Assessment of Artificial Neural Network through Drought Indices

Abstract

Prediction of potential evapotranspiration (PET) using an artificial neural network (ANN) with a different network architecture is not uncommon. Most researchers select the optimal network using statistical indicators. However, there is still a gap to be filled in future applications in various drought indices and of assessment of location, duration, average, maximum and minimum. The objective was to compare the performance of PET computed using ANN to the Penman–Monteith technique and compare drought indices standardized precipitation index (SPI) and standardized precipitation evapotranspiration index (SPEI), using two different computed PET for the durations of 1, 3, 6, 9, and 12–months. Statistical performance of predicted PET shows an RMSE of 9.34 mm/month, RSR of 0.28, R2 of 1.00, NSE of 0.92, and PBIAS of −0.04. Predicted PET based on ANN is lower than that the Penman–Monteith approach for maximum values and higher for minimum values. SPEI–Penman–Monteith and SPI have a monthly correlation of greater than 0.95 and similar severity categories, but SPEI is lower than SPI. The average monthly index values for SPEI prediction show that SPEI–ANN captures drought conditions with higher values than SPEI–Penman–Monteith. PET–based ANN, performs robustly in prediction, fails by a degree of severity classification to capture drought conditions when utilized.

Keywords

• artificial neural network
• drought indices
• statistical analysis
• evapotranspiration
• United Kingdom