Artificial neural network response to mixed pixels in coarse-resolution satellite data

摘要: Abstract A feedforward neural network model based on the multilayer perceptron structure and trained using backpropagation algorithm responds to subpixel class composition in both simulated real data. Maps of response surfaces for data illustrate that set outputs successfully characterizes level dominance controlled contain a range mixtures. For Sierra Nevada test site, correspondence between 250 m reference map produced degraded TM depends degree subpixed mixing as determined from coregistered 30 most mislabeled pixels, classification error results confusion first second largest components, outputs. Overall accuracy increases 62% 79% when pixels are reclassified output. Accuracy 84% if, is used reference. subset Plumas complement findings show systematic way changing proportions components. Based our we suggest interpretation complete can provide information relative classes. We outline threshold-based heuristic would allow labeling pure classes, mixed primary secondary types magnitudes two signals.