AnnaData: Design and Development of a Robust Multi-sensor Early Warning System for Bacterial Blight Detection in Rice Crop using Deep Learning Techniques

Abstract

As per FAO estimates, annually around one-third of food produced worldwide is lost or wasted. Plant pests, pathogens and weeds account for a large proportion of global crop production losses in the pre-harvest stages. Bacterial blight, caused by Xanthomonas oryzae, is one of the most devastating diseases in rice that has the potential to destroy up to 70% of a smallholder farmer's seasonal yield. In this paper, we describe "AnnaData" which employs a robust multi-sensor and multilevel fusion model that combines advanced computer vision techniques along with hyperspectral and thermal data processing, to recognize crop abnormalities in the incipient stages and alert farmers regarding potential onset of bacterial blight disease in rice crop. The efficacy of the AnnaData model has been validated in a lab setting by artificially inoculating the pathogen on a research farm in partnership with India's National Rice Research Institute (NRRI) in Odisha state. Compared to standard computer vision models based on visual and near-infrared image markers that delivered 40%-80% detection rates in asymptomatic stages of the disease, AnnaData's multi-sensor model achieved greater than 95% detection accuracy with less than 5% false positive rates. The AnnaData model is currently being pilot-tested on 5 farm sites in disease-endemic districts of Odisha before being productized for wider distribution among rice farmers in the state.