积雪是地球科学领域的重要变量，作为冰冻圈的重要组成要素之一，其分布与变化深刻影响着水循环、辐射能量平衡、气象和气候变化。亚洲水塔是以青藏高原为核心的“第三极”，是除南极和北极以外冰雪储量最大的地区，该地区在全球气候变化中表现更为敏感。随着全球变暖的加剧，亚洲水塔区长时间序列积雪监测对认识亚洲水塔区气候变化乃至全球气候系统都具有重要的意义。

现有光学传感器中，先进的甚高分辨率辐射计（AVHRR）是目前国际上唯一可提供40年长时间序列、大尺度积雪覆盖区域每日观测的重要数据源，但目前基于AVHRR的积雪覆盖产品主要为二值化积雪面积数据，对于混合像元区域将会高估或低估实际的积雪覆盖面积。而积雪覆盖度（FSC）则能够反映像元内的积雪覆盖比例，可以获取更加精准的积雪覆盖信息。本研究针对AVHRR传感器波段特性，依据现有光学遥感积雪监测方法改进应用四种积雪覆盖度反演算法进行亚洲水塔区2015—2019年积雪季AVHRR积雪覆盖度反演，并构建Landsat-8/OLI积雪覆盖度验证数据集对AVHRR积雪覆盖度反演结果进行精度验证与定量评估，探究最适用于制备亚洲水塔区长时序积雪覆盖度产品的反演算法。具体研究工作及成果有以下三个方面：

（1）基于GEE（Google Earth Engine）平台发展了一种针对AVHRR的亚洲水塔区二阶段随机森林（RF）积雪覆盖度反演方法，第一阶段使用随机森林分类算法将像元分为无雪、有雪、纯雪三类，第二阶段使用随机森林回归算法估算有雪像元（积雪覆盖度介于0%至100%之间）的具体积雪覆盖度。其中，选取了研究时段内500 m空间分辨率的MODIS积雪覆盖度影像用于随机森林模型的训练，并利用AVHRR可见光、中红外等波段地表反射率、积雪指数、归一化植被指数等光谱信息、高程、坡度、坡向、地表覆盖类型等地形信息以及日期、经纬度等时空信息作为随机森林模型输入数据，反演亚洲水塔区基于AVHRR的积雪覆盖度。

（2）改进了适用于亚洲水塔区的基于AVHRR影像的三种积雪覆盖度反演方法：针对AVHRR传感器特点，将三种现有光学遥感数据FSC反演方法（Pan等，2022）进行改进，并应用至亚洲水塔区AVHRR积雪覆盖度的反演，包括以下三种AVHRR积雪覆盖度反演算法：基于积雪指数的线性回归算法（FracSI）、基于纯雪-非雪背景的二端元模型算法（TEM）、基于自动端元提取的多端元光谱混合分析算法（MESMA）。

（3）评估亚洲水塔区上述四种反演算法精度。验证结果表明多端元光谱混合分析算法结合了AVHRR三个波段的观测信息，利用逐像元动态端元矩阵提高了像元端元反演精度，总体精度表现最优，而更依赖训练样本确定经验关系的二端元模型与基于积雪指数的回归算法精度次之，基于GEE平台的随机森林方法相对较差。四种AVHRR积雪覆盖度反演算法均在裸地地区的反演效果最好，草地地区次之，山区和森林地区较差，一方面积雪与植被覆盖的混合像元积雪指数较低、森林冠层遮挡了积雪的反射信号；另一方面山区地表异质性大，积雪更为破碎，且存在山体阴影导致积雪信息难以提取。四种AVHRR积雪覆盖度反演算法在低海拔区反演精度较高，随着海拔升高精度随之下降，尤其在4500 m以上的高海拔区，也与高海拔山区分布有关。基于自动端元提取的多端元光谱混合分析算法（MESMA）在亚洲水塔区不同地表覆盖类型及不同海拔区间内均具备较好的适应性和稳定性，可作为后续亚洲水塔区AVHRR长时间序列积雪覆盖度产品的制备算法。

Snow cover, as an important variable in the field of Earth sciences and a crucial component of the cryosphere, profoundly affects water cycle, radiation energy balance, meteorology, and climate change with its distribution and changes. The Asian Water Tower, which is centered on the Qinghai-Tibet Plateau, is the region with the largest snow and ice reserves after the Antarctic and Arctic regions. This area is more sensitive to global climate change. With the exacerbation of global warming, long-term monitoring of snow cover in the Asian Water Tower region is of great significance for understanding climate change in this region as well as the global climate system.

Among the existing optical sensors, AVHRR is the only important data source that can provide daily observations of large-scale snow cover areas with a 40-year long time series. However, the current AVHRR-based snow cover products mainly provide binary snow cover area data, which will lead to overestimation or underestimation of actual snow cover area in mixed pixel areas. In contrast, Fractional Snow Cover (FSC) can reflect the proportion of snow cover within a pixel and obtain more accurate snow cover information. In this study, four FSC retrieval algorithms were improved and applied to the AVHRR FSC retrieval for the Asian Water Tower region during the snow season from 2015-2019 based on the band characteristics of the AVHRR sensor and existing optical remote sensing snow monitoring methods. A Landsat-8/OLI FSC validation dataset was constructed to verify and quantitatively evaluate the accuracy of AVHRR FSC, and explore the most suitable retrieval algorithm for preparing long-term FSC products in the Asian Water Tower region. The specific research work and results are threefold:

(1) A two-stage random forest (RF) method for snow cover retrieval over the Asian Water Tower region using AVHRR data was developed based on the Google Earth Engine (GEE) platform. The first stage represents a classification step that predicts a class of zero, one, or non-boundary values (FSC in the open interval (0,1)). For the non-boundary classification (not zero or one), a second random forest predicts values between zero and one FSC. This method utilized AVHRR visible and middle-infrared band surface reflectance, snow index, normalized difference vegetation index (NDVI), as well as auxiliary data including elevation, slope, aspect, land cover type, and date/longitude/latitude information to train the RF model with MODIS FSC images at 500-m resolution during the study period.

(2) Three existing FSC retrieval methods based on AVHRR were improved and applied to the Asian Water Tower region by considering the characteristics of the AVHRR sensor. The three AVHRR-based FSC retrieval methods are: linear regression algorithm based on NDSI (FracSI), snow/ non-snow two endmember model (TEM), and multiple endmember spectral mixture analysis algorithm (MESMA).

(3) The results of accuracy validation of the four AVHRR FSC retrieval methods in the Asian Water Tower region show that MESMA FSC has the best overall accuracy, followed by TEM FSC and FracSI FSC, while RF FSC is relatively poor. Because the MESMA algorithm combined with the observation information from the three bands of AVHRR. By utilizing a per-pixel dynamic endmember matrix, it improves the accuracy of endmember retrieval for each pixel. The TEM and FracSI models relying more on training samples to determine empirical relationships show lower precision. All four AVHRR-based FSC retrieval methods perform better in bare ground areas, followed by grassland areas, and exhibit inferior performance in mountainous and forested areas. On one hand, the mixed pixels of snow and vegetation have lower snow indexes, and the forest canopy obstructs the reflection signal of snow. On the other hand, mountainous areas exhibit significant surface heterogeneity, with fragmented snow cover and the presence of mountain shadows, making it challenging to extract snow information. The accuracy of all four methods is higher in low-altitude areas and decreases with the increase in altitude, especially in extremely high-altitude regions above 4500 m, which is also related to the distribution of high-altitude mountainous areas. MESMA FSC retrieval algorithm shows good adaptability and stability in different land cover types and altitude ranges in the Asian Water Tower region, and can be used as a preparatory algorithm for long-term AVHRR snow cover products in the future.