MB-EFF: A Multi-Branch Environmental Feature Fusion Deep Learning Model for Crop Classification and Recommendation

Abstract

The motive of this project is to develop an accurate, reliable and scalable toward nature crop classification and recommendation system which examines the scenarios envisaged in the traditional machine learning modeling and integrates the multi heterogeneous soil and environmental data attributes handling capablity of deep learning. However, given the non-normality of this kind of heterogeneities of soil nutrients and climatic factors, the conventional nature models fail to characterize these complicated nonlinear connections as effectively as they can, which may lead to low prediction accuracy and poor generalization. To respond to these challenges this paper proposes a new deep learning architecture: Multi-Branch Environmental Feature Fusion (MB-EFF), which unites chemical soil features and environmental parameters into individual learning branches. The extracted representations are then recalibrated via a squeeze-and-excitation based feature recalibration mechanism, and subsequently subjected to deep nonlinear reasoning layers for enhanced discriminative learning. The model adopts a standardization on input features as pre-processing, trained with adam optimizer + early stopping applied to achieve stable convergence and robustness. This paper proposes an experimentally validated solution based on publicly available crop dataset and achieves 98% accuracy along with macro averaged & weighted averaged precision, recall and F1-score, being 0.98. As it shows an excellent performance near the perfect scores (1.0) for most of the crop categories reflects the strong generalization of the model in providing class-wise performance. The results showed that MB-EFF is a systematic scalable solution to farm precision ag applications with simulation-level precision and can be fused with mechanistic knowledge with data-driven ag decision support with expertise in determining crop recommendation.