基于Landsat 8-OLI的荒漠化地区植被覆盖度反演模型研究

摘要

Percentage vegetation cover is a direct and effective measure used to assess and monitor desertification. In this study, a novel method that combined stepwise regression and linear spectral unmixing analysis was developed to derive an integrated regression model of percentage vegetation cover against image vegetation indices and vegetation fraction. A total of 134 sample plots were systematically selected in the study area–Kangbao County and percentage vegetation cover data that quantified desertification were collected. Landsat 8-OLI image was acquired and a total of 16 vegetation indices were calculated. Linear spectral unmixing analysis was conducted to extract vegetation fraction of mixed pixels that were introduced into stepwise regression. The results were validated using the observations of sample plots and showed that: (1) When individual vegetation indices were used to develop regression models, the vegetation indices that had the highest correlation were normalized difference vegetation index (NDVI) and soil adjusted vegetation index (SAVI). In the case of stepwise regression modeling, soil adjusted vegetation index (SAVI0.5), simple ratio (SRN-R) and enhanced vegetation index (EVI) had significant contributions and were selected. (2) When only vegetation fraction from linear spectral unmixing analysis was used as an independent variable, the coefficient of determination for the obtained model was 0.673, lower than the coefficient from stepwise regression, but higher than those when individual vegetation indices were employed. (3) The results of accuracy assessment showed that the stepwise regression based on the vegetation indices led to the coefficient of determination R2 and accuracy of 0.719 and 86.70%, and the new method that introduced vegetation fraction together with the vegetation indices as independent variables into the stepwise regression resulted in the coefficient of determination R2 and accuracy of 0.807 and 92.37%. This implied that the integration of regression modeling and linear spectral unmixing analysis provided a great potential to increase the accuracy of predicting percentage vegetation cover as a measure of desertification.%基于Landsat 8-OLI影像数据，利用植被指数逐步回归分析和线性混合像元分解的方法，结合134个野外样地调查数据，将线性混合像元分解结果（植被丰度）导入影像植被指数逐步回归模型，建立康保县荒漠化地区植被覆盖度反演混合模型，并进行精度检验。结果表明：（1）在所选16种影像植被指数中，采用单一植被指数进行荒漠化地区植被覆盖度反演建模，与植被覆盖度拟合优度最高的是归一化植被指数（NDVI）和土壤调节植被指数（SAVI），利用植被指数逐步回归分析建模，筛选出的3种最佳影像植被指数是土壤调节植被指数（SAVI0.5），比值植被指数（SRN-R）和增强型植被指数（EVI）；（2）通过线性混合像元分解建立的植被覆盖度反演模型，分解所得植被丰度与植被覆盖度的决定系数为0.673，模型精度低于利用植被指数逐步回归分析法反演的模型精度，但高于单一植被指数与植被覆盖度反演模型的精度；（3）精度检验显示植被指数逐步回归分析法反演的植被覆盖度模型的决定系数（R2）和精度分别为0.719和86.70%，而混合像元分解和植被指数逐步回归分析综合所建的混合模型的决定系数（R2）和精度分别为0.807和92.37%，表明植被指数逐步回归分析与混合像元分解相结合能较好地提高荒漠化地区植被覆盖度反演精度。