Man-Made Objects Classification in Long-Baseline Monostatic–Bistatic SAR Images: Algorithm Training and Testing on Repeat-Pass CSG Images

Highlights: What are the main findings? A complete classification pipeline combining multi-angle CSG SAR imagery and an XGBoost model with different types of features, such as intensity, polarimetric, spatial, and textural, reliably detects man-made objects, achieving stable F1 scores (0.72–0.73) and true positive rates (0.78–0.80). Feature-importance analysis shows that multi-angle Gabor and modified Freeman–Durden descriptors provide most of the discriminative power, confirming the relevance of long-baseline geometry and tailored polarimetric–textural features for urban object detection. What are the implications of the main findings? Multi-angle monostatic–bistatic SAR configurations, such as those foreseen for PLATiNO-1, can supply an additional information layer that mitigates the underestimation of urban areas and small man-made structures in products like CGLS-LC100. The proposed workflow, from data preprocessing to imbalance handling and hyperparameter tuning, can be directly transferred to future real monostatic–bistatic acquisitions and extended to multi-class LULC mapping and other application scenarios. Land cover mapping is a crucial component of the Copernicus Land Monitoring Service, but existing products underestimate urbanized areas and small-scale man-made objects, limiting their ability to capture the complexity of built environments. Long-baseline monostatic–bistatic Synthetic Aperture Radar (SAR) images, such as the ones that will be made available by the upcoming PLATiNO-1 mission, have the potential to contribute to the detection of the mentioned targets, e.g., by traditional supervised classification approaches. Since bistatic measurements from the PLATiNO-1 mission are not yet available, repeat-pass COSMO-SkyMed second generation (CSG) images collected with different incidence angles are employed to emulate the expected diversity of future monostatic–bistatic products. A complete classification pipeline is developed, and a structured dataset of 48 features is built, combining intensity, polarimetric, spatial, and textural descriptors to train an XGBoost model to identify urban targets within a representative area in Italy. The results demonstrate stable performance, with F1 scores around 0.73 and true positive rates close to 80%, showing good agreement with reference data and confirming the feasibility of the proposed methodology. Although conceived as a proof of concept, the study shows that integrating multi-angle information into classification tasks can improve the detection of man-made structures and provide an additional information layer to be integrated with Copernicus services.