Assessing the efficacy of machine learning techniques to characterize soybean defoliation from unmanned aerial vehicles

Abstract

Severe crop defoliation caused by insects and pests is linked to low agricultural productivity. If the root cause is not addressed, severe defoliation spreads, damaging whole crop fields. Understanding which areas are afflicted by severe defoliation can help farmers manage crops. Unmanned Aerial Vehicles (UAV) can fly over whole crop fields capturing detailed images. However, it is hard to characterize crop defoliation from aerial images that include multiple, overlapping plants with confounding effects from shadows and lighting. This paper assesses the efficacy of machine learning techniques to characterize defoliation. Given an UAV image as input, these techniques detect if severe defoliation is present. We created a labeled data set on soybean defoliation that comprises over 97,000 UAV images. We compared machine learning techniques ranging from Naive Bayes to neural networks and assessed their efficacy for (1) correctly characterizing images that contain defoliated crops and (2) avoiding wrong characterizations of healthy crops as defoliated. None of the techniques studied achieved high efficacy on both questions. However, we created DefoNet, a convolutional neural network designed for detecting crop defoliation that produces models that can be efficacious for either question. If adopted in practice, DefoNet models can guide decision making for mitigating crop yield losses due to defoliating insects.