Pub Date : 2026-01-28DOI: 10.1016/j.compag.2026.111472
Xiangyu Shi , Hongzhou Chen , Peipei Gao , Xu Shen , Xiaobo Zou , Jiyong Shi
Transportation-induced mortality from crowding, hypoxia, and handling stress costs aquaculture billions annually. Although anesthetic protocols mitigate these stressors, the anesthetic recovery phase remains critical for timely interventions. Invasive bioimpedance sensors introduce infection risks, computer vision fails in turbid water. Acoustic techniques demonstrate advantages yet face inadequate feature representation and severe dataset scarcity. To address these interdependent challenges confronting acoustic monitoring, this work presents a novel framework combining generative artificial intelligence (GAI) synthesis with multi-domain fusion for recovery stage discrimination. Specifically, Domain-Specific Denoising Diffusion Probabilistic Models (DS-DDPM), synthesize biologically-authentic features across Zero-Crossing Rate (ZCR), Gammatone, Phase, and Wavelet domains, addressing data scarcity while preserving physiological validity. Subsequently, an adaptive progressive fusion strategy organizes features into periodicity and intensity groups, integrating complementary respiratory patterns through attention mechanisms. The lightweight MambaVision-MDF (Multi-Domain Fusion) architecture then processes the augmented fused features through dual-path temporal-spectral scanning mechanism for discrimination. Experimental validation on bighead carp demonstrates that adaptive progressive fusion and DS-DDPM augmentation improve classification accuracy by 2.1 and 3.5 percentage points over direct four-domain combination and baseline training respectively. This advancement achieves 96.8% accuracy with only 5.83 M parameters for edge deployment, facilitating temporal-state dual-driven strategies for smart aquaculture transportation management. Meanwhile, it establishes a replicable paradigm for GAI applications addressing insufficient representation and data constraints in agricultural monitoring.
{"title":"Generative AI-enhanced multi-domain acoustic fusion for real-time recovery monitoring in smart aquaculture transportation","authors":"Xiangyu Shi , Hongzhou Chen , Peipei Gao , Xu Shen , Xiaobo Zou , Jiyong Shi","doi":"10.1016/j.compag.2026.111472","DOIUrl":"10.1016/j.compag.2026.111472","url":null,"abstract":"<div><div>Transportation-induced mortality from crowding, hypoxia, and handling stress costs aquaculture billions annually. Although anesthetic protocols mitigate these stressors, the anesthetic recovery phase remains critical for timely interventions. Invasive bioimpedance sensors introduce infection risks, computer vision fails in turbid water. Acoustic techniques demonstrate advantages yet face inadequate feature representation and severe dataset scarcity. To address these interdependent challenges confronting acoustic monitoring, this work presents a novel framework combining generative artificial intelligence (GAI) synthesis with multi-domain fusion for recovery stage discrimination. Specifically, Domain-Specific Denoising Diffusion Probabilistic Models (DS-DDPM), synthesize biologically-authentic features across Zero-Crossing Rate (ZCR), Gammatone, Phase, and Wavelet domains, addressing data scarcity while preserving physiological validity. Subsequently, an adaptive progressive fusion strategy organizes features into periodicity and intensity groups, integrating complementary respiratory patterns through attention mechanisms. The lightweight MambaVision-MDF (Multi-Domain Fusion) architecture then processes the augmented fused features through dual-path temporal-spectral scanning mechanism for discrimination. Experimental validation on bighead carp demonstrates that adaptive progressive fusion and DS-DDPM augmentation improve classification accuracy by 2.1 and 3.5 percentage points over direct four-domain combination and baseline training respectively. This advancement achieves 96.8% accuracy with only 5.83 M parameters for edge deployment, facilitating temporal-state dual-driven strategies for smart aquaculture transportation management. Meanwhile, it establishes a replicable paradigm for GAI applications addressing insufficient representation and data constraints in agricultural monitoring.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111472"},"PeriodicalIF":8.9,"publicationDate":"2026-01-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080285","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 : 2026-01-27DOI: 10.1016/j.compag.2026.111473
Xuanyue Tong , Pute Wu , Lin Zhang , Xufei Liu , Shoujun Wu
Accurately calculating the flow rate of the emitters for subsurface infiltration irrigation is essential to match crop water requirements and provide a suitable water amount. However, drip emitters usually have a single flow path and are assumed to have a constant flow rate, which does not apply to subsurface infiltration irrigation represented by ceramic emitters (CEs). Therefore, a flow model was developed to accurately calculate the flow rate of CEs considering ceramic pore flow channels and soil types. Five experiment sites were selected for validation, with the average values of MAE, RMSE, and R2 being 0.0064, 0.0095 and 0.8138, respectively. Meanwhile, the relationship between different soil types, porosities, and working water heads on the flow rate of CEs was explored using a dataset with 403 points of all soil types. Results show that the flow rate of CEs increased with the sand content, and the average flow rates of CEs in sand, silt, loam, and clay soil were 0.0525, 0.0453, 0.0257, and 0.003 L h−1, respectively. Furthermore, the flow rate of CEs in more sandy soil was significantly affected by the porosity, which in clay soil was mainly determined by soil type. In addition, the influence of the working water head on the coefficient of variation (CV) for the flow rate of CE diminished significantly beyond approximately 74 cm. The model can accurately calculate the flow rate of CEs in practice, which is conducive to the practical application of subsurface infiltration irrigation systems and the efficient utilization of irrigation water.
准确计算地下渗灌灌水器的流量是满足作物需水量和提供适宜水量的关键。然而，滴灌器通常具有单一的流路，并且假定具有恒定的流量，这并不适用于以陶瓷滴灌器（CEs）为代表的地下渗灌。因此，建立了考虑陶瓷孔流通道和土壤类型的流动模型，以准确计算ce的流量。选取5个实验点进行验证，MAE、RMSE和R2的平均值分别为0.0064、0.0095和0.8138。同时，利用所有土壤类型403个点的数据集，探讨了不同土壤类型、孔隙度和工作水头对ce流量的关系。结果表明，随着含砂量的增加，碳水化合物的流量逐渐增大，在砂土、粉土、壤土和粘土中碳水化合物的平均流量分别为0.0525、0.0453、0.0257和0.003 L h−1。此外，在砂质较多的土壤中，碳碳化合物的流动速率受孔隙度的显著影响，而在粘土中，孔隙度主要由土壤类型决定。此外，工作水头对CE流量变异系数（CV）的影响在约74 cm以上显著减小。该模型能准确地计算出实际中ce的流量，有利于地下渗灌系统的实际应用和灌溉水的有效利用。
{"title":"Calculating flow rate for ceramic emitters in subsurface infiltration irrigation under various soil types based on fractal capillary bundle model","authors":"Xuanyue Tong , Pute Wu , Lin Zhang , Xufei Liu , Shoujun Wu","doi":"10.1016/j.compag.2026.111473","DOIUrl":"10.1016/j.compag.2026.111473","url":null,"abstract":"<div><div>Accurately calculating the flow rate of the emitters for subsurface infiltration irrigation is essential to match crop water requirements and provide a suitable water amount. However, drip emitters usually have a single flow path and are assumed to have a constant flow rate, which does not apply to subsurface infiltration irrigation represented by ceramic emitters (CEs). Therefore, a flow model was developed to accurately calculate the flow rate of CEs considering ceramic pore flow channels and soil types. Five experiment sites were selected for validation, with the average values of <em>MAE</em>, <em>RMSE</em>, and <em>R<sup>2</sup></em> being 0.0064, 0.0095 and 0.8138, respectively. Meanwhile, the relationship between different soil types, porosities, and working water heads on the flow rate of CEs was explored using a dataset with 403 points of all soil types. Results show that the flow rate of CEs increased with the sand content, and the average flow rates of CEs in sand, silt, loam, and clay soil were 0.0525, 0.0453, 0.0257, and 0.003 L h<sup>−1</sup>, respectively. Furthermore, the flow rate of CEs in more sandy soil was significantly affected by the porosity, which in clay soil was mainly determined by soil type. In addition, the influence of the working water head on the coefficient of variation (CV) for the flow rate of CE diminished significantly beyond approximately 74 cm. The model can accurately calculate the flow rate of CEs in practice, which is conducive to the practical application of subsurface infiltration irrigation systems and the efficient utilization of irrigation water.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111473"},"PeriodicalIF":8.9,"publicationDate":"2026-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079955","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 : 2026-01-27DOI: 10.1016/j.compag.2026.111440
Yanzhe Hu , Shaozhong Kang , Risheng Ding , Herman N.C. Berghuijs , Iris Vogeler , Yuan Qiu , Leonardo A. Monteiro , Marcos Lana
Irrigation is expected to playing a pivotal role against climate change and cropping systems intensification, whilst the secondary soil salinization caused by imprecise irrigation is posing a serious challenge to crop production. Despite increasing attention has been paid to food crops, a more profound understanding of water deficit and soil salinity constraints on forage production is greatly desired, and the response of forage growth to enhanced water stress raised by salinity needs to be considered in crop models. For this purpose, the water absorption module in the APSIM-Lucerne model was extended with two modules that calculate the reduction of the water extraction coefficient (KL) from the chloride concentration (Cl) in soil to enable the simulation of inhibited plant growth under enhanced water stress due to soil salinity. Both modules assume that KL decreases with Cl above a threshold Cl. In the first module, the decrease is exponential (exponential KL modifier), whilst in the second module, KL decreases according to a power law (power KL modifier) until it reaches zero at another higher threshold Cl. In field experiments, soil water content, leaf area index and biomass were measured for alfalfa grown under different combinations of irrigation amounts and salinity levels. The performance of the modified models (exponential and power KL modifiers) and the original model (no KL modifier) to reproduce these data were compared. Results reveal that both modified models showed improved prediction of canopy development and biomass accumulation, while the modified model with the power KL modifier exhibited a comparatively higher predictability under high salinity level, with a relative root mean square error of 23%-27% for biomass, better than 24%-31% of the exponential model and 43%-45% of the original model. The soil water dynamics were not well predicted by the modified models due to an underestimation of soil evaporation which requires further investigation. The study improved the predictability of crop models for forage crop development and production under coupling soil water and salt stresses via the optimization of the dynamic plant water extraction process, thus can be used to chart more reliable irrigation strategies under various pedoclimatic conditions.
{"title":"Modifying water absorption process to enhance model performance on biomass accumulation under soil water and salt stresses","authors":"Yanzhe Hu , Shaozhong Kang , Risheng Ding , Herman N.C. Berghuijs , Iris Vogeler , Yuan Qiu , Leonardo A. Monteiro , Marcos Lana","doi":"10.1016/j.compag.2026.111440","DOIUrl":"10.1016/j.compag.2026.111440","url":null,"abstract":"<div><div>Irrigation is expected to playing a pivotal role against climate change and cropping systems intensification, whilst the secondary soil salinization caused by imprecise irrigation is posing a serious challenge to crop production. Despite increasing attention has been paid to food crops, a more profound understanding of water deficit and soil salinity constraints on forage production is greatly desired, and the response of forage growth to enhanced water stress raised by salinity needs to be considered in crop models. For this purpose, the water absorption module in the APSIM-Lucerne model was extended with two modules that calculate the reduction of the water extraction coefficient (<em>KL</em>) from the chloride concentration (<em>Cl</em>) in soil to enable the simulation of inhibited plant growth under enhanced water stress due to soil salinity. Both modules assume that <em>KL</em> decreases with <em>Cl</em> above a threshold <em>Cl</em>. In the first module, the decrease is exponential (exponential <em>KL</em> modifier), whilst in the second module, <em>KL</em> decreases according to a power law (power <em>KL</em> modifier) until it reaches zero at another higher threshold <em>Cl</em>. In field experiments, soil water content, leaf area index and biomass were measured for alfalfa grown under different combinations of irrigation amounts and salinity levels. The performance of the modified models (exponential and power <em>KL</em> modifiers) and the original model (no <em>KL</em> modifier) to reproduce these data were compared. Results reveal that both modified models showed improved prediction of canopy development and biomass accumulation, while the modified model with the power <em>KL</em> modifier exhibited a comparatively higher predictability under high salinity level, with a relative root mean square error of 23%-27% for biomass, better than 24%-31% of the exponential model and 43%-45% of the original model. The soil water dynamics were not well predicted by the modified models due to an underestimation of soil evaporation which requires further investigation. The study improved the predictability of crop models for forage crop development and production under coupling soil water and salt stresses via the optimization of the dynamic plant water extraction process, thus can be used to chart more reliable irrigation strategies under various pedoclimatic conditions.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111440"},"PeriodicalIF":8.9,"publicationDate":"2026-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079801","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 : 2026-01-27DOI: 10.1016/j.compag.2026.111456
Qixin Kang , Guozhong Zhang , Chao Ji , Zhuangzhuang Zhao , Liming Chen , Lingxiao Lan
<div><div><em>Allium chinense</em> is a specialty crop valued for its distinctive flavor and medicinal properties, with substantial exports to Japan, South Korea, and international markets. Despite its economic importance, cultivation remains labor-intensive due to strict agronomic demands for precise seed orientation. A key obstacle to mechanization is collision-induced bouncing during sowing, which compromises placement quality. This study investigates the effects of seed-discharge parameters and seedbed geometry on directional sowing by analyzing seed–soil collision mechanisms. DEM (Discrete Element Method) models based on HM-JKR (Hertz-Mindlin with JKR) model were constructed, with parameters calibrated via P-B (Plackett-Burman) screening and B-B (Box-Behnken) optimization. Targeted at an AOR (angle of repose) of 41.67°, the optimal parameters were determined as restitution coefficient (<em>R<sub>e</sub></em>) = 0.051, rolling friction coefficient (<em>C<sub>r</sub></em>) = 0.027, and surface energy (<em>K<sub>i</sub></em>) = 0.180 J/m<sup>2</sup> between soil and seeds, yielding a relative error of 1.8 %. Single-factor simulation experiments were conducted to analyze seed-discharge height (<em>h<sub>d</sub></em>), velocity (<em>v<sub>d</sub></em>) and posture (<em>p<sub>d</sub></em>) on flat and V-shaped seedbeds from an impact-energy perspective. Uniform-design experiments were conducted to develop multivariate regression prediction models, which were then validated by simulations and bench tests. Results indicate that on flat seedbeds, <em>h<sub>d</sub></em>, <em>v<sub>d</sub></em>, and <em>p<sub>d</sub></em> significantly influence maximum bounce height (<em>h<sub>b</sub></em>) and rebound velocity (<em>v<sub>b</sub></em>), whereas planar displacement (<em>e<sub>p</sub></em>) is primarily governed by <em>h<sub>d</sub></em> and <em>v<sub>d</sub></em>. On V-shaped seedbeds, <em>v<sub>d</sub></em> exhibits a highly significant effect on <em>v<sub>b</sub></em> but no significant effect on <em>h<sub>b</sub></em> or <em>e<sub>p</sub></em>. As <em>v<sub>d</sub></em> increases, <em>h<sub>b</sub></em> and <em>v<sub>b</sub></em> follow a “first increase, then decrease” trend. In contrast, seed bounce generally intensifies on flat seedbeds. Validation tests showed no significant differences between model predictions, simulations, and bench tests (paired t-tests, p > 0.05), supporting model reliability. Under sandy loam conditions, using lower seed-discharge kinetic energy combined with matched furrow geometry and seed-discharge posture effectively mitigates bounce and drift of <em>Allium chinense</em> seeds. The optimized parameters are <em>h<sub>d</sub></em> = 180 mm and <em>v<sub>d</sub></em> = 0.36 m/s, with seeds oriented facing the furrow and aligned either parallel or perpendicular to the furrow axis. These findings provide actionable guidelines for mechanized oriented sowing systems targeting <em>Allium chinense</em> and analogous bulbous
中国Allium chinense是一种特色作物，因其独特的风味和药用价值而受到重视，大量出口到日本，韩国和国际市场。尽管具有重要的经济意义，但由于严格的农艺要求精确的种子定位，种植仍然是劳动密集型的。机械化的一个关键障碍是播种过程中碰撞引起的弹跳，这会影响播种质量。通过对种子-土壤碰撞机理的分析，探讨了种子流量参数和苗床几何形状对定向播种的影响。基于HM-JKR （Hertz-Mindlin with JKR）模型构建DEM （Discrete Element Method）模型，通过P-B （Plackett-Burman）筛选和B-B （Box-Behnken）优化对参数进行标定。以休止角（AOR）为41.67°为目标，确定了土壤与种子间的最优参数：恢复系数(Re) = 0.051，滚动摩擦系数(Cr) = 0.027，表面能(Ki) = 0.180 J/m2，相对误差为1.8%。通过单因素模拟实验，从冲击能的角度分析了平床和v型床的排种高度（hd）、排种速度（vd）和排种姿态（pd）。采用均匀设计实验建立多元回归预测模型，并通过模拟和台架试验对模型进行验证。结果表明，在平坦的苗床上，hd、vd和pd显著影响最大反弹高度（hb）和反弹速度（vb），而平面位移（ep）主要受hd和vd的影响。在v型苗床上，vd对vb的影响极显著，对hb和ep的影响不显著。随着vd的增加，hb和vb呈现先增加后减少的趋势。相反，在平坦的苗床上，种子反弹通常会加剧。验证检验显示，模型预测、模拟和台架检验之间无显著差异（配对t检验，p > 0.05），支持模型可靠性。在砂壤土条件下，采用较低的排种动能，配合匹配的畦型和排种姿态，可有效缓解葱种子的弹跳和漂移。优化后的参数为hd = 180 mm, vd = 0.36 m/s，种子朝向犁沟，平行或垂直于犁沟轴线。这些发现为中国葱及类似球茎作物的机械化定向播种系统提供了可操作的指导。
{"title":"Coupled effects of seed-discharge parameters and seedbed geometry on Allium chinense seeds–soil bounce","authors":"Qixin Kang , Guozhong Zhang , Chao Ji , Zhuangzhuang Zhao , Liming Chen , Lingxiao Lan","doi":"10.1016/j.compag.2026.111456","DOIUrl":"10.1016/j.compag.2026.111456","url":null,"abstract":"<div><div><em>Allium chinense</em> is a specialty crop valued for its distinctive flavor and medicinal properties, with substantial exports to Japan, South Korea, and international markets. Despite its economic importance, cultivation remains labor-intensive due to strict agronomic demands for precise seed orientation. A key obstacle to mechanization is collision-induced bouncing during sowing, which compromises placement quality. This study investigates the effects of seed-discharge parameters and seedbed geometry on directional sowing by analyzing seed–soil collision mechanisms. DEM (Discrete Element Method) models based on HM-JKR (Hertz-Mindlin with JKR) model were constructed, with parameters calibrated via P-B (Plackett-Burman) screening and B-B (Box-Behnken) optimization. Targeted at an AOR (angle of repose) of 41.67°, the optimal parameters were determined as restitution coefficient (<em>R<sub>e</sub></em>) = 0.051, rolling friction coefficient (<em>C<sub>r</sub></em>) = 0.027, and surface energy (<em>K<sub>i</sub></em>) = 0.180 J/m<sup>2</sup> between soil and seeds, yielding a relative error of 1.8 %. Single-factor simulation experiments were conducted to analyze seed-discharge height (<em>h<sub>d</sub></em>), velocity (<em>v<sub>d</sub></em>) and posture (<em>p<sub>d</sub></em>) on flat and V-shaped seedbeds from an impact-energy perspective. Uniform-design experiments were conducted to develop multivariate regression prediction models, which were then validated by simulations and bench tests. Results indicate that on flat seedbeds, <em>h<sub>d</sub></em>, <em>v<sub>d</sub></em>, and <em>p<sub>d</sub></em> significantly influence maximum bounce height (<em>h<sub>b</sub></em>) and rebound velocity (<em>v<sub>b</sub></em>), whereas planar displacement (<em>e<sub>p</sub></em>) is primarily governed by <em>h<sub>d</sub></em> and <em>v<sub>d</sub></em>. On V-shaped seedbeds, <em>v<sub>d</sub></em> exhibits a highly significant effect on <em>v<sub>b</sub></em> but no significant effect on <em>h<sub>b</sub></em> or <em>e<sub>p</sub></em>. As <em>v<sub>d</sub></em> increases, <em>h<sub>b</sub></em> and <em>v<sub>b</sub></em> follow a “first increase, then decrease” trend. In contrast, seed bounce generally intensifies on flat seedbeds. Validation tests showed no significant differences between model predictions, simulations, and bench tests (paired t-tests, p > 0.05), supporting model reliability. Under sandy loam conditions, using lower seed-discharge kinetic energy combined with matched furrow geometry and seed-discharge posture effectively mitigates bounce and drift of <em>Allium chinense</em> seeds. The optimized parameters are <em>h<sub>d</sub></em> = 180 mm and <em>v<sub>d</sub></em> = 0.36 m/s, with seeds oriented facing the furrow and aligned either parallel or perpendicular to the furrow axis. These findings provide actionable guidelines for mechanized oriented sowing systems targeting <em>Allium chinense</em> and analogous bulbous ","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111456"},"PeriodicalIF":8.9,"publicationDate":"2026-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079878","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 : 2026-01-27DOI: 10.1016/j.compag.2026.111475
Shuo Yang, Junxian Guo, Jiangtong Yu
This study proposes a generative AI-driven framework integrating CNN, VLM and LLM, aiming to provide intelligent solutions for the diagnosis of crop diseases and the generation of control strategies. The framework comprises four core modules: a data enhancement and style transfer module based on CycleGAN, a CNN-based disease detection module, a VLM-based visual semantic description module, and an LLM-based control strategy generation module. In the data enhancement module, CycleGAN is utilized to perform style transfer on the original dataset. This process makes the image features more realistic under natural conditions and improves the model’s generalization ability in real-world agricultural production environments. In the disease detection module, the Yolo-CDDet model is developed, which adopts a cascaded feature learning architecture. This architecture consists of a deformable convolution backbone network, a global–local feature pyramid pooling neck network, and a decoupled prediction structure detection head, enabling precise identification and classification of disease regions. In the visual semantic description module, the MultiTask-CLIP model is constructed, featuring a multi-task classification head. The model outputs text descriptions with fixed feature combinations, providing detailed visual evidence for subsequent control strategy formulation. In the control strategy generation module, the Falcon-40B large language model serves as the core component. By leveraging web crawlers to collect open-source professional agricultural literature and applying the LoRA fine-tuning method to optimize model parameters, the model is optimized. It generates scientifically grounded and practical control recommendations tailored to specific disease characteristics. Experimental results demonstrate that the Yolo-CDDet model achieves superior performance on both the original dataset and the dataset enhanced by style transfer. The model exhibits high Recall, Average Precision, and other excellent metrics. The MultiTask-CLIP model outperforms competing models across multiple evaluation criteria, particularly excelling in CIDEr scores. Additionally, the control strategy generation mechanism based on Falcon-40B surpasses baseline models in terms of Recall, Precision, ROUGE-L, and other quantitative analysis indicators, producing high-quality control strategy texts. This study offers a novel and effective approach for the intelligent diagnosis and integrated management of crop diseases.
{"title":"A generative AI-Driven framework integrating CNN-VLM-LLM for intelligent crop disease diagnosis and control strategy generation","authors":"Shuo Yang, Junxian Guo, Jiangtong Yu","doi":"10.1016/j.compag.2026.111475","DOIUrl":"10.1016/j.compag.2026.111475","url":null,"abstract":"<div><div>This study proposes a generative AI-driven framework integrating CNN, VLM and LLM, aiming to provide intelligent solutions for the diagnosis of crop diseases and the generation of control strategies. The framework comprises four core modules: a data enhancement and style transfer module based on CycleGAN, a CNN-based disease detection module, a VLM-based visual semantic description module, and an LLM-based control strategy generation module. In the data enhancement module, CycleGAN is utilized to perform style transfer on the original dataset. This process makes the image features more realistic under natural conditions and improves the model’s generalization ability in real-world agricultural production environments. In the disease detection module, the Yolo-CDDet model is developed, which adopts a cascaded feature learning architecture. This architecture consists of a deformable convolution backbone network, a global–local feature pyramid pooling neck network, and a decoupled prediction structure detection head, enabling precise identification and classification of disease regions. In the visual semantic description module, the MultiTask-CLIP model is constructed, featuring a multi-task classification head. The model outputs text descriptions with fixed feature combinations, providing detailed visual evidence for subsequent control strategy formulation. In the control strategy generation module, the Falcon-40B large language model serves as the core component. By leveraging web crawlers to collect open-source professional agricultural literature and applying the LoRA fine-tuning method to optimize model parameters, the model is optimized. It generates scientifically grounded and practical control recommendations tailored to specific disease characteristics. Experimental results demonstrate that the Yolo-CDDet model achieves superior performance on both the original dataset and the dataset enhanced by style transfer. The model exhibits high Recall, Average Precision, and other excellent metrics. The MultiTask-CLIP model outperforms competing models across multiple evaluation criteria, particularly excelling in CIDEr scores. Additionally, the control strategy generation mechanism based on Falcon-40B surpasses baseline models in terms of Recall, Precision, ROUGE-L, and other quantitative analysis indicators, producing high-quality control strategy texts. This study offers a novel and effective approach for the intelligent diagnosis and integrated management of crop diseases.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111475"},"PeriodicalIF":8.9,"publicationDate":"2026-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079824","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}
Manual fruit harvesting remains essential in modern horticulture for its ability to preserve fruit quality and reduce mechanical damage. However, the increasing labour intensity of these operations highlights the need for efficient assistive technologies to sustain productivity and ensure sustainability. Although fully automated systems offer alternatives, their limited adaptability in unstructured environments and high implementation costs restrict large-scale deployment. To address these limitations, a following collaborative robot harvesting-assisted transport system (FCR-HATS) is developed, incorporating a novel collaborative region constrained model (CRCM) to enhance harvesting efficiency and reduce manual workload through human-robot collaboration. CRCM introduces a decision-making framework that transitions the robot’s behaviour from conventional passive following to adaptive cooperation by enabling three execution modes: following, stopping, and monitoring. CRCM overcomes the trajectory deviations inherent in the traditional direction-following constraints model (DFCM) by applying collaborative region constraints, allowing the FCR to better align with the human trajectory. Field experiments demonstrate that CRCM significantly improves the accuracy of human-robot trajectory overlap, achieving an average matching cost (AMC) ranging from 0.14 to 0.25 across four benchmark paths. This outperforms traditional DFCM models and reduces collision risk without requiring additional obstacle avoidance strategies. With a unit shrinkage distance of 0.05 m, the FCR achieves precise stopping with lateral and longitudinal errors of 0.17 m and 0.15 m, respectively. Validation in a greenhouse peach orchard further confirms the robustness and continuous operational capability of the proposed system under real harvesting conditions.
{"title":"A following collaborative robot Harvesting-Assisted transport system with collaborative region constraints model for fruit harvesting","authors":"Hengda Li, Ying Chen, Zhenghao Li, Liang Sun, Haichao Li, Pingyi Liu","doi":"10.1016/j.compag.2026.111467","DOIUrl":"10.1016/j.compag.2026.111467","url":null,"abstract":"<div><div>Manual fruit harvesting remains essential in modern horticulture for its ability to preserve fruit quality and reduce mechanical damage. However, the increasing labour intensity of these operations highlights the need for efficient assistive technologies to sustain productivity and ensure sustainability. Although fully automated systems offer alternatives, their limited adaptability in unstructured environments and high implementation costs restrict large-scale deployment. To address these limitations, a following collaborative robot harvesting-assisted transport system (FCR-HATS) is developed, incorporating a novel collaborative region constrained model (CRCM) to enhance harvesting efficiency and reduce manual workload through human-robot collaboration. CRCM introduces a decision-making framework that transitions the robot’s behaviour from conventional passive following to adaptive cooperation by enabling three execution modes: following, stopping, and monitoring. CRCM overcomes the trajectory deviations inherent in the traditional direction-following constraints model (DFCM) by applying collaborative region constraints, allowing the FCR to better align with the human trajectory. Field experiments demonstrate that CRCM significantly improves the accuracy of human-robot trajectory overlap, achieving an average matching cost (AMC) ranging from 0.14 to 0.25 across four benchmark paths. This outperforms traditional DFCM models and reduces collision risk without requiring additional obstacle avoidance strategies. With a unit shrinkage distance of 0.05 m, the FCR achieves precise stopping with lateral and longitudinal errors of 0.17 m and 0.15 m, respectively. Validation in a greenhouse peach orchard further confirms the robustness and continuous operational capability of the proposed system under real harvesting conditions.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111467"},"PeriodicalIF":8.9,"publicationDate":"2026-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079796","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 : 2026-01-27DOI: 10.1016/j.compag.2026.111449
Zepeng Zhang , Zhongxiang Zhu , Liquan Lu , Guanglin Zhang , Haoran Yang , Zhenghe Song
High-quality tractor rotary tillage operations face significant challenges under complex environmental conditions. To overcome this, a front-axle direct torque-coupled hybrid tractor powertrain (FADTC-HTP) is proposed, aiming to improve energy efficiency and path stability via coordinated control of permanent magnet synchronous motor (PMSM) regenerative braking and an electronically controlled limited slip differential (eLSD). Based on the proposed powertrain, a hierarchical control strategy is developed, combining dual-stage model predictive control with reinforcement learning. First, a hidden Markov model (HMM)-based dual-stage linear time-varying MPC(HMM-DLTV-MPC) is designed as a supervisory controller to enable high-precision nonlinear dynamics approximation, adaptive parameter updating, and dynamic decoupling. To address the challenge of achieving optimal control of a multi-input multi-output (MIMO) nonlinear tractor system under random disturbances, an enhanced Soft Actor-Critic (SAC)-based subordinate controller is proposed to improve robustness and efficiency. Specifically, the controller incorporates expert experience guidance to inject prior knowledge and introduces a control target self-correcting mechanism to adaptively mitigate control errors caused by environmental uncertainties. Furthermore, by integrating long short-term memory (LSTM) networks with self-attention modules, the controller substantially strengthens the agent’s ability to capture complex multidimensional temporal dependencies, thereby significantly boosting both control performance and training efficiency. Agent training is conducted in a simulation environment constructed from real-world field traction data and soil cone index distributions. Hardware-in-the-loop (HIL) experiments show that the proposed method improves energy efficiency by up to 11.14% and reduces lateral path deviation by 76.47%. It also significantly reduces dependence on front-wheel steering compensation. These results demonstrate that the proposed powertrain and control strategy can effectively enhance both energy efficiency and operational stability in hybrid tractor applications. This study provides a new perspective on novel hybrid tractor powertrains and intelligent control strategies.
{"title":"Hierarchical control strategy for improving rotary tillage stability and efficiency of hybrid tractors using regenerative braking","authors":"Zepeng Zhang , Zhongxiang Zhu , Liquan Lu , Guanglin Zhang , Haoran Yang , Zhenghe Song","doi":"10.1016/j.compag.2026.111449","DOIUrl":"10.1016/j.compag.2026.111449","url":null,"abstract":"<div><div>High-quality tractor rotary tillage operations face significant challenges under complex environmental conditions. To overcome this, a front-axle direct torque-coupled hybrid tractor powertrain (FADTC-HTP) is proposed, aiming to improve energy efficiency and path stability via coordinated control of permanent magnet synchronous motor (PMSM) regenerative braking and an electronically controlled limited slip differential (eLSD). Based on the proposed powertrain, a hierarchical control strategy is developed, combining dual-stage model predictive control with reinforcement learning. First, a hidden Markov model (HMM)-based dual-stage linear time-varying MPC(HMM-DLTV-MPC) is designed as a supervisory controller to enable high-precision nonlinear dynamics approximation, adaptive parameter updating, and dynamic decoupling. To address the challenge of achieving optimal control of a multi-input multi-output (MIMO) nonlinear tractor system under random disturbances, an enhanced Soft Actor-Critic (SAC)-based subordinate controller is proposed to improve robustness and efficiency. Specifically, the controller incorporates expert experience guidance to inject prior knowledge and introduces a control target self-correcting mechanism to adaptively mitigate control errors caused by environmental uncertainties. Furthermore, by integrating long short-term memory (LSTM) networks with self-attention modules, the controller substantially strengthens the agent’s ability to capture complex multidimensional temporal dependencies, thereby significantly boosting both control performance and training efficiency. Agent training is conducted in a simulation environment constructed from real-world field traction data and soil cone index distributions. Hardware-in-the-loop (HIL) experiments show that the proposed method improves energy efficiency by up to 11.14% and reduces lateral path deviation by 76.47%. It also significantly reduces dependence on front-wheel steering compensation. These results demonstrate that the proposed powertrain and control strategy can effectively enhance both energy efficiency and operational stability in hybrid tractor applications. This study provides a new perspective on novel hybrid tractor powertrains and intelligent control strategies.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111449"},"PeriodicalIF":8.9,"publicationDate":"2026-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079800","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 : 2026-01-27DOI: 10.1016/j.compag.2026.111486
Demin Xu , Xinguang Zhang , Michael Henke , Liang Wang , Jinyu Zhu , Fang Ji , Yuntao Ma
Light is essential for photosynthesis and directly influences crop yield. During winter and spring, limited natural light makes well-managed supplemental lighting crucial for greenhouse production. Traditional lighting design methods, which rely on manual measurements, are inefficient for optimizing light distribution and energy use. This study proposes a 3D simulation framework to optimize supplemental lighting in greenhouses. The virtual model incorporates the spectral power distribution (SPD) and propagation characteristics of light-emitting diode (LED) modules, the optical properties of greenhouse materials, and the greenhouse’s geometric structure to simulate artificial light environments. Validation of the model demonstrated high accuracy, with an R2 of 0.982 and a RMSE of 14.38 μmol·m−2·s−1. Based on simulation outputs, the spatial layout of supplemental lighting modules was determined, and the hourly light integral (HLI) was used as a control variable to develop a time-segmented lighting strategy. For this study, the production performance of tomato was evaluated under four lighting treatments: HLI-driven fixed supplementary lighting (HFS), HLI-driven mobile supplementary lighting (HMS), nighttime timed supplementary lighting (TS), and only natural light (CK). The optimal lighting configuration was achieved when fixtures were positioned 1.7 m above the planting troughs. Tomato yield per plant under the HFS treatment increased by 25.2% compared to CK and by 21.6% compared to TS. While HMS showed higher energy-use efficiency and quantum yield, its yield improvement was relatively modest. Overall, HFS enhanced light energy-use efficiency and quantum yield by 5.5% and 55.3%, respectively, compared to TS. This study provides a practical decision-support tool for greenhouse lighting management, enabling data-driven optimization of light distribution and energy use. The proposed 3D modeling framework not only improves light-thermal synergy but also offers strong scalability for different greenhouse structures and crops. By integrating physical modeling and intelligent control, it contributes to the development of sustainable and smart agricultural production systems.
{"title":"Exploring the application mode of artificial light sources in solar greenhouses based on functional-structural plant model","authors":"Demin Xu , Xinguang Zhang , Michael Henke , Liang Wang , Jinyu Zhu , Fang Ji , Yuntao Ma","doi":"10.1016/j.compag.2026.111486","DOIUrl":"10.1016/j.compag.2026.111486","url":null,"abstract":"<div><div>Light is essential for photosynthesis and directly influences crop yield. During winter and spring, limited natural light makes well-managed supplemental lighting crucial for greenhouse production. Traditional lighting design methods, which rely on manual measurements, are inefficient for optimizing light distribution and energy use. This study proposes a 3D simulation framework to optimize supplemental lighting in greenhouses. The virtual model incorporates the spectral power distribution (SPD) and propagation characteristics of light-emitting diode (LED) modules, the optical properties of greenhouse materials, and the greenhouse’s geometric structure to simulate artificial light environments. Validation of the model demonstrated high accuracy, with an R<sup>2</sup> of 0.982 and a RMSE of 14.38 μmol·m<sup>−2</sup>·s<sup>−1</sup>. Based on simulation outputs, the spatial layout of supplemental lighting modules was determined, and the hourly light integral (HLI) was used as a control variable to develop a time-segmented lighting strategy. For this study, the production performance of tomato was evaluated under four lighting treatments: HLI-driven fixed supplementary lighting (HFS), HLI-driven mobile supplementary lighting (HMS), nighttime timed supplementary lighting (TS), and only natural light (CK). The optimal lighting configuration was achieved when fixtures were positioned 1.7 m above the planting troughs. Tomato yield per plant under the HFS treatment increased by 25.2% compared to CK and by 21.6% compared to TS. While HMS showed higher energy-use efficiency and quantum yield, its yield improvement was relatively modest. Overall, HFS enhanced light energy-use efficiency and quantum yield by 5.5% and 55.3%, respectively, compared to TS. This study provides a practical decision-support tool for greenhouse lighting management, enabling data-driven optimization of light distribution and energy use. The proposed 3D modeling framework not only improves light-thermal synergy but also offers strong scalability for different greenhouse structures and crops. By integrating physical modeling and intelligent control, it contributes to the development of sustainable and smart agricultural production systems.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111486"},"PeriodicalIF":8.9,"publicationDate":"2026-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079799","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 : 2026-01-27DOI: 10.1016/j.compag.2026.111474
Mengjie Liu , Yanlong Miao , Yida Li , Wenyi Sheng , Ruicheng Qiu , Minjuan Wang , Han Li , Man Zhang
<div><div>Maize leaf phenotypic parameters effectively reflect the photosynthesis and growth information of maize plants, which is crucial for breeding superior maize varieties. Current challenges include separating stems and leaves from a single maize plant and accurately measuring the phenotypic parameters of maize leaves. This study proposes a stem-leaf segmentation method based on region growing, incorporating adaptive cuboid region growing and slice region growing, alongside techniques for measuring phenotypic parameters of maize leaves. First, terrestrial laser scanning (TLS) was employed to obtain three-dimensional (3D) point cloud data of maize at the five-leaf (V5) and six-leaf (V6) stages. The point cloud data were then preprocessed to isolate single plant point clouds. Next, the maize point clouds were pre-segmented into three categories—central point clouds, partially expanded leaf point clouds, and unexpanded leaf point clouds—using center-edge segmentation, statistical filtering, and leaf classification. Adaptive cuboid region growing was applied to segment the unexpanded leaf point clouds, while slice region growing was used for partially expanded leaves, with Euclidean clustering optimizing the leaf point clouds, completing the segmentation process. Finally, various methods—including clustering counting, point-to-point distance accumulation, point-to-line distance, vector angle, point cloud triangulation, and triangle area accumulation—were utilized to automatically measure the number of maize leaves, leaf length, leaf width, leaf inclination angle, and leaf area. Compared with other point cloud stem-leaf segmentation methods based on geometric features and common 3D point cloud deep learning models (PointNet++, PointTransformer), the method proposed in this paper performs better. The segmentation results indicated that the Precision (<em>P</em>), Recall (<em>R</em>) and <em>F<sub>1</sub></em>-Score (<em>F<sub>1</sub></em>) for stem-leaf segmentation of all maize plants at the V5 stage exceeded 92.00%, with average values of 96.87%, 97.08%, and 96.97%, respectively. At the V6 stage, <em>P</em>, <em>R</em>, and <em>F<sub>1</sub></em> exceeded 95.00%, with averages of 97.73%, 97.01%, and 97.67%, respectively. The algorithm accurately measured the number of leaves at the V5 stage, while a small error was noted at the V6 stage, yielding a percentage error (<em>PE</em>) of 0.93%. Measurement accuracy for leaf length, width, and area at both growth stages was greater than 93.80%, 92.80%, and 89.50%, respectively. Measurement accuracy for leaf inclination angle was lower, at 82.00% and 88.02% for the V5 and V6 stages, respectively. The proposed methods for stem-leaf segmentation and measurement of leaf phenotypic parameters are fast and accurate, providing technical support for high-quality breeding and intelligent management of maize. Our point cloud data of maize and source code is available from https://github.com/lmj-cau/stem-leaf-se
{"title":"A stem-leaf segmentation method of maize plant point cloud based on region growing and leaf phenotypic parameters measurement","authors":"Mengjie Liu , Yanlong Miao , Yida Li , Wenyi Sheng , Ruicheng Qiu , Minjuan Wang , Han Li , Man Zhang","doi":"10.1016/j.compag.2026.111474","DOIUrl":"10.1016/j.compag.2026.111474","url":null,"abstract":"<div><div>Maize leaf phenotypic parameters effectively reflect the photosynthesis and growth information of maize plants, which is crucial for breeding superior maize varieties. Current challenges include separating stems and leaves from a single maize plant and accurately measuring the phenotypic parameters of maize leaves. This study proposes a stem-leaf segmentation method based on region growing, incorporating adaptive cuboid region growing and slice region growing, alongside techniques for measuring phenotypic parameters of maize leaves. First, terrestrial laser scanning (TLS) was employed to obtain three-dimensional (3D) point cloud data of maize at the five-leaf (V5) and six-leaf (V6) stages. The point cloud data were then preprocessed to isolate single plant point clouds. Next, the maize point clouds were pre-segmented into three categories—central point clouds, partially expanded leaf point clouds, and unexpanded leaf point clouds—using center-edge segmentation, statistical filtering, and leaf classification. Adaptive cuboid region growing was applied to segment the unexpanded leaf point clouds, while slice region growing was used for partially expanded leaves, with Euclidean clustering optimizing the leaf point clouds, completing the segmentation process. Finally, various methods—including clustering counting, point-to-point distance accumulation, point-to-line distance, vector angle, point cloud triangulation, and triangle area accumulation—were utilized to automatically measure the number of maize leaves, leaf length, leaf width, leaf inclination angle, and leaf area. Compared with other point cloud stem-leaf segmentation methods based on geometric features and common 3D point cloud deep learning models (PointNet++, PointTransformer), the method proposed in this paper performs better. The segmentation results indicated that the Precision (<em>P</em>), Recall (<em>R</em>) and <em>F<sub>1</sub></em>-Score (<em>F<sub>1</sub></em>) for stem-leaf segmentation of all maize plants at the V5 stage exceeded 92.00%, with average values of 96.87%, 97.08%, and 96.97%, respectively. At the V6 stage, <em>P</em>, <em>R</em>, and <em>F<sub>1</sub></em> exceeded 95.00%, with averages of 97.73%, 97.01%, and 97.67%, respectively. The algorithm accurately measured the number of leaves at the V5 stage, while a small error was noted at the V6 stage, yielding a percentage error (<em>PE</em>) of 0.93%. Measurement accuracy for leaf length, width, and area at both growth stages was greater than 93.80%, 92.80%, and 89.50%, respectively. Measurement accuracy for leaf inclination angle was lower, at 82.00% and 88.02% for the V5 and V6 stages, respectively. The proposed methods for stem-leaf segmentation and measurement of leaf phenotypic parameters are fast and accurate, providing technical support for high-quality breeding and intelligent management of maize. Our point cloud data of maize and source code is available from https://github.com/lmj-cau/stem-leaf-se","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111474"},"PeriodicalIF":8.9,"publicationDate":"2026-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080284","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 : 2026-01-27DOI: 10.1016/j.compag.2026.111484
Xiaopeng Li , Shuo Yang , Shuqin Li
Agriculture is the foundation of human survival and social development. The accuracy of crop disease diagnosis directly impacts food security and economic sustainability. Current Large Language Model (LLM) applications in this field face several issues, including open-loop designs, hallucinations, and insufficient multimodal data utilization. To address these problems, this paper proposes KiwiGuard, a collaborative kiwifruit disease diagnosis system based on a visual model and an LLM.
Firstly, we propose the KDI-Transformer, a disease severity identification model, to establish a closed-loop design. This model combines Multi-scale Patch Embedding (MPE), Enhanced Shortcut (ESC), and Adaptive Patch Fusion (APF). It provides robust input for the diagnosis system, even in complex environments. Secondly, the system mitigates the LLM hallucination problem found in vertical domains. It combines external knowledge sources with the LLM via Retrieval-Augmented Generation (RAG). This method enhances the knowledge base and improves generated output quality for kiwifruit disease identification. To further optimize the system, we propose the Multimodal Knowledge Graph-enhanced Answer Selection Model (MKGASM). This module addresses the underutilization of multimodal data. Through iterative semantic control and knowledge supplementation, MKGASM enhances both the quality and interpretability of generated responses. Experimental results demonstrate that the visual identification model achieved 89.57% accuracy, outperforming baseline models. Furthermore, the system reduced the hallucination rate to 8% and achieved a Fact-Consistency score of 0.86. Both expert and objective evaluations indicate that the system generates highly accurate responses with superior user interactivity. Together, these contributions represent a valuable exploration of applying LLMs in agriculture.
{"title":"From identification to prevention: a kiwi disease diagnosis system integrating vision and large language models","authors":"Xiaopeng Li , Shuo Yang , Shuqin Li","doi":"10.1016/j.compag.2026.111484","DOIUrl":"10.1016/j.compag.2026.111484","url":null,"abstract":"<div><div>Agriculture is the foundation of human survival and social development. The accuracy of crop disease diagnosis directly impacts food security and economic sustainability. Current Large Language Model (LLM) applications in this field face several issues, including open-loop designs, hallucinations, and insufficient multimodal data utilization. To address these problems, this paper proposes KiwiGuard, a collaborative kiwifruit disease diagnosis system based on a visual model and an LLM.</div><div>Firstly, we propose the KDI-Transformer, a disease severity identification model, to establish a closed-loop design. This model combines Multi-scale Patch Embedding (MPE), Enhanced Shortcut (ESC), and Adaptive Patch Fusion (APF). It provides robust input for the diagnosis system, even in complex environments. Secondly, the system mitigates the LLM hallucination problem found in vertical domains. It combines external knowledge sources with the LLM via Retrieval-Augmented Generation (RAG). This method enhances the knowledge base and improves generated output quality for kiwifruit disease identification. To further optimize the system, we propose the Multimodal Knowledge Graph-enhanced Answer Selection Model (MKGASM). This module addresses the underutilization of multimodal data. Through iterative semantic control and knowledge supplementation, MKGASM enhances both the quality and interpretability of generated responses. Experimental results demonstrate that the visual identification model achieved 89.57% accuracy, outperforming baseline models. Furthermore, the system reduced the hallucination rate to 8% and achieved a Fact-Consistency score of 0.86. Both expert and objective evaluations indicate that the system generates highly accurate responses with superior user interactivity. Together, these contributions represent a valuable exploration of applying LLMs in agriculture.</div></div>","PeriodicalId":50627,"journal":{"name":"Computers and Electronics in Agriculture","volume":"244 ","pages":"Article 111484"},"PeriodicalIF":8.9,"publicationDate":"2026-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079795","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