基于光谱-环境随机森林回归模型的MODIS积雪面积比例反演研究

积雪面积比例（Fractional Snow Cover， FSC）数据能在亚像元尺度上定量的描述像元内积雪覆盖的程度，相比二值积雪面积数据可以更加精确地估计积雪覆盖的面积。基于机器学习的随机森林回归模型可以表示高维的非线性关系，可显著提高MODIS FSC的反演精度。采用随机森林回归模型结合光谱、环境信息构建了一个新的回归模型——光谱-环境随机森林回归（Spectral Environment Random Forest Regressor， SE-RFR）模型，用于MODIS数据反演中国区域的FSC。利用中国典型积雪区内由Landsat 8地表反射率数据获取的FSC数据作为参考值，对SE-RFR模型的反演精度进行评估。研究表明，利用“SE-RFR”获取的FSC数据RMSE、MAE分别为0.160、0.104，精度较高。此外，根据SE-RFR模型与未加入环境信息的随机森林回归（S-RFR）模型比较结果可知，加入环境信息的随机森林回归模型提高了FSC反演的精度，特别是在受环境信息影响较大的青藏高原地区，RMSE从0.200降低到0.181。最后，将SE-RFR模型与目前使用广泛的MODIS FSC反演模型FSC_NDSI、MODSCAG和SSEmod进行了比较，结果表明SE-RFR模型的RMSE与FSC_NDSI、MODSCAG和SSEmod模型的RMSE相比，平均RMSE分别提高了12.0%、8.3%和5.5%。总体来说，SE-RFR模型可以准确地提取MODIS FSC，对于区域乃至全球FSC产品制备具有广泛的应用前景。

Research on retrieval of MODIS fraction snow cover based on spectral environmental random forest regression model

The fractional snow cover （FSC） data can quantitatively describe the extent of snow cover in a pixel on the sub-pixel scale， and can estimate the area of snow cover more accurately than binary snow area data. The random forest regression model based on machine learning can represent high-dimensional nonlinear relationships， which can significantly improve the inversion accuracy of MODIS FSC. In this study， a new regression model， Spectral Environment Random Forest Regressor （SE-RFR） model， was constructed using random forest regression model combined with spectral and environmental information， which was used to retrieve the FSC from MODIS data in China. We used the FSC obtained from Landsat 8 surface reflectance data in a typical snow area in China as a reference value to evaluate the inversion accuracy of the SE-RFR model. Research shows that the RMSE and MAE of FSC data obtained by SE-REF are 0.160 and 0.104， respectively， which has high accuracy. The SE-RFR model is compared with the Spectral Random Forest Regressor （S-RFR） without environmental information. It shows that the random forest regression model with environmental information improves the accuracy of FSC inversion， especially in the Qinghai-Tibet Plateau region， which is influenced by environmental information， and the RMSE decreased from 0.200 to 0.181. Finally， the SE-RFR model was compared with the currently widely used MODIS FSC inversion models FSC_NDSI， MODSCAG and SSEmod. The results showed that the average RMSE of the SE-RFR model is increased by 12.0%， 8.3% and 5.5%， respectively， compared with the RMSE of the FSC_NDSI， MODSCAG and SSEmod models. In general， the SE-RFR model can accurately extract MODIS FSC， which has wide application prospects for the preparation of regional and even global FSC products.