Assessment of a fully soft classification approach using back propagation neural network in estimating of rice growing area

摘要

Mixed pixels are a major problem as a conventional classification will force the allocation of a mixed pixel to one class, which need not even be one of the component classes of that pixel. Since the conventional classification output is "hard", comprising of only the code of the allocated class, such techniques cannot therefore be used appropriately to represent mixed pixels. The fully soft classifications were used to accommodate mixed pixel problem at each stage of classification. More than 90% of rice is planted in southern China where population density is very high and rice planting is often conducted by unit of single firmly, thus the size of paddy field patches are very small and the shape of those are not often irregular. For estimating rice-growing field area using remotely sensing data, the mixed pixel problems are more severe. In this study, an approach to achieve such a fully soft classification using back propagation neural network (BPN) in the rice growing region was assessed. The remote sensing data used in this study is a simulated imagery from TM data and a rice field map investigated by GPS. It was found this approach can improve significantly classification accuracy for rice-growing field harden mapping and total area estimating at sub-pixel level.