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

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