Multilayer feedforward networks are universal approximators
Neural Networks
Neural networks and the bias/variance dilemma
Neural Computation
International Journal of Remote Sensing - Satellite observations of the atmosphere, ocean and their interface in relation to climate, natural hazards and management of the coastal zone
Improved estimation of surface solar insolation using a neural network and MTSAT-1R data
Computers & Geosciences
| Hi-index | 0.00 |
Surface solar insolation (SSI) is an important parameter for interpreting ocean-atmosphere interactions, climate change, surface heat flux, and the Earth's radiation budget. The successful calculation of SSI from satellite data depends strongly on how cloud attenuation is described because most clouds have large spatial and temporal variability and complicated physical characteristics. Moreover, the accuracy of SSI estimation under cloudy conditions is substantially lower than under clear skies. We have generated a neural network (NN)-based cloud factor retrieval system that improves SSI estimation accuracy under cloudy conditions. We used a multilayer feedforward NN with Levenberg-Marquardt backpropagation and an early stopping method to avoid overfitting. The number of hidden nodes was determined by trial and error because a too complicated network is apt to overfit, whereas a too simple network makes training the network difficult. Validation of the estimated SSI using the NN-based cloud factor was performed with pyranometer data obtained from 22 meteorological stations over the Korean peninsula. This SSI estimation for cloudy conditions showed good agreement with ground-based measurements: root mean square error (RMSE) = 67.38 W m-2; standard error (SE) = 54.78 W m-2. This accuracy indicates that the use of an NN-based cloud factor improves SSI estimation over the previous cloud factor system (SSIT: RMSE = 78.03 W m-2, SE = 52.64 W m-2) and the multiple regression-based cloud factor (SSIMR: RMSE = 79.20 W m-2, SE = 67.55 W m-2).