Detecting dryland forests across contrasting vegetation structures: object- and pixel-based classification methods...

Dryland forests are among the most threatened forest types worldwide, making their accurate detection essential for monitoring, conservation, and land-use planning. However, their open and discontinuous structure challenges remote sensing, as individual pixels often contain mixed land cover components. This study evaluates the performance of pixel-based and object-based classification methods for detecting open xeric forests using Sentinel-2 imagery. Four sites in the Monte region (Mendoza, Argentina), representing contrasting forest structures, were analyzed using the Random Forest classification under a binary forest/non-forest scheme. The object-based method combined image segmentation with object-level spectral, SATVI, and texture descriptors, whereas the pixel-based method relied on pixel-level spectral information and the SATVI index. Results showed that the object-based method achieved higher overall accuracy and lower allocation disagreement at sites where tree cover forms spatially coherent patches, either due to greater canopy continuity or stronger contrast with surrounding land covers. In contrast, the pixel-based method performed better in disturbance-driven systems with highly heterogeneous vegetation patterns, offering lower computational complexity. These findings indicate that classification performance is strongly influenced by forest structure and spatial organization and that selecting an appropriate classification strategy depends on landscape characteristics. Our results provide practical guidance for improving dryland forest detection using medium-resolution satellite imagery, supporting large-scale environmental monitoring in dryland ecosystems.