Detecting coffee leaf rust with UAV-based vegetation indices and decision tree machine learning models

Abstract

Coffee leaf rust (CLR) is one of the most devastating leaf diseases in coffee plantations. By knowing the symptoms, severity, and spatial distribution of CLR, farmers can improve disease management procedures and reduce losses associated with it. Recently, Unmanned Aerial Vehicles (UAVs)-based images, in conjunction with machine learning (ML) techniques, helped solve multiple agriculture-related problems. In this sense, vegetation indices processed with ML algorithms are a promising strategy. It is still a challenge to map severity levels of CLR using remote sensing data and an ML approach. Here we propose a framework to detect CLR severity with only vegetation indices extracted from UAV imagery. For that, we based our approach on decision tree models, as they demonstrated important results in related works. We evaluated a coffee field with different infestation classes of CLR: class 1 (from 2% to 5% rust); class 2 (from 5% to 10% rust); class 3 (from 10% to 20% rust), and; class 4 (from 20% to 40% rust). We acquired data with a Sequoia camera, producing images with a spatial resolution of 10.6 cm, in four spectral bands: green (530–570 nm), red (640–680 nm), red-edge (730–740 nm), and near-infrared (770–810 nm). A total of 63 vegetation indices was extracted from the images, and the following learners were evaluated in a cross-validation method with 10 folders: Logistic Model Tree (LMT); J48; ExtraTree; REPTree; Functional Trees (FT); Random Tree (RT), and; Random Forest (RF). The results indicated that the LMT method contributed the most to the accurate prediction of early and several infestation classes. For these classes, LMT returned F-measure values of 0.915 and 0.875, thus being a good indicator of early CLR (2 to 5% of rust) and later stages of CLR (20 to 40% of rust). We demonstrated a valid approach to model rust in coffee plants using only vegetation indices and ML algorithms, specifically for the disease's early and later stages. We concluded that the proposed framework allows inferring the predicted classes in remaining plants within the sampled area, thus helping the identification of potential CLR in non-sampled plants. We corroborate that the decision tree-based model may assist in precision agriculture practices, including mapping rust in coffee plantations, providing both an efficient non-invasive and spatially continuous monitoring of the disease.