A Novel Method to Estimate Subpixel Temperature by Fusing Solar-Reflective and Thermal-Infrared Remote-Sensing Data With an Artificial Neural Network

Among the multisource data fusing methods, the potential advantages of remote sensing of solar-reflective visible and near-Infrared [(VNIR); 400-900 nm] data and thermal-infrared (TIR) data have not been fully mined. Usually, a linear unmixed method is used for the purpose, which results in low estimation accuracy of subpixel land-surface temperature (LST). In this paper, we propose a novel method to estimate subpixel LST. This approach uses the characteristics of high spatial-resolution advanced spaceborne thermal emission and reflection radiometer (ASTER) VNIR data and the low spatial-resolution TIR data simulated from ASTER temperature product to generate the high spatial-resolution temperature data at a subpixel scale. First, the land-surface parameters (e. g., leaf area index, normalized difference vegetation index (NDVI), soil water content index, and reflectance) were extracted from VNIR data and field measurements. Then, the extracted high resolution of land-surface parameters and the LST were simulated into coarse resolutions. Second, the genetic algorithm and self-organizing feature map artificial neural network (ANN) was utilized to create relationships between land-surface parameters and the corresponding LSTs separately for different land-cover types at coarse spatial-resolution scales. Finally, the ANN-trained relationships were applied in the estimation of subpixel temperatures (at high spatial resolution) from high spatial-resolution land-surface parameters. The two sets of data with different spatial resolutions were simulated using an aggregate resampling algorithm. Experimental results indicate that the accuracy with our method to estimate land-surface subpixel temperature is significantly higher than that with a traditional method that uses the NDVI as an input parameter, and the average error of subpixel temperature is decreased by 2-3 K with our method. This method is a simple and convenient approach to estimate subpixel LST from high spatial-temporal resolution data quickly and effectively.