GA based heuristic approach to evaluate optimal hyperparameters of RNN-LSTM model for effective crop yield prediction

Crop yield prediction is an important and essential field of agriculture. An accurate and timely prediction of crop yield can be of great help to farmer and government of a nation in ensuring food for masses. The advancements in the field of deep learning have led to an accurate and précised crop yield predictions. Recurrent Neural Networks embedded with Long Short-Term Memory (RNN-LSTM), one of the recent deep learning models, are quite efficient in dealing with time series data henceforth making them one of the most suitable models for crop yield prediction. The efficiency of the model largely depends on various hyperparameters used in the model such as window size and number of neurons. The selection of a correct window size and an optimum number of neurons in hidden layers increases the efficacy of model without unnecessarily increasing the size of the network. The present study proposes a Genetic Algorithm (GA) based heuristic approach to (i) Identify optimum values for two hyperparameters (window size and number of neurons) used in RNN-LSTM model and to (ii) Train the model with optimal hyperparameters to predict wheat crop yield for one of the Punjab regions. The prediction results of wheat crop for optimum pair of hyperparameters: window size and number of neurons in hidden layers, 7 and 65 respectively, exhibited lowest RMSE value of 90 kgha⁻¹. The results proved the effectiveness of proposed model and revealed the importance and significance of finding optimum values of different hyperparameters for more accurate prediction results.