
Arvind kumar
Former Vice- ChancellorRani Lakshmi Bai Central Agricultural Uni., Jhansi, U.P., INDIA

Submitted
2026/07/22First Online
2026/09/10doi
10.18805/ag.D-6558Cite article - M. Lokeshwari, Nivethitha Manavalagan, P. Amrutha, S. Harinee, N.R. Divyasree. (2026). An Optimized Temporal Attention-based Deep Learning Framwork for Accurate Crop Yield Prediction from Multi-source Agricultural Data. Agricultural Science Digest. 46: 107–118. doi: 10.18805/ag.D-6558
Background: In precision agriculture, crop yield prediction is crucial for supporting effective management of crops, timely settlement of farmers crop insurances, food security planning and serve as decision tool for sustainable agricultural achievements. However, prediction of crop yield is difficult due to the complex interaction behavior among diverse conditions in weather, soil properties, multispectral remote sensing observations and the dynamics in crop growth. Despite existing AI techniques like machine learning and deep learning models outperformed in predicting crop yield, they are often not able to capture the temporal dependencies. Furthermore, determining optimal hyperparameters for model training remains a challenging task, significantly influencing prediction accuracy, model convergence and generalization performance.
Methods: To overcome this, the present study aimed to predict the crop yield using temporal attention (TA) based Artificial bee colony (ABC) optimized Bi - LSTM model using multi source agricultural data. The proposed model learns the sequential and temporal dependencies from both past and future time steps simultaneously, from the historical weather variables, soil properties, remote sensing derived normalized difference vegetation indices (NDVI) and historical yield data. The ABC algorithm was employed to automatically optimize the model hyperparameters with minimum run time.
Result: The model is trained using district wise rice yield records of Tamil Nadu, historical weather variables, Sentinel-2 derived NDVI and soil properties. The dataset collected for major rice growing districts from 1995 to 2025 (30 years). The model shows the superior performance over conventional methods like random forest, support vector regression, deep neural network, convolutional neural network, long short-term memory models with lesser root mean square error (RMSE = 298.01 kg ha -1 ) and higher coefficient of determination, R 2 (81.76%). Also, the training time of TA - ABC - Bi-LSTM model is comparatively less than the comparative models because of ABC optimization algorithm tuned the best hyperparameters compared to traditional trial and error manual method.
Distinguished scholars and subject experts committed to editorial excellence, scientific integrity and rigorous peer review.
View Full Board