Multi-Stage Neural Network based Ensemble Learning Approach for Wheat Leaf

Abstract

Wheat is a major food commodity in the world, but it faces serious disease attacks, which have caused huge losses every year. The review summarizes the recent progress in deep learning (DL), especially, multi-stage ensemble and lightweight neural networks, to detect wheat leaf disease automatically. The models such as DenseNet and ResNet are highly accurate but are prone to overfitting, inefficiency in computation and generalization to real-field conditions. We discuss how ensemble learning techniques and architectural advances e.g. feature fusion, attention mechanism and knowledge distillation can be used to make them more robust, better at feature extraction and ameliorate environmental factors (complex backgrounds, occlusion etc.). The above analysis also establishes that multi-stage models are far more effective than single models, however, there is an issue of critical trade-off between accuracy and computational cost. Future efforts should focus on creating light interpretable models based on multisource data to be used on large scale and in real-time in precision agriculture.