基于BP神经网络的地表温度空间降尺度方法

Downscaling of Remotely Sensed Land Surface Temperature with the BP Neural Network

基于统计模型的降尺度方法被广泛用于热红外影像的尺度转换中,然而,大多数算法都会受到复杂地表环境的影响,例如地表覆盖、季节等。为了解决地表温度与光谱指数函数关系的不确定性,提出了一种新型的基于BP神经网络的地表温度降尺度方法。首先,在粗分辨率的情况下,训练得到一个以光谱指数为输入,原始温度为输出的BP神经网络。之后,输入高分辨率的光谱指数进而得到高分辨率的温度结果。实验通过设置多种光谱指数组合和BP网络隐藏层节点数而展开。结果评价时,以原始温度影像为参照,在城镇、植被和水体区域内,该方法的RMSE、R2、Bias及相对精度优于传统的分层线性回归降尺度方法。实测验证表明:该算法的RMSE和Bias分别达0.98℃、0.51℃,明显优于分层线性回归的结果(RMSE为2.9℃,Bias为1.7℃),说明该方法具有较高的降尺度精度,这对于城市热环境的研究具有一定的应用价值。

Downscaling algorithms based on statistical models have been widely utilized to address the issue of coarse-resolution Land Surface Temperature （I,ST）. However, most methods （e.g., TsHARP algorithm） could be affected by land environment,including land cover,seasons.In this study,a Back Propagation （BP） neural network was introduced for LST downscaling in a specific area with complex land covers.The meth- od comprises two steps.First, five reprehensive spectral indices were selected to training according to three typical land cover,including vegetation,building,and water.And the structure of network was trained using coarse-resolution spectral indices and LST. Second,high-resolution spectral indices were input to the net work to get a high-resolution LST.A stratified linear regression downscaling with land-cover classification was conducted for comparative evaluation. The comparative results showed that in urban, vegetation, and water areas, the Root Mean Square Error （RMSE）, determination coefficient （R2 ）, and relative accuracy for the proposed approach （BP neural network） were better than those for stratified linear regression.Finally, the verification results show that RMSE and bias of the algorithm are 0.98℃ and 0.51℃,which is obvi- ously better than the result of stratified linear regression （RMSE is 2.9℃ and Bias is 1.7℃）.It shows that this method has a higher downscaling accuracy.And the approach is potential for producing high-resolution LST for the study on urban thermal environment.