一种基于集成学习和特征融合的遥感影像分类新方法

A novel remotely sensed image classification based on ensemble learning and feature integration

针对多源遥感数据分类的需要,提出了一种基于全极化SAR影像、极化相干矩阵特征、光学遥感影像光谱和纹理的多种特征融合和多分类器集成的遥感影像分类新方法.对全极化PALSAR数据进行预处理和极化相干矩阵特征提取,利用灰度共生矩阵计算光学和SAR影像的对比度、逆差距、二阶距、差异性等纹理特征参数,并与光谱特征结合,形成6种组合策略.利用集成学习方法对随机森林分类器、子空间分类器、最小距离分类器、支持向量机分类器、反向传播神经网络分类器等分类器进行组合,对不同组合策略的遥感影像特征集进行分类.结果表明提出的基于多种特征和多分类器集成的新方法很好地利用了主被动遥感数据在不同地表景观类型提取上的潜力,综合了多种算法的优势,能够有效地提高总体精度和各类别的分类精度.

To make full use of the multi-source remotely sensed data for classification, a novel method was proposed based on the integration of full-polarization SAR （HH, HV, VH, VV） data, features of polarization coherence matrix, spectral features provided by optical data, texture features extracted from optical and SAR data and multi-classifier ensemble. Preprocessing for full-polarization data was performed and polarimetric features are extracted from polarization coherence matrix. Spatial textural features including contrast, dissimilarity, second moment, etc., are extracted from PALSAR full-polarization data and optical image using Grey-level Co-occurrence Matrix （GLCM） method. Features of polarization coherency matrix, full-polarization SAR channels, spectral and textures are integrated by 6 strategies. Some well-known classification techniques, including Support Vector Machine （SVM）, Minimum Distance （MD）, Back Propagation Neural Network （BPNN）, Multi-Layer Perceptron （MLP）, Random Subspace （RSS）, Random Forest （RF） classifiers were selected to test different combination strategies. The parallel and sequential ensemble learning techniques were selected to integrate single classifier for land cover classification. The results indicate that the proposed approach integrating multi-source, multi-features and multi-classifier strategy can make full use of the potential of optical and SAR remotely sensed data for landscape types, and improve the overall accuracy and the accuracy of single land cover type effectively.