Machine learning for land use change analysis in environmental protection areas

Anthropogenic land use and land cover change (LULCC), combined with ongoing climate variability, poses significant challenges to environmental protection areas (EPAs) by altering ecosystem structure, degrading vegetation integrity, and disrupting local climate regulation. Despite their importance, traditional LULCC approaches often fail to incorporate dynamic environmental drivers, limiting their capacity to represent complex landscape-climate interactions. This study investigates the environmental dynamics of the Guaraqueçaba Environmental Protection Area and evaluates the landscape's potential for automated classification and prediction of impacts associated with land use change. A multitemporal dataset spanning 15 years (2009-2023) was analyzed, comprising an original set of approximately 30.2 million records of annual time series of precipitation, maximum and minimum temperatures, evapotranspiration, global solar radiation, relative humidity, wind speed, land use and land cover information, and the Normalized Difference Vegetation Index (NDVI). To address class imbalance, a balanced subset of 3.6 million records was used for modeling. Predictive models were developed using multiple linear regression (MLR), k-nearest neighbors (KNN), and random forest (RF), with performance assessed under both imbalanced and balanced data conditions using accuracy, $R^2$ , precision, recall, and F1-score metrics. The results indicate pronounced local climate changes, including increasing temperatures in anthropogenically modified areas and altered humidity patterns associated with vegetation loss. Among the evaluated models, RF exhibited the highest predictive performance, achieving accuracies of up to 96% and an R $_{}^2$ of 88.6%, effectively capturing the nonlinear interactions between LULCC, climate variables, and vegetation dynamics. Precipitation and NDVI emerged as the most influential drivers of LULCC processes. These findings demonstrate the effectiveness of machine learning approaches for identifying environmental degradation trajectories in protected areas and provide a robust framework to support targeted mitigation strategies and policy development applicable to other EPAs facing increasing anthropogenic pressure and climate variability.