Researchers have released the first assessment of PhiSat-2 imagery for monocular building-height estimation, a task that uses a single satellite image to infer how tall structures are. The study introduces the PhiSat-2–Height Dataset (PHDataset), which combines global PhiSat-2 imagery with open building-height references to create 9,475 co-registered patch pairs from 26 cities.

To make use of this data, the authors propose the Two-Stream Ordinal Network (TSONet). Its multiscale decoupled height representation separates adaptive bin generation from spatial assignment, while footprint-conditioned height reconstruction adds building-footprint guidance at both feature and reconstruction levels.

In city-disjoint tests, TSONet reduced building root-mean-square error by 3.64% and mean absolute error by 4.49% relative to the most accurate comparison method originally designed for building-height estimation, while keeping similar footprint overlap. PhiSat-2 itself reduced building RMSE by 3.73% and 6.05% compared with evaluated Sentinel-2 and Landsat 8 configurations. Using all seven bands produced the lowest errors among tested band subsets. LiDAR-based testing in a held-out city supported the dataset's supervision approach, and selected patch comparisons showed finer urban height patterns than coarser global products.

The results indicate that PhiSat-2's spatial and spectral information is useful for urban building-height mapping. The authors position PHDataset and TSONet as a foundation for further development in this area.