The Backpropagation Neural Networks Modelling Using Probabilistic Neural Network for Monsoon Onset Prediction Based on Global Climate Indices

Abstract

Monsoon-onset is the key factor in agricultural production especially for crops. Monsoon-onset over Indonesia region is influenced by global climate phenomena such as El Nino Southern Oscillation (ENSO), Indian Ocean Dipole- Mode (IOD), Southern Oscillation Index (SOI) and El Nino Modoki (EMI). And thus, we used the global climate indices as predictors for neural networks input model. In this research we build a prediction model using Backpropagation Neural Networks (BPNN) method and Probabilistic Neural Networks (PNN) as classifier for monsoon-onset classification. This research has an objective to predict the monsoon-onset over Indramayu district with agronomic approach definition using Backpropagation Neural Networks and Probabilistic Neural Networks. Because of Indramayu district is one of the rice field center in Indonesia, the agronomic approaches that been used is especially for crops which has determined in Moron et al. (2009) research. This research divide into two scheme. First, monsoon-onset prediction using backpropagation neural networks model in five different architectures. This architectures consist of different hidden layer and hidden neuron combination. And the second is, monsoon-onset prediction using Backpropagation Neural Networks -based on first scheme results- with probabilistic Neural Networks as classifier. The Probabilistic Neural Networks are used for classify the monsoononset class in validation process for each new evaluated data. The first scheme shows that the backpropagation neural networks using 1 hidden layer and 10 hidden neuron as the best results with r 0.89 and RMSE 11.6. This results occurred at Kertasemaya location point. In the second scheme, Kertasemaya location point also gave us the best results in prediction with r , PNN accuration and RMSE as follows 0.85, 87.5% and 10.3. This results give us better number than previous scheme, with some notes. Taylor Diagram was used as a performance analysis tools for models evaluation. The last model (models which is developed in second scheme) become the best prediction model for 3 out 8 location points in the research. And also, we build a correlation map based on location points spreads in order to point out the spatial pattern. The final results shows that monsoon-onset over Indramayu district can be determined using agronomic approach. Monsoon-onset prediction with agronomic approaches has been built using Backpropagation Neural Networks models and Probabilistic Neural Network as classifer, also global climate indices as predictors.