基于超分辨率重建的农田遥感影像分类方法

Method for Classifying Farmland Remote Sensing Images Based on Super-resolution Reconstruction

摘要:

大尺度农业分类制图对保障粮食安全、优化灌溉管理、保护生态环境、高标准农田建设以及推动农业资源可持续发展具有不可替代的价值。中低分辨率卫星影像难以满足高精度田间观测，高分辨率遥感影像获取成本通常较高，为实现精准的农业种植区分类，本文提出了基于超分辨率重建的农田遥感影像分类方法，并系统评估了其提升分类准确率的潜力。聚焦Sentinel-2号数据，设计了一种新型生成对抗网络（HM-PGGan），通过融合多尺度渐进生成与注意力机制，以增强高倍率超分辨率重建的结构保真度。除评估重建后对影像分类结果的改进外，进一步分析了超分辨率重建影像与基于GLCM的纹理特征在支持向量机（SVM）和随机森林（RF）分类器协同作用下综合效应。从视觉分析与定量对比双重视角评估了道路、建筑物、裸土等典型地物类型的分类性能。试验结果表明，针对下采样至40 m的Sentinel-2号影像，经16倍超分辨率重建并结合纹理特征后，总体准确率从重建前的95.45%（SVM）和96.76%（RF）提升至98.19%（SVM）和98.16%（RF）;在跨平台超分辨率影像重建分类试验中，总体准确率从77.79%（SVM）和78.50%（RF）提升至93.28%（SVM）和93.77%（RF）。研究结果显示，超分辨率重建及添加纹理特征作为空间信息的辅助，均可增强原始影像的分类制图精度，但不同地物类别对高倍率超分辨率与特征融合的响应存在差异，为农业种植区遥感影像高精度分类提供了新的技术思路。

Abstract:

Large-scale agricultural land classification mapping plays an irreplaceable role in ensuring food security, optimizing irrigation management, protecting the ecological environment, supporting the construction of high-standard farmland, and promoting the sustainable development of agricultural resources. However, the limited spatial resolution of freely available satellite data (e.g., Sentinel-2) hinders precise field-level observation, while high-resolution imagery remains costly. The potential of 16× super-resolution (SR) reconstruction to overcome this limitation was investigated. High-magnification multi-scale progressive growing generative adversarial network (HM-PGGan), a novel network integrating multi-scale progressive generation and attention mechanisms, was proposed to enhance the structural fidelity of high-magnification SR reconstruction from Sentinel-2 data. The research evaluated not only the classification improvement from SR-reconstructed imagery but also the synergistic effect of combining it with GLCM-based texture features, using both support vector machine (SVM) and random forest (RF) classifiers. Comparative experiments on typical land cover types (roads, buildings, bare soil) showed that 16×SR reconstruction with texture features boosted the overall classification accuracy (OA) for downsampled 40 m Sentinel-2 imagery from 95.45% (SVM) and 96.76% (RF) to 98.19% (SVM) and 98.16% (RF). In a cross-platform SR experiment, OA was improved from 77.79% (SVM) and 78.50% (RF) to 93.28% (SVM) and 93.77% (RF). The results demonstrated that while both SRR and the integration of texture features as ancillary spatial information enhanced classification accuracy, the responses varied significantly across different land cover types. The research result can provide a novel technical pathway for high-precision remote sensing classification of agricultural planting areas.