Detection of Cropland Abandonment through Multi-Temporal Landsat Data and Spatially Independent Machine Learning Validation

Abstract. Cropland abandonment (CA) is a major land-use change with important environmental and socio-economic implications. This study evaluates cropland abandonment using multi-temporal Landsat features and a spatially independent validation framework, comparing the performance and spatial behaviour of Random Forest and XGBoost classifiers. A set of temporally aggregated spectral indices (NDVI, BSI, NDBI, and MNDWI), including multi-year trends and variability measures, was integrated into a 56-band composite dataset. Training and validation samples were generated using 100 × 100-pixel windows centred on land-use parcels, with overlapping areas between different reference classes explicitly excluded to avoid label ambiguity. To reduce spatial autocorrelation, the data were split into separated training (1,582.6 km²) and testing (719.2 km²) areas within the Savinjska statistical region in Slovenia. Random Forest (RF) and XGBoost (XGB) classifiers were trained and evaluated using spatially separated validation data. Classification performance was assessed using standard accuracy metrics. Results indicate that XGB achieved a higher overall accuracy (0.705) compared to RF (0.670) and exhibited strong sensitivity in detecting cropland abandonment, while RF produced more conservative and spatially stable estimates of abandoned cropland area. Spatial error maps and area-based comparisons reveal systematic differences between the two classifiers, particularly in their tendency to overestimate abandonment extent. The findings highlight the importance of spatially explicit validation strategies, careful reference data preparation, and multi-temporal feature design for robust cropland abandonment mapping. The main contribution lies in the systematic assessment of model behaviour, spatial error patterns, and area estimates under strict spatial separation of training and testing data.