Abstract
The enhanced thematic mapper (plus)(ETM+) instrument on Landsat 7 shares the same basic design as the TM sensors on Landsats 4 and 5, with some significanti mprovements. In common are six multispetral bands with a 30-m ground projected instantaneous field of view (GIFOV). However, the thermal-IR (TIR) band now has a 60-m GIFOV, instead of 120- m. Also, a 15-m panchromatic band has been added. The artificial neural network (NN) image sharpening method described here uses data form the higher spatial resolution ETM+ bands to enhance (sharpen) the spatial resolution of the TIR imagery. It is based on an assumed correlation over multiple scales of resolution, between image edge contrast patterns in the TIR band and several other spectral bands. A multilayer, feedforward NN is trained to approximate TIR data at 60m, given degraded (from 30-m to 60-m) spatial resolution input from spectral bands 7,5, and 2. After training, the NN output for full-resolution input generates an approximation of a TIR image at 30-m resolution. Two methods are used to degrade the spatial resolution of the imagery used for NN training, and the corresponding sharpening results are compared. One degradation method uses a published sensor transfer function (TR) for Landsat 5 to simulate sensor coarser resolution imagery form higher resolution imagery. For comparison, the second degradation method is simply Gaussian low pass filtering and subsampling, wherein the Gaussian filter approximates the full width at half maximum amplitude characteristics of the TF-based spatial filter. Two fixed-size Nns (that is, number of weights and processing elements) were trained separately with the degraded resolution data, and the sharpening results compared. The comparison evaluates the relative influence of the degradation technique employed and whether or not it is desirable to incorporate a sensor TF model. Preliminary results indicate some improvements for the sensor model-based technique. Further evaluation using a higher resolution reference image and strict application of sensor model to data is recommended.
