基于改进PatchSVDD的田间异常区域检测方法  

Farmland anomaly detection based on improved PatchSVDD

作　　者：陈祖强 庞立欣 郭娜炜 蔡金金[4] 么炜[1,2] 刘博 CHEN Zuqiang;PANG Lixin;GUO Nawei;CAI Jinjin;YAO Wei;LIU Bo(College of Information Science and Technology,Hebei Agricultural University,Baoding 071001,China;Hebei Key Laboratory of Agricultural Big Data,Hebei Agricultural University,Baoding 071001,China;Science Research Academy,Hebei Agricultural University,Baoding 071001,China;College of Mechatronical&Electrical Engineering,Hebei Agricultural University,Baoding 071001,China)

机构地区：[1]河北农业大学信息科学与技术学院,河北保定071001 [2]河北省农业大数据重点实验室,河北保定071001 [3]河北农业大学科学技术研究院,河北保定071001 [4]河北农业大学机电工程学院,河北保定071001

出　　处：《河北农业大学学报》2024年第1期106-114,共9页Journal of Hebei Agricultural University

基　　金：国家自然科学基金项目(61972132);河北省自然科学基金项目(F2020204009);河北省重点研发计划项目(21327215D);河北省省级科技计划资助项目(20327404D、20327401D、21327404D);河北省引进留学人员资助项目(C20190342)。

摘　　要：利用无人机遥感技术对农田进行监测并及时发现田间异常对保证农业生产具有重要意义。目前田间异常区域检测需要标注大量的正常与异常样本。然而,异常样本在整个农田区域中占比过小且无法充分收集。特别是农田异常的多样性和不可预知性,对检测方法的适用性提出了更高的要求。针对以上问题,本文提出基于改进PatchSVDD (Patch-level Support Vector Data Description)模型的田间异常区域检测方法,该方法仅使用田间正常区域的标注信息,即可对田间异常区域进行检测和定位。首先,改进方法引入不相邻图像块之间的边界损失函数,从而提升了正常与异常样本边界的判别性,改进了检测的鲁棒性;第二,引入外部记忆组件,通过压缩存储正常样本特征,从而在保证检测精度的基础上有效减少了测试阶段的时间和空间消耗;第三,构建了包含杂草簇、种植缺失、障碍物、双倍种植和积水共5个异常类型的田间异常数据集。本文方法在平均检测AUC(Area Under Curve)值和平均定位AUC值上分别达到了96.9%和94.6%,相比于原算法分别提升1.2%和1.6%,从而验证了方法的有效性。It is of significant importance that UAV remote sensing technology was adopted to identify farmland anomalies to ensure agricultural production.Most current farmland anomaly detection methods require labeling a large number of normal and abnormal samples.However,the abnormal samples in the farm area are too small to be collected adequately.In particular,the diversity and unpredictability of farmland anomalies require advanced detection methods.To address these issues,this paper proposed an improved PatchSVDD(Patch-level Support Vector Data Description) model for detecting abnormal areas in the farm, which only utilized the labeling information ofnormal areas in the farm without labeling the abnormal ones. First, the improved method introduced a margin-basedloss function between non-adjacent image patches to improve the discriminability of the boundary between normaland abnormal samples and enhance the robustness of the detector. Second, an external memory module was adoptedto store the compressed regular sample features to effectively reduce the time and space consumption in the testingphase while ensuring detection accuracy. Third, a farmland anomaly detection dataset was constructed containing atotal of five anomaly types, i.e., weed clusters, missing planting, obstacles, double planting and standing water. Theproposed method achieved 96.9 % and 94.6 % in the average detection AUC (Area Under Curve) value and averagelocalization AUC value, respectively, which demonstrated improvements of about 1.2 % and 1.6 % compared with theoriginal PatchSVDD, suggesting the effectiveness of the method.

关 键 词：农田监测 异常检测 无人机遥感 三元损失函数 核心集 

分 类 号：S127[农业科学—农业基础科学] TP391[自动化与计算机技术—计算机应用技术]