Estimation of physical variables from multichannel remotely sensed imageryusing a neural network: Application to rainfall estimation

Satellite-based remotely sensed data have the potential to provide hydrolog ically relevant information about spatially and temporally varying physical variables. A methodology for estimating such variables from multichannel r emotely sensed data is presented; the approach is based on a modified count erpropagation neural network (MCPN) and is both effective and efficient at building complex nonlinear input-output function mappings from large amount s of data. An application to high-resolution estimation of the spatial and temporal variation of surface rainfall using geostationary satellite infrar ed and visible imagery is presented. Test results also indicate that spatia lly and temporally sparse ground-based observations can be assimilated via an adaptive implementation of the MCPN method, thereby allowing on-line imp rovement of the estimates.