Forecasting Paddy Yield in Sri Lanka Using Back-propagation Learning in Artificial Neural Network Model

Climate change has a direct and indirect impact on food production, and food production depends on the available resources such as the quantity of seeds sown, climate, soil moisture, solar radiation, expected carbon, fertilizers, pesticides, government policies, etc. Paddy yield is one of the major contributors to food production, and there are several studies on forecasting paddy yield production in Sri Lanka using common algorithms in ANNs, this study focused on forecasting the paddy yield in Sri Lanka based on some climate factors using the selected best steepest descent optimizer algorithm for Backpropagation learning (BP) in the Artificial Neural Network (ANN) model. The input and output for the ANN model were selected as climate factors and average seasonal paddy yield respectively. The seasonal data from 1981-2020 were collected from the website of the Department of Census and Statistics, Sri Lanka, and the NASA POWER project which provides solar and meteorological data collected from 1981-2023. The min-max normalization process was used to rescale the data in [0,1], and the parameters of the ANN model were selected by trial and error. First, built-up an ANN model to forecast the yield of paddy in key districts in Sri Lanka. The forecasting accuracy of the steepest descent optimizer algorithms which are the Batch Gradient Descent Method (BGD), Stochastic Gradient Descent Method (SGD), and Mini-batch Gradient Descent Method (MBGD), for BP learning in the ANN model was measured by Mean Square Error. The results show that the ANN model with BP learning is suitable for Paddy yield prediction, and BGD was the best optimization algorithm for BP learning in this study. Hence, paddy yields of key districts in Sri Lanka were forecasted from 2021-2023 based on some the climate factors.