基于深度学习与几何特征的田块变化遥感检测方法

REMOTE SENSING DETECTION METHOD FOR PLOT CHANGES BASED ON DEEP LEARNING AND GEOMETRIC FEATURES

关键词:  田块平整  深度学习  几何特征匹配  田块变化识别  变化分析

Key words:  farmland leveling  deep learning  geometric feature matching  farmland plot change detection  change analysis

摘要:

目的 传统田块变化研究受限于高精度田块空间数据获取效率低和检测分析手段单一，难以深入探究田块时空演变及精准评估田块平整工程成效。方法 文章提出了一种基于深度学习与几何特征的田块变化遥感检测方法，旨在精准地分析田块变化特征并科学衡量田块平整工程建设成效。首先，研究基于亚米级的高分辨率卫星遥感影像，采用BsiNet模型对江苏、浙江和重庆3个研究区的田块进行智能提取，高效获取了高精度的田块数据；随后，通过将田块平整前后的田块数据进行几何特征匹配，来精确识别发生变化的田块变化区域；最后，从建设成效的田块规模、空间形态、耕作效率三方面，对田块变化区域进行全面量化对比分析，进而实现对田块时空变化的有效检测。结果 （1）BsiNet模型实现了高精度田块数据的高效提取，平均准确率达到90%，为田块平整工程成效评估提供了高精度的田块空间数据；（2）基于几何特征匹配的变化识别方法充分考虑田块形状和大小的变化，实现了田块平整背景下多区域变化田块的高效识别与空间定位；（3）3个研究区的田块变化均表现为规模扩大、形态规整化和耕作效率提升。江苏、浙江、重庆3个研究区的平均田块规模增幅分别475%、153%、43%；平均耕作效率分别提升28%、21%、14%。结论 该研究提出的田块变化遥感检测方法，有效解决了现有田块变化研究存在的高精度田块空间数据获取效率低、检测分析手段单一的问题，实现了对田块时空演变的有效检测，能够为田块平整工程的成效评估提供切实可行的技术支持。

Abstract:

Traditional studies on plot changes are constrained by low efficiency in acquiring high-precision plot spatial data and limited detection and analysis methods, making it difficult to thoroughly investigate the spatiotemporal evolution of plots and accurately assess the effectiveness of plot leveling projects. This study proposed a remote sensing detection method for plot changes based on deep learning and geometric features, aiming to precisely analyze plot change characteristics and scientifically evaluate the effectiveness of plot leveling projects. First, high-resolution sub-meter satellite remote sensing images were utilized, and the BsiNet model was employed to extract high-precision plot boundaries in three study areas: Jiangsu, Zhejiang, and Chongqing. Subsequently, plot data before and after leveling were matched based on geometric features to precisely identify areas of change. Finally, the identified areas were comprehensively analyzed in terms of plot scale, spatial morphology, and cultivation efficiency, thereby achieving effective detection of the spatiotemporal changes in plots. The results showed that: (1) The BsiNet model achieved efficient and high-precision extraction of plot data, with an average accuracy of 90%, providing reliable spatial data for evaluating the effectiveness of plot leveling projects. (2) The change detection method based on geometric feature matching fully considered changes in plot shape and size, enabling efficient identification and spatial localization of plot changes under land leveling conditions across multiple regions. (3) The results analysis revealed that plot changes in all three study areas exhibited trends of increased scale, improved shape regularity, and enhanced cultivation efficiency. The average increase in plot scale in the Jiangsu, Zhejiang, and Chongqing study areas was 475%, 153%, and 43%, respectively; the average improvement in cultivation efficiency was 28%, 21%, and 14%, respectively. Therefore, the proposed remote sensing detection method for plot changes effectively addresses the limitations of low efficiency in acquiring high-precision plot spatial data and single-method monitoring approaches in existing research. It enables effective monitoring of the spatiotemporal evolution of plots and provides practical technical support for evaluating the effectiveness of plot leveling projects.