Comparative Assessment of High-Resolution Satellite and UAV Imagery for Land Cover Classification of the Yerevan Botanical Garden (Armenia) Using Machine Learning and Deep Learning Models

Accurate land-cover mapping of urban botanical gardens is essential for biodiversity assessment, ecosystem monitoring, and sustainable landscape management. However, their high structural complexity creates significant challenges for remote sensing-based classification. This study evaluates the influence of spatial resolution on classification performance by comparing 3 m PlanetScope satellite imagery and centimeter-scale unmanned aerial vehicle (UAV) data within a 28 ha heterogeneous area of the Yerevan Botanical Garden (YBG) in Yerevan, Armenia. Three pixel-based machine learning (ML) algorithms, Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosting (GB), were compared with the spatial-context-based U-Net deep learning (DL) semantic segmentation model. The results revealed a strong relationship between sensor resolution and classifier performance. For satellite imagery, U-Net achieved the highest statistical accuracy, with an Overall Accuracy (OA) of 78.56% and a Kappa coefficient (κ) of 0.726, demonstrating the benefit of contextual feature extraction for reducing pixel-level noise. However, spatial evaluation indicated that RF generated land-cover patterns with better preservation of local landscape structures and class boundaries. In contrast, UAV-based classification showed superior performance for pixel-based ML methods, with GB achieving the highest accuracy (OA = 85.9%, κ = 0.839), followed by SVM (OA = 84.7%, κ = 0.825) and RF (OA = 84.3%, κ = 0.821). Area comparison revealed that medium-resolution satellite imagery overestimated continuous tree canopy coverage (30.17% vs. 16.53% from UAV mapping) and underestimated fragmented vegetation classes due to mixed-pixel effects. The findings highlight that classifier selection should be adapted to sensor characteristics and landscape complexity. For satellite imagery, U-Net achieved the highest overall accuracy while RF preserved landscape structure and class boundaries more effectively; for UAV imagery, GB achieved the best overall performance, confirming that spectral-based pixel classifiers perform strongly at fine spatial resolutions. This is probably due to reduced mixed-pixel effects rather than enriched spatial information, as the latter was not leveraged in the classification process used.