Computers and Electronics in Agriculture最新文献

{"title":"Capped honey segmentation in honey combs based on deep learning approach","authors":"Francisco J. Rodriguez-Lozano ,&nbsp;Sergio R. Geninatti ,&nbsp;José M. Flores ,&nbsp;Francisco J. Quiles-Latorre ,&nbsp;Manuel Ortiz-Lopez","doi":"10.1016/j.compag.2024.109573","DOIUrl":"10.1016/j.compag.2024.109573","url":null,"abstract":"<div><div>Honey is the food stored by honey bees for periods when it is scarce in the field as well as being a product that is consumed worldwide by humans. Each hive generates different amounts of honey depending on the population of the bee hive, health state or environmental factors. In fact, the reserves of honey provide beekeepers with a double function: to predict the amount of honey that can be obtained and to analyze the state of the bee colonies. The assessment of honey reserves is commonplace in scientific research related to the health of bee colonies, genetic improvement or environmental issues, and emerging technologies can provide useful tools to evaluate honey stored in hives. In this context, this work proposes a methodology to detect the honey areas in high resolution photographs automatically using methods based on deep learning. Specifically, the methodology follows a <em>“divide and conquer”</em> approach where the images are separated into tiles with overlapping areas that are used by a semantic segmentation algorithm based on Feature Pyramid Network (FPN), detecting the honey in each tile to finally merge the tiles back into the complete image. The proposal has been compared with different feature extractors (backbones) and other semantic segmentation models, obtaining on average accurate results above 90% and 87% in the Dice score and IOU metrics respectively.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"227 ","pages":"Article 109573"},"PeriodicalIF":7.7,"publicationDate":"2024-11-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142571484","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}

{"title":"Multi-modal machine learning for the early detection of metabolic disorder in dairy cows using a cloud computing framework","authors":"Rafael E.P. Ferreira ,&nbsp;Maria Angels de Luis Balaguer ,&nbsp;Tiago Bresolin ,&nbsp;Ranveer Chandra ,&nbsp;Guilherme J.M. Rosa ,&nbsp;Heather M. White ,&nbsp;João R.R. Dórea","doi":"10.1016/j.compag.2024.109563","DOIUrl":"10.1016/j.compag.2024.109563","url":null,"abstract":"<div><div>In precision livestock farming (PLF), wearable sensors, computer vision, and genomic tests generate large amounts of data, which can be challenging to integrate and analyze jointly due to their diverse nature. However, incorporating both genomic and phenotypic data together can be beneficial for developing predictive models in animal biology. The development of automated and modular data pipelines using scalable solutions such as cloud computing can be an effective strategy to integrate and analyze animal-level information in real-time. The objectives of this study were (1) to propose a cloud computing-based framework to automate the processing and integration of phenotypic and genotypic data, and (2) to assess different data fusion strategies (early and late fusion, and cooperative learning) for the early detection of subclinical ketosis (SCK) in dairy cows, integrating wearable sensors, imaging systems, and genotypic data in livestock farms. We developed a modular pipeline for image analysis, which includes body segmentation, frame quality assessment, animal identification, and body condition score (BCS), which were crucial for producing the features used in SCK detection. The body segmentation module achieved a Dice similarity coefficient of 0.990, the frame quality assessment module reached 99.1 % accuracy, the animal identification module attained 93.2 % accuracy, and the BCS module achieved accuracies of 81.1 % and 96.2 % when allowing up to 0.25 and 0.50 prediction error, respectively. For SCK detection, early fusion and cooperative learning achieved the lowest mean absolute errors in predicting plasma beta-hydroxybutyrate as a continuous variable (as low as 0.242). Late fusion, combined with an ordinary least squares regression, achieved the highest F₁ scores for binary SCK prediction (up to 0.750). These results suggest that data fusion techniques can be effectively used to integrate genotypic and phenotypic data from multiple sensors. Additionally, SCK detection can be performed on dairy farms using the proposed cloud computing-based framework, which was implemented with modular, independent services that can be customized and reused for a variety of tasks.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"227 ","pages":"Article 109563"},"PeriodicalIF":7.7,"publicationDate":"2024-11-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142571485","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}

{"title":"Ground-based on-line weed control using computer vision: Analyzing the inference time-accuracy dilemma","authors":"Saad Abouzahir ,&nbsp;Essaid Sabir ,&nbsp;Mohamed Sadik ,&nbsp;Mohamed Abouzahir ,&nbsp;Fouad Agramelal","doi":"10.1016/j.compag.2024.109577","DOIUrl":"10.1016/j.compag.2024.109577","url":null,"abstract":"<div><div>The detrimental effects of weeds on crop growth and yield present substantial challenges to the agribusiness sector, necessitating the deployment of robust control strategies. The rapid advancement of Computer vision (CV) techniques has driven the integration of ground-based imaging sensors to enable site-specific weed management. The main challenge in weed management revolves around on-line weed detection, which demands a careful balance between inference time and detection accuracy. Finding this balance is very important, as prioritizing a higher number of frames per second (fps) might reduce the detection precision. However, the real-time constraint for on-line weed control remains relatively unexplored. This paper addresses this gap by categorizing proposed approaches based on ground-vehicle configuration and evaluating the real-time requirements for on-line weed control. We comprehensively examine the different components of ground-vehicles including the travel speed, camera settings, and weeding tools to understand the fps required for seamless weed control operation. Results show that for travel speeds below <span><math><mrow><mn>4</mn><mspace></mspace><mi>k</mi><mi>m</mi><mspace></mspace><msup><mrow><mi>h</mi></mrow><mrow><mo>−</mo><mn>1</mn></mrow></msup></mrow></math></span>, Deep Neural Networks (DNNs) operating at fps rates lower than 10 Hz are suitable for effective on-line weed detection. However, at higher speeds or with smaller Fields of View, fps demands increase. Our findings further reveal that the relatively relaxed fps requirements of on-line weed control create opportunities for deploying larger DNNs, such as NASNet-A-Large, which can significantly enhance detection accuracy. The operational latency introduced by certain weeding tools further provides additional processing time for DNNs. The continuous advancement of larger DNNs and improvements in hardware offer promising prospects for precise and effective weed management. Future research should leverage these developments, focusing on enhancing detection accuracy rather than optimizing for faster inference times, given the relaxed real-time constraints of ground-based weed control.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"227 ","pages":"Article 109577"},"PeriodicalIF":7.7,"publicationDate":"2024-11-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142571486","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}

{"title":"Tea bud DG: A lightweight tea bud detection model based on dynamic detection head and adaptive loss function","authors":"Lu Jianqiang ,&nbsp;Luo Haoxuan ,&nbsp;Yu Chaoran ,&nbsp;Liang Xiao ,&nbsp;Huang Jiewei ,&nbsp;Wu Haiwei ,&nbsp;Wang Liang ,&nbsp;Yang Caijuan","doi":"10.1016/j.compag.2024.109522","DOIUrl":"10.1016/j.compag.2024.109522","url":null,"abstract":"<div><div>Tea bud detection plays a crucial role in early-stage tea production estimation and robotic harvesting, significantly advancing the integration of computer vision and agriculture. Currently, tea bud detection faces several challenges such as reduced accuracy due to high background similarity, and the large size and parameter count of the models, which hinder deployment on mobile devices. To address these issues, this study introduces the lightweight Tea Bud DG model, characterized by the following features: 1) The model employs a Dynamic Head (DyHead), which enhances tea bud feature extraction through three types of perceptual attention mechanisms—scale, spatial, and task awareness. Scale awareness enables the model to adapt to objects of varying sizes; spatial awareness focuses on discriminative regions to distinguish tea buds against complex backgrounds; task awareness optimizes feature channels for specific tasks, such as classification or localization of tea buds. 2) A lightweight C3ghost module is designed, initially generating basic feature maps with fewer filters, followed by simple linear operations (e.g., translation or rotation) to create additional “ghost” feature maps, thus reducing the parameter count and model size, facilitating deployment on lightweight mobile devices. 3) By introducing the α-CIoU loss function with the parameter α, the loss and gradient of objects with different IoU scores can be adaptively reweighted by adjusting the α parameter. This approach emphasizes objects with higher IoU, enhancing the ability to identify tea buds in environments with high background similarity. The use of α-CIoU focuses on accurately differentiating tea buds from surrounding leaves, improving detection performance. The experimental results show that compared with YOLOv5s, the Tea Bud DG model reduces the model size by 31.41 % and the number of parameters by 32.21 %. Compared with YOLOv7_tiny, the size and parameters are reduced by 18.94 % and 23.84 %, respectively. It achieved improvements in [email protected] by 3 %, 3.9 %, and 5.1 %, and in [email protected]_0.95 by 2.6 %, 3.2 %, and 4 % compared with YOLOv5s, YOLOv8s, and YOLOv9s, respectively. The Tea Bud DG model estimates the tea yield with an error range of 10 % to 16 %, providing valuable data support for tea plantation management.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"227 ","pages":"Article 109522"},"PeriodicalIF":7.7,"publicationDate":"2024-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142571283","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}

{"title":"A grapevine trunks and intra-plant weeds segmentation method based on improved Deeplabv3 Plus","authors":"Shuming Yang ,&nbsp;Zheng Cui ,&nbsp;Maoqiang Li ,&nbsp;Jinhai Li ,&nbsp;Dehua Gao ,&nbsp;Fulong Ma ,&nbsp;Yutan Wang","doi":"10.1016/j.compag.2024.109568","DOIUrl":"10.1016/j.compag.2024.109568","url":null,"abstract":"<div><div>Accurate identification of grapevine trunks and interplant weeds is crucial for the intelligent development of weeding sessions in vineyards. Challenges arise due to the nonuniform planting of wine grapes, obscuration of grapevine trunks by interplant weeds, and variations in trunk characteristics across different growth stages, complicating the accurate segmentation of grapevine trunks and intraplant weeds. This study presents a new identification model that employs an improved Deeplabv3 Plus framework with lightweight Mobilenetv2 as its central network, supplemented by a coordinate attention block to boost feature extraction capabilities. The model was deployed using the robot operating system (ROS) on a crawler robot for field operations. We developed datasets for grapevine trunks and intraplant weeds, and upon training and testing, the model achieved a mean intersection over union (MIoU) of 84.4 % and a pixel accuracy of 92.03 %. Field trials integrating the ROS system demonstrated a grapevine trunk miss detection rate of 3.6 %, a false detection rate of 2.4 %, and a detection speed of 22 frames per second (FPS). The results show that our method effectively balances recognition accuracy and speed, offering valuable technical support for developing intelligent field weeders for wine grape cultivation.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"227 ","pages":"Article 109568"},"PeriodicalIF":7.7,"publicationDate":"2024-10-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142552658","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}

{"title":"Determination and quantitative evaluation of early postharvest hidden damage in fresh strawberry fruit based on coupling of dynamic finite element method and response surface methodology","authors":"Jie Guo ,&nbsp;Yufei Liu ,&nbsp;Manoj Karkee ,&nbsp;Xuping Feng ,&nbsp;Zichen Huang ,&nbsp;Yuwei Wang ,&nbsp;Wenkai Zhang ,&nbsp;Xiaoli Li ,&nbsp;Yong He","doi":"10.1016/j.compag.2024.109588","DOIUrl":"10.1016/j.compag.2024.109588","url":null,"abstract":"<div><div>Collision damage is the most common type of damage during the mechanized harvesting, stacking and transportation of strawberries. Aiming at the problem that hidden bruises on fruits caused by collision behaviors are difficult to detect and accurately quantify in the early stage of damage, this paper carried out simulation and experimental research on the bruise susceptibility of strawberry fruits at the moment of collision based on dynamic finite element method and response surface method. By measuring the physical characteristics parameters of three different varieties of strawberry fruits, the three-dimensional solid model of the fruit including cortex, central pith, and achene was established. The multi-scale finite element model of the fruit was further established based on the mechanical property parameters obtained based on the quasi-static compression experiment. A total of 240 different experimental scenarios were set up in this paper, and the cloud diagrams of fruit’s equivalent stress and the changing law of system energy under different conditions were obtained. The experimental results showed that the equivalent stress and contact force inside the fruit vary due to differences in contact material, drop height and impact angle. In order to further obtain the comprehensive effects of contact material, drop height and impact angle on bruise susceptibility, four empirical models for predicting bruise susceptibility were established by using response surface methodology. By comparing the experimental results with the predicted results of the model, it was found that under the conditions where the contact material was steel and the drop height was 1 m, the relative error between the measured value and the predicted value was the smallest (1.38 %) when the impact angle was −19°; the relative error between the measured value and the predicted value was the largest (6.43 %) when the impact angle was −31°. The results of this study showed that the predicting models of strawberry fruit’s bruise susceptibility based on response surface methodology were reasonable and correct. These models can be used to determine the potential mechanical damage of strawberry fruits during mechanized harvesting, stacking and transportation, and can provide a basis for the development of end-effectors/manipulators in strawberry or other fruit and vegetable picking robots, the formulation of picking/harvesting strategies, and the design of packaging container structures.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"227 ","pages":"Article 109588"},"PeriodicalIF":7.7,"publicationDate":"2024-10-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142561128","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}

{"title":"Nested sequential feed-forward neural network: A cumulative model for crop yield prediction","authors":"N. Andy Kundang Chang ,&nbsp;Shouvik Dey ,&nbsp;Dushmanta Kumar Das","doi":"10.1016/j.compag.2024.109562","DOIUrl":"10.1016/j.compag.2024.109562","url":null,"abstract":"<div><div>This paper contends that framing crop yield prediction as a time-series problem imposes significant limitations. The varying climatic conditions, along with the distinct time frames associated with different stages of crop cultivation – such as sowing, vegetation growth, flowering, and harvest – present substantial challenges for accurately predicting crop yields. Additionally, the evolving climatic conditions over the years further complicate the prediction process. To address these challenges, this study introduces a novel perspective termed the ’Time-Dimension Based (TDB) Problem,’ offering a conceptual framework that redefines how crop yield prediction should be approached. The TDB framework guides the modeling architecture into two layers: one for capturing the varying climatic conditions and the other for accumulating their impact on crops to determine the final yield. To implement this approach, the paper introduces the ”Nested Sequential Feed-Forward Neural Network (NSFFNet),” a novel neural network architecture. NSFFNet features key components, including an innovative ’Nested Sequential Feed-Forwarding of Inputs’ using feed-forward neural network for capturing Earth’s climatic patterns over time, and a ’Neural Cache Layer’ that utilizes cache memory to accumulate the cumulative impact of these patterns on crop yield. To validate this approach, a comprehensive evaluation of NSFFNet was conducted against traditional time-series models. The model was assessed for accuracy, generalizability, and robustness, particularly in estimating yields during drought years. NSFFNet consistently outperforms established models like RNN, 1D CNN, LSTM, GRU, and Transformer. These findings suggest that redefining crop yield prediction as a TDB problem is a highly effective strategy.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"227 ","pages":"Article 109562"},"PeriodicalIF":7.7,"publicationDate":"2024-10-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142561125","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}

{"title":"Integrating remote sensing assimilation and SCE-UA to construct a grid-by-grid spatialized crop model can dramatically improve winter wheat yield estimate accuracy","authors":"Qiang Li ,&nbsp;Maofang Gao ,&nbsp;Sibo Duan ,&nbsp;Guijun Yang ,&nbsp;Zhao-Liang Li","doi":"10.1016/j.compag.2024.109594","DOIUrl":"10.1016/j.compag.2024.109594","url":null,"abstract":"<div><div>Grain yield estimation remains a focal point in agricultural research. It’s well known that crop models have very high accuracy in field application, but their scalability to a regional level encounters formidable constraints attributed to stringent input parameter demands, challenges in data acquisition, and complexities in parameter calibration. In a concerted effort to overcome these aforementioned challenges, this study endevours to formulate a spatialized crop growth model, organized grid by grid, propelled by a myriad of data sources encompassing diverse remote sensing and statistical inputs. Our approach involves the integration of a machine learning technique—the shuffled complex evolution algorithm (SCE-UA) to propose an automatic parameter optimization method for model calibration, alongside two remote sensing assimilation methods: a four-dimensional variational assimilation algorithm (4Dvar) and ensemble Kalman filter (Enkf) to optimising model trajectories to improve crop yield estimation accuracy. This innovative methodology addresses the intricacies associated with regional-scale simulation and bridges the gap between the inherent limitations of conventional crop models and the demand for high-precision yield estimations. The results show that: (1) we improved the accuracy of the regional crop model from 0.53 to 0.94 for the coefficient of determination (R²) and from 824.82 kg/ha to 148.48 kg/ha for root mean square error (RMSE), which greatly improved the accuracy of winter wheat yield estimation; (2) after comparing different optimization and assimilation strategies, the simulation strategy of complex shuffling algorithm (SCE-UA) combined with the four-dimensional variational algorithm (4Dvar) can enable the grid-by-grid model to estimate yield to achieve the highest simulation accuracy, with R² of 0.94 and RMSE of 148.48 kg/ha; (3) we evaluated the simulation effectiveness of the algorithm and discuss the shortcomings and uncertainties of the grid-by-grid model. This study has important practical implications for the development of spatialized models for estimating winter wheat yields and bolstering our capacity for informed decision-making in the realm of food production and agricultural management.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"227 ","pages":"Article 109594"},"PeriodicalIF":7.7,"publicationDate":"2024-10-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142552660","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}

{"title":"Generalized few-shot learning for crop hyperspectral image precise classification","authors":"Hao-tian Yuan ,&nbsp;Ke-kun Huang ,&nbsp;Jie-li Duan ,&nbsp;Li-qian Lai ,&nbsp;Jia-xiang Yu ,&nbsp;Chao-wei Huang ,&nbsp;Zhou Yang","doi":"10.1016/j.compag.2024.109498","DOIUrl":"10.1016/j.compag.2024.109498","url":null,"abstract":"<div><div>Hyperspectral remote sensing technology, with its advantage of acquiring a substantial amount of spectral information across different bands, has provided a robust tool for crop monitoring and management in the agricultural field. However, a prevalent challenge persists, namely the limited number of labels required for crop classification. We propose a new method named Generalized Few-Shot Learning (GFSL) to address the small-sample problem and get better classification of crops. The proposed GFSL first maps the embedding features extracted by a convolutional neural network to a Hilbert space by an implicit nonlinear mapping with a kernel trick. Then, GFSL maximizes the kernel similarity between each sample and its class mean as much as possible, and minimizes the kernel similarity between each sample and the means of other classes as much as possible at the same time. To give a more meaningful balance between intra-class similarity and inter-class similarity, GFSL defines the negative of intra-class similarity plus the logarithm of the sum of exponential functions of inter-class similarities as the loss function. We conducted experiments on three publicly available crop hyperspectral datasets: WHU-Hi-HanChuan, Salinas, and Indian Pines, and results show that the proposed approach exhibits an improvement in classification accuracy of 11.46%, 6.86%, and 14.49% on the three datasets, respectively, in comparison to some state-of-the-art methods, which demonstrates the superiority of the proposed method for crop hyperspectral image classification with limited training samples. The Python source code is available at <span><span>https://github.com/kkcocoon/GFSL</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"227 ","pages":"Article 109498"},"PeriodicalIF":7.7,"publicationDate":"2024-10-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142561126","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}

{"title":"Visual large language model for wheat disease diagnosis in the wild","authors":"Kunpeng Zhang ,&nbsp;Li Ma ,&nbsp;Beibei Cui ,&nbsp;Xin Li ,&nbsp;Boqiang Zhang ,&nbsp;Na Xie","doi":"10.1016/j.compag.2024.109587","DOIUrl":"10.1016/j.compag.2024.109587","url":null,"abstract":"<div><div>Early detection of symptoms in wheat plants is crucial for mitigating disease effects and preventing their spread. Prompt phytosanitary treatment minimizes yield losses and enhances treatment efficacy. In recent years, numerous image analysis-based methodologies for automatic disease identification have been developed, with Convolutional Neural Networks (CNNs) achieving notable success in visual classification tasks. The existing methods often lack the necessary intelligence and reasoning for real-world applications. This study introduces an advanced wheat disease diagnosis approach using a Visual Language Model (VLM), named the Wheat Disease Language Model (WDLM). The WDLM first leverages the modified Segment Anything Model (SAM) to isolate key wheat features from complex wild environments. To enhance the logical reasoning abilities, the WDLM integrates a reasoning chain to generate clear, reasoned explanations for its diagnosis. By employing dedicated prompt engineering, this study establishes the Wheat Disease Semantic Dataset (WDSD) to fine-tune the VLM. The WDSD, which includes a diverse set of wheat images from various sources, bridges the gap between advanced VLM technology and wheat pathology. Tailored with task-specific data, the WDLM demonstrates superior intelligence by providing accurate classification of wheat diseases and suggesting potential treatment options. Compared to CNN-based models, Transformer-based models, and other VLMs, the WDLM shows improved performance in various scenarios. Integrated with mobile applications, the WDLM approach is readily applicable in the field, representing a promising advancement in the intelligent diagnosis of wheat diseases.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"227 ","pages":"Article 109587"},"PeriodicalIF":7.7,"publicationDate":"2024-10-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142552659","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}