augmentation techniques primarily modify the color channels, often neglecting the depth channel. Addressing this gap, we introduce three methods for augmenting RGB-D images: RGB-D-Aug, Recompose3D, and Compose3D. We conducted experiments utilizing a multi-modal fusion network tailored for semantic segmentation of different plant species, namely ESANet. RGB-D-Aug introduces artificial depth sensor noise in addition to commonly used geometric transformations and color variations. Recompose3D and Compose3D generate augmented RGB-D images and corresponding ground-truth labels by composing background images and a set of foreground plant snippets. Recompose3D rearranges plants from a given training image, while Compose3D employs all plant snippets available in the training dataset. In our experiments designed to evaluate generalization performance, we tested our three methods and compared them not only to the augmentation technique used in ESANet, which consists of geometric transformations and color channel variations, but also to an extended version of the Copy-Paste method, an image composition technique originally introduced for RGB images. All three of our proposed methods outperformed the ESANet augmentation. The image composition methods, Copy-Paste, Recompose3D, and Compose3D, performed significantly better, with Compose3D achieving the highest generalization performance of all methods tested. In addition to improving model robustness, Compose3D allows the creation of realistic agronomic image scenes. Our research is an important step towards developing robust and generalizable models for different applications in arable farming.
{"title":"Beyond Color: Advanced RGB-D data augmentation for robust semantic segmentation in crop farming scenes","authors":"Florian Kitzler , Alexander Bauer , Viktoria Kruder-Motsch","doi":"10.1016/j.compag.2026.111432","DOIUrl":"10.1016/j.compag.2026.111432","url":null,"abstract":"<div><div>The emergence of smart farming in recent years has substantially increased the importance of artificial vision systems in crop production. Data augmentation is essential for developing robust semantic segmentation models when dealing with small datasets, such as in selective weed control. Due to advances in multi-modal data fusion, RGB-D image datasets contribute substantially to improve model performance. However, most data augmentation techniques primarily modify the color channels, often neglecting the depth channel. Addressing this gap, we introduce three methods for augmenting RGB-D images: RGB-D-Aug, Recompose3D, and Compose3D. We conducted experiments utilizing a multi-modal fusion network tailored for semantic segmentation of different plant species, namely ESANet. RGB-D-Aug introduces artificial depth sensor noise in addition to commonly used geometric transformations and color variations. Recompose3D and Compose3D generate augmented RGB-D images and corresponding ground-truth labels by composing background images and a set of foreground plant snippets. Recompose3D rearranges plants from a given training image, while Compose3D employs all plant snippets available in the training dataset. In our experiments designed to evaluate generalization performance, we tested our three methods and compared them not only to the augmentation technique used in ESANet, which consists of geometric transformations and color channel variations, but also to an extended version of the Copy-Paste method, an image composition technique originally introduced for RGB images. All three of our proposed methods outperformed the ESANet augmentation. The image composition methods, Copy-Paste, Recompose3D, and Compose3D, performed significantly better, with Compose3D achieving the highest generalization performance of all methods tested. In addition to improving model robustness, Compose3D allows the creation of realistic agronomic image scenes. Our research is an important step towards developing robust and generalizable models for different applications in arable farming.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111432"},"PeriodicalIF":8.9,"publicationDate":"2026-01-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025086","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-18DOI: 10.1016/j.compag.2026.111447
Xin-Dong Ni , Mao-Lin Li , Fang-Lei Li , Zi-Xuan Dai , Chun-Xiao Xing , Feng Wang , Yan-Xin Yin , Du Chen , Zhi-Zhu He
Accurate and stable grain flow monitoring plays a critical role in yield estimation and closed-loop control of combine harvesters. To address the limitations of existing flow sensors under dynamic and noisy field conditions, this study proposes a novel grain flow sensing method based on a magnetoelastomer–Hall array structure. A multilayer flexible composite structure, comprising a NdFeB magnetic film and a PDMS elastomeric substrate, was developed to enable direct, in-situ transduction of grain flow impact into magnetic field variations. The sensor was structurally designed using discrete element method (DEM) simulations to align with the spatial impact distribution patterns of free-falling grains. A 3 × 3 magnetoelastomer array was arranged to capture the spatial characteristics of grain impacts. And a maximum-response extraction strategy was adopted to enhance signal robustness under heterogeneous grain flow. Laboratory tests confirmed that 3 × 3 array architecture preserves the mechanical compliance and sensitivity of the magnetoelastomer units without performance trade-offs. Grain flow detection performance was evaluated on a grain flow test bench and experimental results revealed a strong linear relationship between sensor output and actual grain flow (), with a mean yield estimation error of 6.07%. Therefore, the proposed sensing system provides a flexible, sensitive, and reliable solution for real-time grain flow monitoring, and establishes a foundation for introducing flexible sensing into next-generation intelligent agricultural machinery.
{"title":"Magnetoelastomer-based grain flow sensor for combine harvesters","authors":"Xin-Dong Ni , Mao-Lin Li , Fang-Lei Li , Zi-Xuan Dai , Chun-Xiao Xing , Feng Wang , Yan-Xin Yin , Du Chen , Zhi-Zhu He","doi":"10.1016/j.compag.2026.111447","DOIUrl":"10.1016/j.compag.2026.111447","url":null,"abstract":"<div><div>Accurate and stable grain flow monitoring plays a critical role in yield estimation and closed-loop control of combine harvesters. To address the limitations of existing flow sensors under dynamic and noisy field conditions, this study proposes a novel grain flow sensing method based on a magnetoelastomer–Hall array structure. A multilayer flexible composite structure, comprising a NdFeB magnetic film and a PDMS elastomeric substrate, was developed to enable direct, in-situ transduction of grain flow impact into magnetic field variations. The sensor was structurally designed using discrete element method (DEM) simulations to align with the spatial impact distribution patterns of free-falling grains. A 3 × 3 magnetoelastomer array was arranged to capture the spatial characteristics of grain impacts. And a maximum-response extraction strategy was adopted to enhance signal robustness under heterogeneous grain flow. Laboratory tests confirmed that 3 × 3 array architecture preserves the mechanical compliance and sensitivity of the magnetoelastomer units without performance trade-offs. Grain flow detection performance was evaluated on a grain flow test bench and experimental results revealed a strong linear relationship between sensor output and actual grain flow (<span><math><mrow><msup><mrow><mi>R</mi></mrow><mn>2</mn></msup><mo>=</mo><mtext>0.98</mtext></mrow></math></span>), with a mean yield estimation error of 6.07%. Therefore, the proposed sensing system provides a flexible, sensitive, and reliable solution for real-time grain flow monitoring, and establishes a foundation for introducing flexible sensing into next-generation intelligent agricultural machinery.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111447"},"PeriodicalIF":8.9,"publicationDate":"2026-01-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025152","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-18DOI: 10.1016/j.compag.2026.111455
Jian Jiang , Xichen Yang , Hui Yan , Jia Liu , Yifan Chen , Zhongyuan Mao , Tianshu Wang
Chrysanthemum is a traditional Chinese medicinal herb that is widely employed in various medical applications. The medicinal value of chrysanthemums varies among different types, and this value is highly correlated with their specific classification. Therefore, the classification of chrysanthemums is essential for ensuring their proper medicinal use. However, existing traditional classification methods are often time-consuming and costly, making them less suitable for practical applications. To overcome these limitations, a novel chrysanthemum classification method based on multi-stream deep color space feature fusion is proposed. Firstly, chrysanthemum images are transformed from the RGB color space to the HSV and LAB color spaces. And the classification-aware features are extracted from the H, S, and L channels, respectively. Secondly, a multi-stream deep network is designed, which employing both 1D and 2D networks. The 1D network focus on further analyzing the features from the H, S, and L channels. And, the 2D network extracts deep classification-aware features from the original image. Thirdly, an effective feature fusion method is designed to descript the characteristics of different chrysanthemums more efficiently, which take both inter-layer interaction and inter-path interaction into consideration. Finally, all the features extracted from different streams of the network are fused to gain the deep color space feature. The performance comparisons are conducted on the dataset which contain 4276 real chrysanthemum images of 18 categories. Experimental results demonstrate that the proposed method is more accurate and stable than tested methods, achieving an accuracy of 95.45%, which is approximately 2.07% higher than the best tested method.
{"title":"Chrysanthemum classification method via multi-stream deep color space feature fusion","authors":"Jian Jiang , Xichen Yang , Hui Yan , Jia Liu , Yifan Chen , Zhongyuan Mao , Tianshu Wang","doi":"10.1016/j.compag.2026.111455","DOIUrl":"10.1016/j.compag.2026.111455","url":null,"abstract":"<div><div>Chrysanthemum is a traditional Chinese medicinal herb that is widely employed in various medical applications. The medicinal value of chrysanthemums varies among different types, and this value is highly correlated with their specific classification. Therefore, the classification of chrysanthemums is essential for ensuring their proper medicinal use. However, existing traditional classification methods are often time-consuming and costly, making them less suitable for practical applications. To overcome these limitations, a novel chrysanthemum classification method based on multi-stream deep color space feature fusion is proposed. Firstly, chrysanthemum images are transformed from the RGB color space to the HSV and LAB color spaces. And the classification-aware features are extracted from the H, S, and L channels, respectively. Secondly, a multi-stream deep network is designed, which employing both 1D and 2D networks. The 1D network focus on further analyzing the features from the H, S, and L channels. And, the 2D network extracts deep classification-aware features from the original image. Thirdly, an effective feature fusion method is designed to descript the characteristics of different chrysanthemums more efficiently, which take both inter-layer interaction and inter-path interaction into consideration. Finally, all the features extracted from different streams of the network are fused to gain the deep color space feature. The performance comparisons are conducted on the dataset which contain 4276 real chrysanthemum images of 18 categories. Experimental results demonstrate that the proposed method is more accurate and stable than tested methods, achieving an accuracy of 95.45%, which is approximately 2.07% higher than the best tested method.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111455"},"PeriodicalIF":8.9,"publicationDate":"2026-01-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025084","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-17DOI: 10.1016/j.compag.2026.111431
Mengjie Ye , Yong Cheng , De Yu , Yongqi Yuan , Ge Jin , Huan Wang
Effective feature representation and heterogeneous fusion are essential for plant leaf recognition. However, existing methods have several limitations, such as insufficient comprehensiveness and distinctiveness in feature representation, as well as a lack of full consideration for the compatibility and complementarity in heterogeneous fusion. In the end, we propose a discriminative shape representation named the bag of multiscale curvature angle cuts (BMCAC) to capture fine curvature and spatial distribution characteristics, an advanced deep representation called the progressive salient deep representation (PSDR) to fully exploit deep convolutional features, and an effective fusion framework termed the K-weighted shape and deep feature fusion (KWFF) to aggregate the local context and global importance of heterogeneous features. Specifically, BMCAC is derived from the curvature angle cuts (CAC), multiscale analysis, and the bag of visual words (BoVW) model; PSDR is constructed by applying progressive downsampling and hierarchical pooling operations to deep convolutional features; and KWFF is developed by encoding neighboring information using homogeneous distance measures while incorporating globally weighted contributions from heterogeneous distance measures. Extensive experiments on four well-known benchmark leaf datasets demonstrate that the proposed shape and deep representations can efficiently extract leaf image features, and the fusion framework can effectively integrate heterogeneous features, outperforming state-of-the-art methods. The source code is available at https://github.com/Mumuxi1123/BMCAC_PSDR_KWFF.
{"title":"Discriminative feature representations and heterogeneous fusion for plant leaf recognition","authors":"Mengjie Ye , Yong Cheng , De Yu , Yongqi Yuan , Ge Jin , Huan Wang","doi":"10.1016/j.compag.2026.111431","DOIUrl":"10.1016/j.compag.2026.111431","url":null,"abstract":"<div><div>Effective feature representation and heterogeneous fusion are essential for plant leaf recognition. However, existing methods have several limitations, such as insufficient comprehensiveness and distinctiveness in feature representation, as well as a lack of full consideration for the compatibility and complementarity in heterogeneous fusion. In the end, we propose a discriminative shape representation named the bag of multiscale curvature angle cuts (BMCAC) to capture fine curvature and spatial distribution characteristics, an advanced deep representation called the progressive salient deep representation (PSDR) to fully exploit deep convolutional features, and an effective fusion framework termed the K-weighted shape and deep feature fusion (KWFF) to aggregate the local context and global importance of heterogeneous features. Specifically, BMCAC is derived from the curvature angle cuts (CAC), multiscale analysis, and the bag of visual words (BoVW) model; PSDR is constructed by applying progressive downsampling and hierarchical pooling operations to deep convolutional features; and KWFF is developed by encoding neighboring information using homogeneous distance measures while incorporating globally weighted contributions from heterogeneous distance measures. Extensive experiments on four well-known benchmark leaf datasets demonstrate that the proposed shape and deep representations can efficiently extract leaf image features, and the fusion framework can effectively integrate heterogeneous features, outperforming state-of-the-art methods. The source code is available at <span><span>https://github.com/Mumuxi1123/BMCAC_PSDR_KWFF</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111431"},"PeriodicalIF":8.9,"publicationDate":"2026-01-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025083","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号