Abstract
Abstract
The fungal pathogen Magnaporthe oryzae is the cause of rice blast disease and is a major challenge to the world rice production as it decreases the yield and quality of rice. To manage the disease properly it is important to accurately evaluate the severity of the disease but the conventional manual method of inspection is long, subjective and inconsistent. In this paper, a deep learning-based multi-scale feature fusion model is suggested to achieve automated evaluation of the severity of rice blast disease using leaf images. The suggested model combines characteristics obtained in various convolutional layers to both fine-grained lesion characteristics and global contextual information, which allows effective severity prediction. The framework can be used to bias classification-based severity prediction (mild, moderate, severe) and regression-based estimation of infected leaf area. After Resizing, Normalization and Data Augmentation the Convolutional NN Backbone extracts the Hierarchical Feature Maps from shallow convolutions, intermediate convolutions and deep convolutions. These maps are used in finer maps to highlight fine mapping of lesion boundaries, texture characteristics, and global disease patterns. Then, these features are spatially aligned using upsampling or downsampling, and fused in multi-scale feature-fusion module to obtain a consolidated spatial representation of the size, distribution, intensity and contextual information of the lesion. The fused features are then processed by global average pooling and fully connected layers.The fused features are fed to the global average pooling and fully connected layers. The output layers are a softmax layer to classify disease severity as either mild, moderate and severe, and a regression layer which estimates the percentage of leaf area that is infected. The experimental findings indicate that the suggested strategy has a classification accuracy of about 93% and the error in prediction is low with a loss value of about 0.2, which is better than the conventional single-scale models. Moreover, the model has a good generalization performance, with an accuracy of validation of approximately 92% and AUC of almost 0.88. The findings substantiate the claim that multi-scale feature fusion is an effective way of improving the accuracy and reliability of the rice blast severity assessment. The suggested system can be adopted as a useful means of precision agriculture to identify the situation in a timely manner and take appropriate measures to reduce the losses of crops.