Attention-Guided Bidirectional Temporal Modelling with Graph-Based Regional Spatial Context for Bajra Crop Yield Prediction

Abstract: Bajra (pearl millet) is a very important crop in Rajasthan, India, since it is drought-resistant, nutritious, and culturally important. But its productivity is becoming vulnerable to changes in climate, such as erratic rain and temperature changes, and thus precise estimation of yield is vital. Crop Yield Prediction (CYP) indicators like soil decomposition, rainfall and meteorological patterns are slowly evolving, exhibiting long-term temporal dependency and propagating over time. Conventional cropping prediction algorithms based on artificial intelligence process the historical data and these indicators in a unidirectional manner. While mapping the temporal dependencies, these algorithms consider each year independently and do not capture the delayed effect, like salt degradation. To address this issue, the study proposes a region-based spatiotemporal model with an attention-guided Bidirectional LSTM (Long-Short Term Memory) framework for CYP, termed as G-BiLSTM. The proposed model reproduces the spatial relationships between districts via GCN (Graph Convolution Network) -based immediate neighbour extraction. Further, a Bidirectional LSTM is used to model multi-year CYP temporal features, allowing each annual observation to be encoded using both past and future temporal context. A variance-reduced and comprehensible representation is produced by integrating an attention mechanism to adaptively highlight the most informative years within a temporal window. Using 15 agroenvironmental characteristics, including understudied elements like saline and alkaline soil composition, the framework is assessed on a dataset that includes 32 districts in Rajasthan over 13 years (2007–2019). The suggested attention-enhanced BiLSTM consistently outperforms traditional temporal models, achieving lower prediction error and better generalisation, according to experimental results analysis using a three-year sliding temporal window. For regional crop yield forecasting, the suggested method offers a scalable solution.

Mamta Kumari, Suman and Devendra Prasad. “Attention-Guided Bidirectional Temporal Modelling with Graph-Based Regional Spatial Context for Bajra Crop Yield Prediction”. International Journal of Advanced Computer Science and Applications (IJACSA) 17.2 (2026). http://dx.doi.org/10.14569/IJACSA.2026.01702107

@article{Kumari2026,
title = {Attention-Guided Bidirectional Temporal Modelling with Graph-Based Regional Spatial Context for Bajra Crop Yield Prediction},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2026.01702107},
url = {http://dx.doi.org/10.14569/IJACSA.2026.01702107},
year = {2026},
publisher = {The Science and Information Organization},
volume = {17},
number = {2},
author = {Mamta Kumari and Suman and Devendra Prasad}
}