Neural network method for sharpening Landsat thermal data from higher resolution multispectral data

Abstract

In this study, a sharpening method based on a neural network (NN) approximation technique is described to increase the spatial resolution of thematic mapper (TM) thermal-infrared (T-IR) data. Sharpening is derived from a learned input-output mapping of edge contrast patterns between TIR and higher resolution TM bands. This method is similar to a reported adaptive least squares (LS) method to estimate TM T-IR data at a higher resolution. However, there are two major differences: use of NN approximation instead of LS estimation, and application or reported multiresolution technique to adaptively combine spatial information from the original image and its highresolution estimate. With training examples from reduced resolution data, a multilayer feedforward NN is trained to approximate T-IR data samples on the basis of a possible nonlinear combination of data samples from three other TM bands. The trained NN output for full-resolution input data is an estimate of T-IR image at 30-m resolution. A potential benefit of this sharpening method is that the NN approximation technique can be developed from only a subset of image scene samples, and yet be applied to the entire scene. Preliminary examples show sharpening at four times higher resolution, but further evaluation with a high-resolution reference image is recommended. Although results are encouraging, a different training strategy to improve network generalization is suggested as a way to improve the sharpening process. Any use of trade, product, or firm names is for descriptive purposes only and does not imply endorsement by the U.S. Government.