Efficient Attention-Based Hybrid Deep Learning Architecture for Multi-Crop Plant Disease Recognition

Early and accurate diagnosing of crops that contract diseases is critical in sustaining agricultural production and managing economic losses. Despite the massive success of the deep learning in the automated diagnosis of plant disease, new practices are largely only applicable to specific crops, and also need to be in controlled conditions and not in the field. In response to the aforementioned problems, a new Efficient Attention-based Hybrid Deep Learning (EA-HDL) has been suggested in this paper to perform the classification of multi-crop leaf diseases using real-field images. The architecture is based on an EfficientNetV2 backbone pretrained and has an attention-based pooling mechanism to encourage the use of discriminative features by the effective synthesis of information of the disease-relevant areas and the elimination of background noise. It is a tested, validated and benchmarked framework that was experimented on four of the most crucial crops: cotton, chickpea (chana), Black Gram and wheat in different field conditions. Strong and consistent results have been obtained in experiment work with a 100% record of classification accuracy in the cotton case, 98.64% in the chickpea case, 97.53% in the wheat case and competitive results in the Black Gram case in spite of difficult visual variability. It can be compared to the latest state-of-the-art deep learning models to prove that our approach is more accurate, as it generalizes and works with a variety of crops. The results are evidence that attention-based hybrid deep learning models have a tremendous potential of enhancing accuracy in disease classification in real-life agricultural 1 conditions. The EA-HDL is an effective and scalable platform to real-world crop disease surveillance and precision agriculture system.