Multilayer optimized deep learning model to analyze spectral indices for predicting the condition of rice blast disease

Abstract

Rice blast disease is one of the most destructive infectious diseases that affects world food security. Proper monitoring and an accurate decision-making process can assist in disease management strategy. Ground surveys and sampling are the less accurate, expensive, and time-consuming processes that are ineffective to check epidemic. Satellite data-driven approach might be an ideal cost and time-efficient technique that can provide an accurate result due to its revisit across farmland. Temperature variation is a salient feature of this disease trajectory. Hence, land surface temperature can be a cardinal property for disease risk estimation. Spectral indices-based analysis can be more efficient for tracking the disease density. In this study, the MODIS satellite-based Land Surface Temperature (LST) parameter is used to indicate the disease in the field. The indicated risk estimation is also examined using ground truth observation to provide less erroneous labeling. Various spectral combination based remote sensing indices were accumulated to audit the disease states. Remote sensing indices such as Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), Enhanced Vegetation Index (EVI), Normalized Difference Moisture Index (NDMI), and Moisture Stress were obtained from the Sentinel-2 archive. These images, depicting the various indices, are processed through a novel optimized deep learning model to predict the disease condition of farmland. The model is developed using various residual networks with $L2$ regularization and batch normalization to enhance the performance of the model. A combination of convolution layers is used to extract crucial spectral information from the remote sensing images and processed through fully connected layers to prognosticate the state of the disease. The model can predict with 89.67% accuracy using the EVI parameters for different geographical positions compared with other remote sensing parameters and has less chance of erroneous possibilities. The proposed system will lead to improved agricultural monitoring management for the incidence of leaf blast disease in real-time.