Addressing coffee crop diseases: forecasting Phoma leaf spot with machine learning

Coffee production is significantly impacted by various diseases, predominantly those caused by fungi. One such notable disease in coffee crops is caused by the fungus Phoma spp. This pathogen leads to several symptoms detrimental to coffee plants, such as leaf lesions, drying of branches, and rotting of flowers and fruits. These symptoms often result in the dropping of the affected parts, subsequently leading to a decrease in the overall yield of the coffee crop. In response to this challenge, our objective was to develop a forecasting model for the incidence of Phoma leaf spot in Brazilian coffee crops, utilizing advanced machine learning algorithms. This approach is intended to predict disease outbreaks, thereby enabling timely and effective management strategies to mitigate the impact on coffee yield. The study was conducted in two stages: (1) calibration of machine learning models for locations (Boa Esperança, Carmo de Minas, Muzambinho, Varginha, Araxá, Araguari, and Patrocínio) with field data between 2010 and 2022; (2) Phoma leaf spot incidence forecast in municipalities of coffee-producing states in Brazil [Paraná (PR), São Paulo (SP), Rio de Janeiro (RJ), Espírito Santo (ES), Minas Gerais (MG), Goiás (GO), and Bahia (BA)]. Thirty-year climate data were retrieved from the NASA/POWER platform. Reference evapotranspiration was estimated by the Penman–Monteith method, generating water balance according to Thornthwaite and Mather (1955). To understand the effect of climate variables on the disease incidence, Pearson’s univariate correlation was performed for each location. We used six algorithms to forecast the disease incidence, considering a 7-day latency period to define input variables. It is noteworthy that the evaluated locations present similar climatic conditions. Summer was the hottest and rainiest period, while winter was the coldest and driest. Annual averages of air temperature, cumulative rainfall, potential evapotranspiration, soil water storage, and incident radiation were 21.1 °C, 1208.9 mm, 1283.2 mm, 58.0 mm, 435.7 mm, and 18.1 MJ m2 day−1, respectively. The XGBoost model demonstrated superior performance for both high- and low-yield coffee trees, achieving an impressive precision (R2fit) of 0.46 and 0.51, respectively. Additionally, it exhibited high accuracy, with Root Mean Square Error (RMSE) values of 3.45% for high-yielding and 3.16% for low-yielding trees. In contrast, the multilayer perceptron (MLP) model displayed suboptimal results under both yield conditions. Given these findings, the XGBoost model proves effective in predicting the incidence of the disease at least 7 days ahead, based on the parameters applied in this study.