Design of a Multilayer Perceptron Network Based on the Normalized Histogram to Develop Yields Predictive Model

Abstract

This article presents a multilayer perceptron (MLP) for patio-temporal remote sensing analysis of satellite image time series for a yield prediction model. This paper illustrates the model for crop yield prediction from the spatial cumulation of normalized difference vegetation index (NDVI) image time series at the province level. The methodological framework comprises the transformation of each NDVI image into a histogram to ensure no loss of information in the mapping of high- dimensional/unstructured NDVI images into pixels consideration, to be then used in training the MLP. The research work also includes an analysis of several activation functions for the hidden layer and testing their consequences on accuracy, including Radial-Basis (RadBas), Logarithmic-Sigmoid (LogSig), Hyperbolic-Tangent Sigmoid (TanSig), which depend mainly on exponential functions and limit the amplitude of the output. The proposed approach was utilized to predict the winter crop yield in Diwaniyah-Iraq province, one of the main agriculture regions, whose economy considerably depends on crop production. It can also be extended to other crops and other regions in Iraq. The proposed methodology showed the ability to predict the crop yield around seven-nine weeks before harvest. It also, outperformed the performance of traditional approaches by transforming the input into a more convenient form that reflects more useful information. The results show that the proposed model provides efficient accuracy (determination coefficient R2 > 0.85 and error level <0.24).