基于多季相光谱混合分解和决策树的干旱区土地利用分类

该文利用2015年Landsat 8多季相数据，以民勤县为例探讨了干旱区土地覆被/利用分类方法。首先进行了研究区多季相遥感数据的主成分变换，确定了Landsat OLI光谱空间的内在维数。其次通过对各季相主成分空间的分析，确定了光谱混合分解的端元类型及各自的代表性季相。获取端元光谱之后，采用全约束线性光谱混合模型分解各个季相的遥感影像，估计各端元组分占像元的面积百分比（端元丰度值）。最后利用端元丰度值的多季相估计结果，依据先验知识和训练样本建立决策树分类规则进行土地覆被/利用分类。分类方法的总体精度达到90.94%，Kappa系数为0.90。研究表明：将物理意义明确的端元丰度值作为分类变量，能够快速、有效地获取分类规则，生成树的结构简单、合理、清晰，对训练数据的依赖性较低，因此具有应用于整个旱地系统土地覆被/利用分类的潜力。

Land use classification in arid region based on multi-seasonal linear spectral mixture analysis and decision tree method

Abstract: Land-cover/land-cover pattern of arid ecosystem in China, which determines ecosystem service value, is undergoing accelerated changes due to natural and anthropogenic disturbances. To provide a basis for ecosystem service assessment, in this article, we adopted 2015 multi-seasonal data of Landsat 8 for exploring the land cover/use classification method in arid region using the Minqin oasis as a case study area. We firstly conducted a principle component transform on the multi-seasonal remote data for identifying the inherent dimensionality of Landsat OLI spectral space. Then endmember classes and respective representative season were determined through the analysis of principal component feature space for each season. After extracting the endmember spectra by averaging the reflectance of 200~400 endmember pixels, fully-constrained linear spectral mixture model was performed on each seasonal image for yielding quantitative estimates of the areal abundance of endmembers within each pixel. At last, these endmember abundance estimates were used for land cover/use classification in Minqin study area by using the decision tree method. According to the natural environment and land-use characters of study area and given the resolution of remote sensing data and applicability for ecosystem service assessment, in this research, we developed the two-level classification system. Exposed surface, crop land, forest/shrub land, grassland, impervious surface and water area were defined as first-level classes. The exposed surface was subdivided into moving sand, Gobi/hill/bare-land, salinized moving sand, and saline-alkaline land. Crop land was subdivided into spring crop, summer crop, perennial crop based on seasonal growth characteristics. Similarly, forest/shrub land was subdivided into spring forest/shrub, summer forest/shrub, and evergreen forest/shrub. Decision tree was designed based on the seasonality pattern of feature endmember abundance of each target class. The first step in the classification procedure was to overlay the training data on the three-seasonal abundance composite images for identifying the seasonality pattern of each class. The second step was to measure the feature endmember abundance distribution of each class within training samples. Aided by the histogram distribution, the segmenting boundary of each node was established by an interactive process. The results showed that sand, salt, green vegetation, dark materials, and water were five endmember classes used for multi-seasonal linear spectral mixture analysis. But their representative seasons were different. So, we extracted endmember reflectance for sand, salt, and green vegetation from early winter, spring, and summer, respectively. The spectral reflectance of dark material and water endmembers were derived from spring as well as salt. The mean RMSE (root-mean-square error) values were all lower than 0.01, which meant good fitness of linear spectral mixture model for each seasonal image. From the statistics of endmember abundance estimates, we found sand and dark materials were matrix of the study area. The areal percentages of water and salt were relatively low, and their values were not uniform. The green vegetation cover was the lowest, and corresponding abundance varied most. Through analyzing the rate of endmember abundance change between seasons, it seemed that the abundances of sand and dark materials did not vary with seasons. Green vegetation increased significantly from spring to summer, and the overall amounts of water and salt somewhat decreased during the same period. The overall precision of the classification method developed in this study was 90.94%, and the Kappa coefficient was 0.90. It suggested the reliability of the 2015 land cover/use classification results of Minqin. The supplement use of multi-seasonal remote data can help acquire the comprehensive information of target surface feature, which was important for the stability of land cover/use classification results. Owing to the obvious physical meaning of endmember abundance, the design of decision tree can depend on the prior knowledge instead of vast training data. Thus, the classification method developed in this study may have the potential in the land cover/use classification of the whole drylands in China.