Spatiotemporal variation of deforestation in the Piauí Caatinga under a panel and machine learning approach

Abstract

The Piauí Caatinga has shown a marked rise in deforested area in recent years, but the factors associated with this dynamic remain poorly investigated under a spatiotemporal approach at the municipal scale. This study identified the factors associated with the spatiotemporal variation of deforestation in the biome across 140 municipalities of Piauí state between 2017 and 2024, integrating fixed-effects panel regression (FE), a spatial autoregressive model (SAR), and Random Forest (RF), based on deforestation data (PRODES/INPE), land use (MapBiomas), socioeconomic indicators (IBGE), rural credit (SICOR/BACEN), and protected areas (CNUC/MMA). The time series revealed a 105.3% increase in mean annual deforested area from 2021 onward, without statistically significant confirmation of either a trend (Mann-Kendall) or a change point (Pettitt) at the 5% level. Moran's index confirmed significant spatial dependence, intensified in the post-2021 period. The FE model identified agricultural land-use mosaics as the strongest correlate of intramunicipal temporal variation (β = 1.032; p < 0.001), a result potentially confounded by overlap with the PRODES response variable, whereas the SAR revealed significant spatial dependence (Rho = 0.294; p < 0.001) and associations of productive-pressure and infrastructure variables not captured by the FE. Random Forest, under municipality-grouped cross-validation, explained 50.98% of the variance in deforestation, identifying pasture, cattle herd size, distance to hydrography, and protected area coverage as additional structural predictors. The results evidence methodological complementarity among the three approaches and point to land-use conversion as the main correlate of deforestation in the biome, supporting the territorial prioritization of environmental command-and-control actions.