A combination of temporal thresholding and neural network methods for classifying multiscale remotely-sensed image data

A thresholding technique applied to a time series of 1 km, biweekly composited AVHRR-NDVI data stratifies a Sierra Nevada test site into broad vegetation classes based on temporal habit. These classes provide a basis for developing strata specific neural network classifiers which operate on single date 250 m and 500 m data simulated from Landsat Thematic Mapper, as well as 1000 m digital elevation data. The results of this combined hierarchical approach are compared to results from a network developed for the site as a whole without prior stratification. The artificial neural networks are feedforward models based on the multilayer perceptron structure trained by a backpropagation algorithm. The combined approach appears to perform slightly better than the sitewide model, although the results are comparable. The results from both approaches produced cover-type proportions which are closer to the proportions in the original 30 m reference class map than to the aggregated 250 m map used for training and testing the network models