中国农业大学学报2025，Vol.30Issue(2)：163-175,13.DOI:10.11841/j.issn.1007-4333.2025.02.15
基于残差神经网络的水稻氮磷钾元素营养诊断
Nutritional diagnosis of nitrogen,phosphorus and potassium in rice based on the residual neural network
摘要
Abstract
To achieve rapid and accurate diagnosis and recognition of three major nutrient element deficiency types in rice,one late rice variety"Huanghuazan"was selected as the research object for field cultivation trials.Different treatments are set as follows:Four nitrogen application levels are N0(0 kg/hm2),N1(130 kg/hm2),N2(260 kg/hm2)and N3(390 kg/hm2);Four phosphorus application levels are P0(0 kg/hm2),P1(300 kg/hm2),P2(600 kg/hm2),P3(780 kg/hm2);Four potassium application levels are K0(0 kg/hm2),K1(90 kg/hm2),K2(180 kg/hm2)and K3(270 kg/hm2).During the tillering and jointing stages of rice growth,the high-resolution image data of the top three fully expanded leaves from each tiller were scanned,and after generative adversarial network(GAN)reconstruction of rice image with super-resolution,the data underwent normalization and expansion through image preprocessing;Additionally,attention mechanisms and soft thresholding functions were integrated into the residual block while preserving the backbone structure of residual neural network Resnet34;Furthermore,pre-trained weights obtained from the ImageNet dataset were transferred to the model with an improved residual structure.The results indicate that:The improvement of Resnet34 network yielded extraordinary results,exhibiting a remarkable accuracy of 98.98％and 98.10％within the severe fertilizer deficiency gradients during these pivotal growth phases.Resnet34 network demonstrated robustness in moderate deficiency scenarios,and achieved the accuracies of 97.99％and 95.90％,respectively.Regarding excessive fertilizer application,the Resnet34 model maintained a commendable performance and achieved accuracies of 91.87％and 88.49％across the respective gradients.Compared with the pre-improvement network,the three-element stress classification accuracy of the rice image test set at the tillering stage and jointing stage increased by 29.58％and 29.75％;The analysis of the confusion matrix indicated that the identification accuracy of nitrogen stress was superior,and the training curve demonstrated faster convergence speed.In conclusion,the model established in this study exhibits exceptional proficiency in diagnosing nutrient deficiency stress during the pivotal tillering and jointing stages of rice growth,and can accurately predict the rice nutritional status.This study provides a scientific reference for the nutritional diagnosis of nitrogen,phosphorus and potassium in rice.关键词
水稻/氮磷钾营养诊断/Resnet34/注意力机制/软阈值化Key words
rice/nitrogen,phosphorus and potassium nutritional diagnosis/Resnet34/attention mechanism/soft thresholding分类
信息技术与安全科学引用本文复制引用
孔杰,杨红云,黄淑梅,吴正,孙爱珍..基于残差神经网络的水稻氮磷钾元素营养诊断[J].中国农业大学学报,2025,30(2):163-175,13.基金项目
国家自然科学基金项目(62162030,61562039). （62162030,61562039）
