A data-driven framework for identifying productivity zones and the impact of agricultural droughts in sugarcane using SPI and unsupervised learning

One significant impact of climate change in agricultural areas is the increase in frequency and potential impacts of droughts. The Standardized Precipitation Index (SPI) is an important meteorological drought index. It is commonly used as a proxy for agricultural drought. However, this use ignores the complex plant, soil, and atmosphere interactions. This work proposes a data-driven framework for identifying productivity zones and the impact of agricultural droughts based on unsupervised learning. It also contains a case study to identify sugarcane productivity zones during a severe drought event for 393 cities in São Paulo state, Brazil. Temporal and spatio-temporal versions of the k-means clustering model were implemented, considering historical productivity, SPI, and geolocation data as inputs. The resulting productivity zone labels for each city were compared with the target labels and the prediction using only the SPI. The best model configuration considered spatial and temporal features, historical productivity, and SPI values. This work presented three main contributions: (i) the proposed framework, which can be applied with other data sources, models, crops, and areas; (ii) the in-depth comparison with the traditional methodology used for identifying agricultural droughts; and (iii) an in-depth study of the use of the framework for a specific crop.