一种基于归一化扰动模型的积雪和植被覆盖度反演方法

积雪和植被的覆盖范围对于研究气候变化和水资源平衡、生态环境状况具有重要的意义,但它们的光谱曲线具有较强的时空变异性,难以获取精确的覆盖度产品。针对线性混合像元分解算法在积雪和植被覆盖度反演中噪声和光谱变异带来的误差,本文提出了一种基于归一化扰动模型的积雪和植被覆盖度反演方法,并选用了3个不同的区域（单独的积雪覆盖区、单独的植被覆盖区、积雪和植被混合的覆盖区）来验证所提出框架的可行性。研究结果表明：① 该方法单独反演积雪覆盖度的均方根误差为0.172,单独植被覆盖度反演均方根误差为0.223,积雪和植被覆盖度混合反演的均方根误差分别为0.185和0.249,3种方案均有较高的精度;② 对影像与端元组进行归一化后,降低了光谱异质性,在此方法下的扰动混合模型可以有效地减弱MODIS影像光谱变化和噪声带来的误差;③ 针对MODIS影像,该框架获取的积雪覆盖度相对于植被覆盖度具有更高的精度。今后将进一步发展类似的积雪覆盖度与雪粒径协同反演算法。

Snow and Vegetation Cover Fractions Mapping Using a Perturbed Mixing Model based on Spectral Normalization

Snow and vegetation cover fractions are important for studying climate change, water resource balance, and eco-environmental conditions. Yet, it is difficult to acquire accurate cover products due to the high spatiotemporal variability of snow and vegetation cover fractions. The enemember variability can result from complex terrains, atmospheric influences, and the intrinsic variability of features such as the chlorophyll concentration in plants, snow particle size, and snow contamination. Supervised spectral unmixing algorithms often assume that the endmembers are known exactly. However, in practice, the endmembers are extracted from real spectral images that may be affected by measurement noises or errors. In addition, available knowledge of the endmembers might not exactly match the actual endmembers of the spectral image at hand, since the spectral signature of the same material may be slightly altered in different images or because distinct but confusingly similar spectral signatures may be mixed up. To solve this problem, this paper proposed a perturbed mixing model （PMM） based on spectral normalization. The PMM attempted to reduce the errors caused by spectral changes and noises by introducing disturbed factors （both image and endmember matrix are perturbed）. The PMM model could capture the small noise and endmember variations, yet its accuracy would decrease when the spectrum changed enormously. To solve the problem, the spectral normalization was used to reduce the differences between the endmembers and the spectrum matrix. Spectral normalization did not change the correlation between endmembers and the relative position of high dimensional feature space, but aggregated spectral features of higher correlation coefficients and decreases spectral changes. Then, the PMM was used to quantify the spectral variation and measurement noise/error in order to improve the accuracy of the snow and vegetation cover mapping. Finally, three different areas （snow-dominated region, vegetation-dominated region, and region where snow and vegetation are mixed） were selected to validate the feasibility of the proposed framework. Results show: （1） The root mean square error （RMSE） of snow-dominated region was 0.172, the RMSE of vegetation-dominated region was 0.223, and the RMSE of snow and vegetation mixed region were 0.185 and 0.249, respectively. Relatively high accuracy was achieved in the three types of areas. （2） After normalizing the endmembers and images, the spectral heterogeneity was obviously decreased and the overall accuracy of the three algorithms was better than before normalization. （3） The snow coverage fraction obtained by the framework had higher accuracy than the vegetation coverage fraction.