Rice is one of the significant crops and Karnataka is the key regions in India for rice production, which produces more rice yields based on its fertile soil and climatic conditions. For agricultural planning in ensuring food security, there is a need for an accurate prediction of accelerated urbanization. The linear statistics approach was used earlier to predict rice yield that faces limitations in analyzing the nonlinear, complex relationship regarding the variability of crop, climate and soil. In addition, for the dynamic environment, it is not adaptable and highly dependent on the historical data, which often limits the accuracy and leads to poor decision-making for real-time agricultural data. Therefore, this work develops an efficient model for predicting the rice yield by leveraging real-time data sources, optimizing performance, interpreting model predictions, and allowing the management and planning of resources more efficiently, thus contributing to food security and sustainable agricultural practices. Initially, the model collected the real-time weather data of the agricultural region in different districts in Karnataka. Then, the collected input data are individually provided to the pre-processing unit. The pre-processing processes, such as filling missing values, min-max and standard scaling normalization, process the provided multi-modal data individually and enhance the quality of the data by minimizing the potential artefacts and provide three distinct pre-processed outcomes. Further, the three distinct pre-processed outcomes are provided separately to the Temporal Feature analysis module that comprises of a 1-Dimensional Convolutional Layer and Transformer layers (Conv1D+Trans Layers), which yield features such as F1, F2, and F3. These acquired features are fed to the Multi-Head Cross-Attention Fusion (MH-CAF) network for feature fusion. Then, the fused features are further analyzed in the newly developed Adaptive Residual Multi-Objective Loss Spatio-Temporal Network (ARes-MOL-STNet) for rice yield prediction. To enhance the performance of the prediction model, an Enriched Calving season of Elk Herd Optimizer (ECEHO) strategy is applied to optimize the novel parameters of the developed ARes-MOL-STNet model. Here, the rice yield prediction has been done as a regression task, and thus the MAE of the proposed model attained a minimal value compared to the state-of-the-art models with values 70% of Random Forest, 40% of TCN, 60% of CNN-LSTM, and 20% of Res-MOL-STNet, respectively. This indicates that the proposed deep learning model can achieve stable prediction results across regional datasets, and also aid in agricultural management.
G et al. (Fri,) studied this question.