基于多层感知器模型的MODIS地表温度降尺度研究

LST（land surface temperature，地表温度）是一个极为重要的自然地理参数，能够表征地球表层系统的多个自然地理过程，同时与人类生产生活密切关联.为解决LST获取中时空分辨率不能自动互补的问题，以北京市为研究案例区，构建MLP（multilayer perceptron，多层感知器）模型，对MODIS（Moderate Resolution Imaging Spectroradiometer）1 km LST数据进行降尺度研究.结果表明:预测白天LST的MLP模型的训练、测试和整体决定系数分别为0.810、0.796和0.807，预测夜间LST的MLP模型的训练、测试和整体决定系数分别为0.702、0.705和0.701，预测残差值均服从正态分布，构建的模型可靠性较高.与回归模型和支撑向量机等模型相比，MLP模型具有拟合优度高、误差小等优点，此外，MLP模型测试数据集的拟合度也较高，说明模型的泛化推广能力较强.MLP模型的LST降尺度结果能够清晰地反映下垫面地表热环境的空间异质性和昼夜差异.平原地区的建设用地区为LST的高值区，山区远郊县为LST的低值区，白天LST显著高于夜间LST.将TM影像反演的LST的空间分辨率聚合到250 m，使之与MLP模型的LST降尺度结果的空间分辨率相同，并通过随机采50 000个样点比较TM影像反演的LST和MLP模型的LST降尺度结果.检验结果表明，尽管二者在具体数值上有些差异，但其空间结构高度相似，协方差为正，相关系数可达0.730，误差呈现正态分布.研究显示，人工神经网络模型在LST降尺度方面具有较大的应用前景. 

Downscaling MODIS Land Surface Temperatures Using Multilayer Perceptron Model

Land surface temperature (LST) is an important geographic parameter which can indicate multiple geographical processes of the land surface system. It is also closely related to human production and life. In order to synchronize the spatial and temporal resolution of remotely sensed land surface temperatures, a multi-layer perceptron (MLP) was constructed to downscale MODIS land surface temperature data by taking Beijing as a study area. Compared with regression models and support vector machines, the results of MLP showed higher model fitness and smaller error. The determination coefficients of the MLP''s train, test and all datasets were 0.810, 0.796 and 0.807 in the day, and 0.702, 0.705 and 0.701 at night, respectively. The downscaled LST could clearly reflect the spatial heterogeneity and differences between the day and night of the underlying surface thermal environment. In order to validate the performance of the MLP model, retrieved LST based on TM images with spatial resolution 60 m was aggregated to 250 m, which is the same resolution of the downscaled LST based on MLP. By random sampling and comparison, both spatial patterns were found to be highly similar, with a correlation coefficient of 0.730. The errors were normally distributed, although some LST values at specific pixels were different. This research suggests that the artificial neural network model has great application potential in the field of LST downscaling.