ARTMAP Neural Network Classification of Land Use Change
Byron M. Shock, Gail A. Carpenter, Sucharita Gopal, and C.E. Woodcock · World Congress of Computers in Agriculture and Natural Resources, Proceedings of the 2002 Conference · 2013
The ability to detect and monitor changes in land use is essential for assessment ofthe sustainability of development. In the next decade, NASA will gatherhigh-resolution multi-spectral and multi-temporal data, which could be used fordetecting and monitoring long-term changes. Existing methods are insufficient fordetecting subtle long-term changes from high-dimensional data. This projectemploys neural network architectures as alternatives to conventional systems forclassifying changes in the status of agricultural lands from a sequence of satelliteimages. Landsat TM imagery of the Nile River delta provides a testbed for these land usechange classification methods. A sequence of ten images was taken, at various timesof year, from 1984 to 1993. Field data were collected during the summer of 1993 at88 sites in the Nile Delta and surrounding desert areas. Ground truth data for 231additional sites were determined by expert site assessment at the Boston UniversityCenter for Remote Sensing. The field observations are grouped into classesincluding urban, reduced productivity agriculture, agriculture in delta, desert/coastreclamation, wetland reclamation, and agriculture in desert/coast. Reclamationclasses represent land use changes. A particular challenge posed by this database isthe unequal representation of various land use categories: urban and agriculture indelta pixels comprise the vast majority of the ground truth data available in thedatabase. A new, two-step training data selection method was introduced to enable unbiasedtraining of neural network systems on sites with unequal numbers of pixels. Datawere successfully classified by using multi-date feature vectors containing data fromall of the available satellite images as inputs to the neural network system.