基于多特征优化的PolSAR数据农作物精细分类方法

Crop Classification Based on PolSAR Data Using Multiple Feature Optimization

摘要:

农作物精细分类在农业资源调查、农作物种植结构监管等诸多领域具有重要意义。极化合成孔径雷达（Polarimetric synthetic aperture radar, PolSAR）能够有效探测伪装和穿透掩盖物，提取多种散射特征信息，获取覆盖农作物生长关键物候阶段的连续时序信息，有效提升表达作物遥感特征的丰富度，在农作物分类中独具优势。但多时相和多特征的引入必然导致模型运算量剧增，不利于工程应用。针对上述问题，本文提出了一种基于多特征优化的PolSAR数据农作物精细分类方法，首先对PolSAR数据进行多种极化目标分解及参数提取以获得多个散射特征；然后使用基于栈式稀疏自编码网络和ReliefF优选的方法进行特征增强与优化，获取最优特征集；最后构建具有2个分支结构的卷积神经网络，融合不同卷积深度输出的特征，完成农作物的高精度分类。通过对单时相数据的特征分析、单时相数据初步分类实验和多时相数据不同特征集结合分类器的对比实验，证明本文所提方法能够在低维特征输入的前提下，最大程度提取不同作物之间的差异性特征，准确高效地实现对农作物的精细分类，最高分类精度和Kappa系数分别达到97.69%和97.24%。

Abstract:

Crop fine classification is of great significance in many fields such as agricultural resources survey and crop planting structure supervision. Polarimetric synthetic aperture radar (PolSAR) can effectively detect camouflage and penetrate masks, extract multiple scattering feature information, obtain continuous time series information covering the key climatic phases of crop growth, and effectively enhance the richness of crop remote sensing features, which is a unique advantage in crop classification. However, the introduction of multi-temporal phases and multi-features inevitably leads to a drastic increase in model arithmetic, which is not conducive to engineering applications. In view of the above problems, a multi-feature optimization-based approach for crop fine classification of PolSAR data was proposed, which firstly carried out multiple polarization target decomposition and parameter extraction of the PolSAR data in order to obtain multiple scattering features, and then a stacked sparse self-coding network based and ReliefF preferred method was used for feature enhancement and optimization to obtain the optimal set of features, and finally a convolutional neural network with two branching structures was constructed to fuse the features output from different convolutional depths to complete the high-precision classification of crops. Through the characterization of single-time-phase data, the preliminary classification experiments of single-time-phase data and the comparison experiments of combining classifiers with different feature sets of multi-time-phase data, it was proved that the method proposed can maximally extract the differential features between different crops under the premise of low-dimensional feature input, and accurately and efficiently realize the fine classification of crops, with the highest classification accuracy and Kappa coefficient reaching 97.69% and 97.24%, respectively.