地表三分之一地区存在着季节性积雪，积雪的堆积和消融是冰冻圈最活跃的变化过程。积雪的多寡与存在形式变化影响着全球的能量平衡和物质转换。被看作地球第三极的青藏高原地区，是我国三大稳定积雪区之一。该地区积雪雪量在空间和时间维度的变化深刻影响着中国乃至高亚洲区域的水循环、能量交换等过程。然而，青藏高原上的积雪受气候、地形等因素影响，消融和堆积过程变化较快，幅度较大。近年来，学者们基于MODIS数据发展了多种积雪覆盖度(Fractional Snow Cover, FSC)算法。这些算法能否准确、高频率地监测积雪，需要进一步评估。 本研究以青藏高原为例，基于30m空间分辨率Landsat-8/OLI数据反演积雪覆盖度“真值数据”建立验证数据集。利用数据集中2013年至2015年149幅影像评估三种国际上较常用的MODIS积雪覆盖度产品精度。产品分别为第六代MODIS官方积雪覆盖度产品MOD10A1 Version 006，NASA-JPL提供的积雪产品MODSCAG (MODIS Snow Covered-Area and Grain Size)，以及施建成利用线性光谱混合分解算法反演的积雪覆盖度数据MODAGE (MODIS Automatic-selected Endmembers Fractional Snow Cover)。精度评估工作分别从二值化积雪判识精度和覆盖度反演精度两个方面进行，并针对不同地表覆盖类型，有效太阳入射角，观测尺度与积雪分布情况（积雪覆盖面积及其破碎程度）等条件分析三个产品的误差特点。精度评价结果显示，由于MODIS数据空间分辨率较低（500m），积雪覆盖度算法仍无法完全消除混合像元的影响。为了减小多种地物反射信号混合给算法带来的尺度效应，本研究尝试基于Landsat-8/OLI数据和混合像元分解算法，通过比较同种算法反演不同空间分辨率积雪覆盖度的结果差异，讨论尺度效应。分析导致积雪覆盖度出现尺度效应的因素，定量表达影响因素建立尺度校正公式，用以提高现有MODIS积雪覆盖度产品精度。 基于上述研究背景，本研究主要研究结果及结论如下： 1. 为规范化中低分辨率积雪算法与产品的精度评估，本研究以青藏高原地区为例，基于空间分辨率相对较高的Landsat-8影像，建立了一套积雪覆盖度验证数据集。首先利用2013年5月—2016年12月共计339幅Landsat-8/OLI地表反射率产品基于混合像元分解算法反演得到30m空间分辨率积雪覆盖度数据，随后利用像元聚合法将其至升尺度至500m建立验证数据集作为积雪覆盖度“真值数据”。验证数据集中还包括地表覆盖类型数据以及MODIS官方第六代地表反射率产品MOD09GA中波段1-波段7地表反射率、质量控制数据。 2. 利用上述验证数据集评估三种MODIS积雪覆盖度产品在青藏高原地区的精度，并分析误差来源。三种产品整体积雪判识精度可达到90%以上。但产品反演斑块状积雪时，易漏判积雪。本研究利用破碎度指标Fragmentation定量表达积雪的斑块状程度，分析发现MOD10A1与MODAGE产品漏判误差与积雪破碎度相关性较高。但当反演地区有效太阳入射角较大或地表为植被覆盖时，MOD10A1和MODSCAG则容易过判积雪。MOD10A1利用基于归一化积雪指数(NDSI)和FSC的线性经验回归关系式反演积雪覆盖度，像元NDSI值易受环境因素变化的影响，判识精度降低。MODSCAG利用混合像元分解算法反演积雪。模型模拟积雪端元时，无法针对不同光照条件分别模拟积雪的反射特征，因此该产品对有效太阳入射角的变化不够敏感。同时该产品的植被校正处理也使其过判植被区积雪。MODAGE直接从每张图像中提取端元。反演积雪破碎程度较高的图像时，算法无法提取到积雪端元信息，只能利用端元信息单一的参考端元库替代，减弱了MODAGE积雪判识能力。 空间分辨率500m时，MODSCAG和MODAGE积雪覆盖度反演精度较高。均方根误差(RMSE)分别为0.157和0.142，小于MOD10A1反演误差（RMSE为0.170）。前两者均采用混合像元分解算法，可对积雪混合像元进行基于亚像元级别的定量成分分析，一定程度上减小混合像元的影响。MOD10A1采用线性经验回归关系式反演积雪覆盖度，易受背景非雪地物的反射信号干扰。当空间分辨率缩小至2km时，三种产品的积雪覆盖度精度均有提升，RMSE减小约30%。但继续聚合至5km，精度再次降低。 3.探讨了积雪覆盖度出现尺度效应的原因，并建立尺度校正公式，有望改进中低分辨率（500m，1km等）的积雪覆盖度反演精度。本研究通过分析对比“先反演再聚合至大观测尺度”和“先聚合至大观测尺度再反演”积雪覆盖度结果差异发现，低空间分辨率下，积雪的不均一分布和不同观测尺度下地表反射率的非线性变化是积雪覆盖度产生尺度效应的主要因素。利用各像元积雪覆盖度标准差和平均值定量表达这两个影响因素，建立尺度校正公式。将校正公式运用于MODAGE 500m/1km/2km空间分辨率产品，RMSE减小约25%，产品精度得到提升，一定程度减少了尺度效应带来的误差。

Snow is one of the most dynamic components in the cryosphere. The amount and variation of snow cover have significant influences on global climate change, water cycling and radiation energy budget. The Tibetan Plateau, which is often called ‘the Third Pole’, is one of the principal perennial snowfields in China. Considering the amount and dynamic change of snow over this area, reliable daily snow cover data which are able to monitor snow variations across the Tibetan Plateau is in great demand. In recent years, researchers develop various fractional snow cover (FSC) algorithm and deliver products based on MODIS data. This paper used Landsat-8/OLI to establish evaluation dataset as ‘true value’, and evaluate three MODIS-based FSC products over the Tibetan Plateau. Including MOD10A1 from MODIS snow products version 6, MODSCAG (MODIS Snow Covered-Area and Grain Size), and MODAGE (MODIS Automatic-selected Endmembers Fractional Snow Cover). Forests, grass, bare soil and the Himalaya are chosen as typical land cover types. The assessment work includes 149 Landsat-8 images spanning a range over the Tibetan Plateau from 2013 to 2015. These three datasets are compared with the ‘true value’ using binary and fractional metrics. Considering the effects brought by observation scales, this paper also explores the scale effect of fractional snow cover. Firstly, using different upscaling methods to simulate the difference between fine and coarse resolutions. Analyzing the factors causing scale effects, then establishing scale adjustment method. Based on the above research, the main contents of this thesis can be concluded as follows: 1. This paper established FSC ‘true value’ evaluation dataset over the Tibetan Plateau. Including 339 Landsat-8 images from May, 2013 to December, 2016. The ‘true value’ is acquired by employing linear spectral mixture analysis to Landsat-8/OLI. Dataset contains Fractional Snow Cover ‘true value’, MOD09GA surface reflectance data (band1-band7), band quality control information and land cover type data. 2. In binary classification assessment, the overall snow identification precision is over 90%. Affected by the larger local illumination angle and vegetation-covered background, MOD10A1 overestimates snow. While it misses identifying snow due to the snow patchiness. MODSCAG shows superiority in correctly identify snow pixels, but overestimates snow at snow and snow-free boundary zone, and while the local illumination angle exceeds 30°. MODAGE could precisely recognize snow-free, but underestimates snow when the snow cover is patchy. In fractional snow cover assessment, MODSCAG and MODAGE obtain smaller root mean square errors (RMSE) than MOD10A1, probably due to the spectral mixture analysis is superior to normalized difference snow index (NDSI) based empirical method in retrieving FSC. Furthermore, while the spatial resolution decreases to 2km, RMSE of all products significantly decline of about 30%, but when data aggregated to 5km, the RMSE increase again. 3. In order to explore the scale effect of fractional snow cover. This research analyzes the differences between two different upscaling methods, this paper found that under coarse resolutions, the mean FSC value and FSC heterogeneity are main factors leading to FSC scale effects. So we developed the scale adjustment method based on these two factors. After applying the scale adjustment method to MODAGE, the retrieval accuracy has been improved. The RMSE is reduced of about 25% under various large observation scales.