Machine Learning Approaches for Temporal Classification of Forest and Shrub Cover in Burned Landscapes in Mediterranean Ecosystems

Abstract Wildfires have altered Mediterranean fire regimes, reshaping vegetation dynamics and challenging ecosystem resilience. This study presents a reproducible Python-based framework for classifying forest vegetation and post-fire shrub states in Mediterranean ecosystems. The workflow also enables the assessment of label-dataset composition, predictor-variable sets, and classifier choice. The proposed workflow is demonstrated through a case study in Aragón, Spain, using harmonized Landsat OLI and Sentinel‑2 MSI multispectral imagery combined with topographic and edaphic variables. Three datasets of training labels were constructed from national inventories, automatic extraction and manual labeling, and five predictor sets were tested using both original spectral variables and principal components. Random Forest (RF) and Support Vector Machines (SVM) classifiers were implemented under multiple resampling and preprocessing pipelines, with performance evaluated through balanced accuracy, Cohen’s kappa, user’s and producer’s accuracy, Shapley Additive Explanations (SHAP), and Moran’s I spatial autocorrelation. SVM consistently outperformed RF, particularly when predictors were selected using an automatic approach that reduced multicollinearity. Dataset composition strongly influenced results: the dataset based on manual labeling achieved the highest accuracies, with most classes exceeding 0.85 in both user’s and producer’s metrics, while automatically expanded datasets introduced imbalance and noise, leading to sharp performance declines. Minority and spectrally similar classes such as Quercus ilex, Pinus pinaster , and Pinus nigra were most frequently misclassified. High Moran’s I values (higher than 0.70) revealed strong spatial clustering, underscoring the risk of inflated accuracy estimates in spatially autocorrelated datasets.