Predicting Active Fire Occurrence in the Brazilian Cerrado Using ConvLSTM Networks, Multi-Source Environmental and Anthropogenic Data

Abstract

Wildfires in the Brazilian Cerrado combine high frequency, large spatial extent, and strong anthropogenic influence, resulting in substantial ecological, climatic, and social impacts. This paper investigates the short-term prediction of daily active fire occurrence in the Cerrado using convolutional long short-term memory (ConvLSTM) networks driven by multi-source environmental and anthropogenic data, with emphasis on satellite-based products. We build a spatiotemporal dataset on a regular grid by integrating active fire detections from the BDQueimadas system (INPE), meteorological variables from INMET and ERA5-Land, MODIS vegetation indices and land surface temperature, MapBiomas land-use and land-cover maps, SRTM topography, and distance-to-infrastructure layers. Each training sample consists of a short window of recent environmental conditions and a binary target map indicating whether at least one new active fire ignition occurs in each grid cell on the following day, thus framing ignition risk as a spatiotemporal forecasting problem. Although raw satellite and meteorological data are available from 2014 to 2024, limitations in memory and computation led us, in this initial prototype, to construct the full 5 km feature cube only for two representative years (2014 and 2017). The ConvLSTM model is trained on windows extracted from 2014 and evaluated on an independent test set from 2017. On this extremely imbalanced test set (≈ 0.03% positive cell–days), the model attains an area under the ROC curve (AUC-ROC) of about 0.89 at the pixel level, while precision–recall analysis and top-k evaluation highlight both nontrivial skill and the challenges of predicting rare fire ignitions. The study documents the data-engineering pipeline, the ConvLSTM configuration, and the evaluation protocol as a step toward operational early-warning tools for the Cerrado biome.