顾及小样本的森林优势树种半监督稀疏表示遥感分类

The application of remote sensing technology in large scale tree species classification plays a crucial role. But its classification precision has been affected by some problems as the small sample, the high dimension data, and mixed spectral pixels in the process of the application development. So multi-source remote sensing images should be made full use to extract features, and put forward a new classifier based on the theory of semi-supervised learning and algorithm of sparse representation. Then those restriction problems could be solved. This project will summarize the methods integrated spectrum and structure, high dimension and low dimension, interpretation and optimization. First, on the basis of the theory of morphological, it establishes the dominance species features set including texture, spectral, canopy characteristics by Gabor filter, wavelet transform, LM Gauss filter methods based on high spatial and hyperspectral data. These methods should face the forest diversified characteristics and complex structure information. Second, based on the theory of pattern recognition with small sample, the research proposes a sparse representation classification method in the dichotomy of species participated not mark sample and tag sample. And it uses semi-supervised learning method to solve the contradiction between feature dimension and sample size. Finally, based on sparse representation theory, this solves the sample noise by the algorithm of adaptive mutation research. It studies a global optimizing SVM classifier fusing holistic and local features in subspaces based on high accuracy level. They realize a sparse representation classification of forest trees with multi-source and feature fusion. This research will expand the remote sensing classification theory and method under the condition of small sample about forest trees. The new tools for fine tree species classification could be provided.

受小样本、数据高维、混合光谱等问题影响，当前基于多源遥感数据的森林树种分类精度不高，缺乏对分类体系的学习能力、泛化能力及鲁棒性等特点的深入研究，导致树种分类结果可信度较低。本项目拟运用光谱与结构、高维与低维、解译与优化相结合的研究方法，在形态学理论基础上，针对特征结构复杂、尺度多元的林分特征及背景信息，基于高空间和高光谱遥感数据，利用Gabor滤波、小波变换、LM高斯滤波方法构建优势树种多特征样本集；针对树种特征维度与样本数量的矛盾，基于小样本模式识别理论，研究新型的半监督学习方式，探讨未标记样本与标记样本参与树种的二分性稀疏表示分类方法；针对树种分类器的设计与优化，基于稀疏表示理论与自适应变异算法，研究基于高可信度融合局部的SVM全局优化分类器，实现多特征的森林树种稀疏表示遥感分类，该成果有望拓展森林树种小样本遥感分类的理论与方法，为树种精细遥感分类研究提供参考。

本课题以多源遥感数据为研究对象，通过对树种特征的样本学习和分类方法研究，构建多源多特征的树种遥感分类体系，设计有效的算法模型和研究方法，解决小样本、信息高维、混合光谱等制约树种分类精度的问题，并通过典型研究区优势树种分类的应用实例进行方法验证。.主要研究成果包括：1、通过从合理配置波长、方向、相对位移、椭圆率等参数的角度改进Gabor滤波方法，提取高分辨率遥感影像中树种的纹理特征。冠幅特征的提取为后续分类提供可靠的空间位置、对象特征、几何形状等信息，针对树种冠幅分布的疏密程度，研究基于高光谱数据的光谱端元分解方法，探讨单立木空间结构，在高斯滤波的理论基础上，设计以高分影像为研究对象的单立木树冠轮廓的检测方法，以生长算法为理论基础拟合树冠轮廓，构建并整合冠幅形状生长边界，实现单立木的识别。2、提出一种基于深度对抗域适应的高分辨率遥感影像跨域分类方法。利用深度卷积神经网络VGG16学习场景影像的深度特征，利用对抗学习方法最小化源域和目标域特征分布差异，从而提高模型对目标域的分类性能和模型对目标域数据集的泛化学习能力。3、提出一种基于分类器约束的深对抗域自适应（CDADA）遥感图像跨域半监督分类方法。该方法利用深度卷积神经网络（DCNN）建立特征表示来描述语义内容适应过程之前的场景，利用对抗域自适应来调整源和目标的特征分布，采用全局平均池化压缩树种影像特征，使用简单线性迭代聚类生成超像素，以超像素为最小分类单元，生成树种分类专题地图。.通过本课题的研究，不仅丰富森林树种遥感分类的理论方法，而且在森林优势树种遥感分类研究上取得了原创性的研究成果。