文献类型：期刊文献

中文题名：基于语义分割网络模型的核桃叶片焦枯程度估计研究

英文题名：Research on the Estimation of Severity of Juglans Leaf Necrosis Based on the Model of Semantic Segmentation

作者：司恒山[1,2,3] 何子奇[1,2,3] 李志鹏[1,2,3] 陆森[1,2,3] 张劲松[1,2,3]

第一作者：司恒山

机构：[1]林木资源高效生产全国重点实验室,中国林业科学研究院林业研究所,北京100091;[2]南京林业大学南方现代林业协同创新中心,江苏南京210037;[3]河南小浪底森林生态系统定位观测研究站,河南济源454650

年份：2025

卷号：38

期号：1

起止页码：28-38

中文期刊名：林业科学研究

外文期刊名：Forest Research

收录：;北大核心:【北大核心2023】；

基金：新疆维吾尔自治区“揭榜挂帅”项目;中央级公益性科研院所基本科研业务费专项项目(CAFYBB2024ZA015)。

语种：中文

中文关键词：核桃叶片;焦叶症;严重程度分级;复杂背景;语义分割

外文关键词：walnut leaf;Juglans leaf necrosis;severity classification;complex background;semantic segmentation

分类号：TP391.4;S76;S75

摘要：[目的]实现核桃叶片焦叶程度的准确定量化,为科学精准治理焦叶症提供科学依据。[方法]以核桃叶片复杂背景图像为研究对象,提出基于语义分割网络模型的核桃叶片焦叶症分级方法。首先对焦叶叶片图像进行分割,主要包括两个阶段,第一阶段采用Segment Anything(SAM)模型在复杂自然背景下提取目标叶片的边缘轮廓,第二阶段分别使用SAM和Mask R-CNN模型,对焦叶叶片进行分割。然后,提出了核桃叶片焦叶程度的分级标准与方法。[结果]SAM和Mask R-CNN模型都具有较好地核桃焦叶叶片识别和分割能力。SAM模型虽然分割时需点选标识目标区域,但该模型无需再次训练即可直接运行,具有较好的可操作性和交互性。相比之下,经训练后的Mask R-CNN模型分割精度更高,其像素精度、平均像素精度、平均交并比分别为98.95%、98.19%、95.94%。同时,基于Mask R-CNN模型的核桃叶片焦叶程度的分级平均准确率达到91.29%。[结论]在复杂自然背景下,采用基于语义分割网络模型的两阶段核桃叶片焦叶程度分级方法,能够准确地对核桃叶片焦叶部位进行识别和分割,为核桃焦叶程度等级划分提供了理论依据,对核桃焦叶症的精准防控提供了技术支撑。
[Objective]The objective of this study is to determine the severity of Juglans leaf necrosis accurately,and provide scientific evidences for the efficient management of this disease.[Method]This study focused on walnut leaf images with complex backgrounds,and proposed a method for grading Juglans leaf necrosis severity with a semantic segmentation network model.Our method had two primary steps.First,the Segment Anything Model(SAM)was used to extract the edges and contours of target leaves in complex natural backgrounds.Second,the SAM and Mask R-CNN(Convolutional Neural Network)models were applied to segment the scorched leaves.Finally,the grading standard and method for evaluating Juglans leaf necrosis severity were developed.[Results]Both the SAM and Mask R-CNN models presented strong capabilities of identifying and segmenting scorched walnut leaves.The trained Mask R-CNN model showed a superior segmentation performance with a pixel accuracy of 98.95%,mean pixel accuracy of 98.19%,and mean Intersection over Union(IoU)of 95.94%.Although the SAM model required manual selection of target area in the segmentation process,the SAM model could be operated without the retraining step and showed excellent usability and interactivity.Alternatively,the average accuracy of grading Juglans leaf necrosis severity with the Mask R-CNN model was 91.29%.[Conclusion]In complex natural background,the two-steps grading method with the semantic segmentation network model can identify and segment the scorched areas of walnut leaves accurately.This method provides a theoretical basis for grading the Juglans leaf necrosis severity,and offers technical support for the precise control of this disease.