Optimization of Rice Yield Prediction for Hyper Parameters of Artificial Neural Network

Abstract

Introduction: Development of neural network models An effective Artificial Neural Network (ANN) requires proper structural design and tuning of hyperparameters. Research in agricultural forecasting shows that ANNs have great potential to manage complex nonlinear relationships between variables. Climate, Hydrology and Environment. Accurate rice yield forecasting is extremely important for Thailand as rice is a major economic crop that plays a role in the country's food security and economic stability.

Objectives: The objectives of this research is to develop a neural network model. An Artificial Neural Network (ANN) that has been optimized for forecasting Thailand's annual rice yield. It focuses on tuning hyperparameters to increase model accuracy and stability, thereby enhancing forecasting efficiency to support agricultural planning and resource management.

Methods: The study uses annual data collected over 12 years (2013–2024), data on 5 government organizations and determined 34 predictor series, under various climate parameters, water resources, solar energy influences – rainfall and temperature, relative humidity, and solar radiation, as well as reservoir sizes. All the data was scaled with Min–Max Scaler which is used to make up the model stable and reduce variation with the range of the data. Testing involved analyzing the ANN structures with 1-2 hidden layers, and tuning the important hyperparameters ie. number of nodes, batch size, dropout rate, learning rate, activation function, optimiser, number of training epochs to reach the structure of the model which produces the best forecasting results.

Results: The optimized ANN model demonstrated excellent predictive performance, achieving an R² value of 0.9757, RMSE of 0.0780, MRAE of 0.0121, MASE of 0.0128, and MAE of 0.0061, with an overall prediction accuracy of 98.40%. These results confirm that appropriate hyperparameter tuning substantially improves ANN forecasting capability for annual rice yield estimation.

Conclusions: The findings indicate that a well-configured and properly tuned ANN model can significantly enhance the accuracy of rice yield forecasting in Thailand. The superior performance achieved using optimizers such as Adam and RMSProp highlights the effectiveness of advanced optimization techniques in ANN training. This modeling approach provides valuable insights for monitoring rice production, supporting water resource planning, and accelerating the development of smart agriculture systems in Thailand.