The First Assessment of PhiSat-2 Imagery for Monocular Building Height Estimation

The gist The study presents an initial open-reference assessment of PhiSat-2 imagery for monocular building height estimation by constructing a PhiSat-2–Height Dataset (PHDataset) and proposing a Two-Stream Ordinal Network (TSONet) to jointly learn dense height estimation and auxiliary footprint prediction

Data Construction and Methodology

The researchers constructed the PHDataset by integrating global PhiSat-2 imagery with open building-height references, resulting in 9,475 co-registered patch pairs from 26 cities worldwide. The TSONet framework is a two-stream model that jointly models auxiliary footprint prediction and dense building height estimation. This framework employs a shared encoder, a footprint stream, and a height stream.

Key Network Components

The TSONet architecture comprises three main components:

1. A shared encoder that adopts a five-level hierarchical design analogous to U-Net to extract multi-level feature representations.

2. A footprint stream which employs a lightweight head with a 3 × 3 convolution to produce footprint logits.

3. A height stream enhanced by a Feature-Enhanced Bin Refinement (FEBR) module, which follows a classification–regression paradigm to explicitly encode the ordinal structure of building heights.

Cross-Task Interaction and Ordinal Modeling

The Cross-Stream Exchange Module (CSEM) is introduced to exchange and refine task-related representations between the two streams. CSEM uses a spatial confidence mask to modulate the exchanged features, ensuring that the footprint branch is modulated by the learned exchange strength while the height branch is jointly constrained by both a gating mask and a spatial confidence mask. The FEBR module refines bin queries through a Coarse-to-Fine Query Readout (CFQR) process, which progressively refines their feature-conditioned representations by aligning them with the hierarchical feature pyramid.

Performance and Assessment

Experiments show that TSONet outperforms representative competing methods, reducing MAE and RMSE by over 13.2% and 9.7%, respectively, while improving IoU and F1-score by over 14.0% and 10.1%. The component-wise ablation study indicates that CSEM mainly improves the footprint-related metrics, whereas FEBR contributes more directly to continuous height estimation accuracy. Furthermore, sensor-information experiments show that both the native 4.75 m spatial detail and the seven-band multispectral observations contribute to height estimation performance.

Limitations and Future Directions

The study concludes that PhiSat-2 imagery contains useful spatial–spectral cues for monocular BHE at intermediate spatial resolution, but the reported results are best understood as an initial empirical assessment rather than an operational global height product. Key factors shaping performance include the long-tailed distribution of building heights and residual geometric inconsistencies in current PhiSat-2 products. Future work is suggested to expand PHDataset with more geographically diverse high-quality PhiSat-2 scenes and further improve preprocessing pipelines, particularly band co-registration, geometric correction, and quality screening.

Independent Validation

A LiDAR-based accuracy audit showed that TSONet remains consistent with the reference labels in the selected cities, achieving an MAE of 3.44 m and an F1-score of 0.621. However, the prediction error increases to an MAE of 5.17 m when evaluated against the independent LiDAR-derived nDSM. This suggests that the prediction–LiDAR difference is associated with uncertainty in open reference labels.

Contextual Relevance

Patch-level qualitative comparison shows that TSONet produces height maps that are visually more consistent with the reference labels, better preserving elongated, regularly arranged, and densely distributed buildings than coarser global products. This suggests that task-oriented learning from PhiSat-2 imagery can provide complementary local height information relative to coarser global products. The results place PhiSat-2-based BHE in the context of existing global height products and indicate its potential relevance for relatively detailed urban vertical structure mapping.

Conclusion

TSONet combines footprint-aware structural guidance with ordinal height modeling, enabling more effective exploitation of PhiSat-2 spatial–spectral information for BHE. The component-wise ablation study shows that CSEM mainly improves structural consistency and height-map-derived building-region performance, whereas FEBR contributes more directly to height estimation accuracy. The task-configuration ablation further indicates the advantage of combining ordinal height refinement with footprint-aware spatial guidance. Overall, the results indicate that PhiSat-2 imagery contains useful spatial–spectral cues for monocular BHE and can provide complementary height information with finer spatial detail than coarser global products. The assessment also reveals key factors shaping the achievable performance, including the long-tailed distribution of building heights, the uncertainty of open reference labels, and residual geometric inconsistencies in current PhiSat-2 products.

Final Summary

TSONet combines footprint-aware structural guidance with ordinal height modeling, enabling more effective exploitation of PhiSat-2 spatial–spectral information for BHE. The component-wise ablation study shows that CSEM mainly improves structural consistency and height-map-derived building-region performance, whereas FEBR contributes more directly to height estimation accuracy; their joint use provides the most balanced performance across continuous height regression and building-region consistency metrics. The task-configuration ablation further indicates the advantage of combining ordinal height refinement with footprint-aware spatial guidance. In addition, the sensor-information experiments show that both the native 4.75 m spatial detail and the seven-band multispectral observations contribute to height estimation performance. The LiDAR-based accuracy audit and contextual comparison with existing global height products further support the physical relevance and practical value of PhiSat-2-based predictions. The overall results indicate that PhiSat-2 imagery contains useful spatial–spectral cues for monocular BHE and can provide complementary height information with finer spatial detail than coarser global products.

1. Bold header: Task-oriented joint modeling of footprint segmentation and height estimation. This system can jointly learn dense building height estimation and auxiliary footprint prediction, allowing it to leverage footprint-aware structural guidance to better exploit spatial–spectral information from PhiSat-2 imagery, as proposed by the TSONet framework.

2. Bold header: Ordinal structure modeling for long-tailed height distributions. The incorporation of the Feature-Enhanced Bin Refinement (FEBR) module allows the system to explicitly encode the ordinal structure of building heights, which is crucial because high-rise buildings are intrinsically more demanding to estimate than low-rise buildings in PHDataset.

3. Bold header: Cross-task feature interaction for structural guidance. The Cross-Stream Exchange Module (CSEM) enables adaptive interaction between the height and footprint streams, specifically modulating the footprint branch via a gating mask, which is designed to enhance building-region structure and boundary consistency.

4. Bold header: Robustness against geometric inconsistencies. By including a controlled inter-band shift experiment, the system can be trained to maintain performance even when input imagery exhibits inter-band misregistration, suggesting that improved band co-registration and automated geometric correction will enhance reliability.

5. Bold header: Performance adaptation to spatial resolution degradation. The model's ability to handle resolution changes, as shown by the results in Fig. 8(b), demonstrates that the native 4.75 m detail provides useful structural cues, allowing the system to maintain reasonable performance even when input becomes coarser, though performance still degrades predictably with increased coarseness.

6. Bold header: Contextual comparison for practical relevance assessment. The system can perform a patch-level qualitative comparison against several publicly available global building height products, identifying that TSONet's predictions show visually coherent building layouts and local height contrasts that are more consistent with the PhiSat-2 image context and the reference labels.

PhiSat-2 provides multispectral observations at a 4.75m spatial resolution across geographically diverse scenes, offering a new source of remote sensing imagery for monocular building-height estimation. To comprehensively assess its utility, this study introduces the PhiSat-2--Height Dataset (PHDataset) and the Two-Stream Ordinal Network (TSONet). PHDataset integrates global PhiSat-2 imagery with open building-height references and contains 9475 co-registered patch pairs from 26 cities worldwide. In TSONet, multiscale decoupled height representation (MDHR) separates adaptive bin generation from spatial assignment, while footprint-conditioned height reconstruction (FCHR) incorporates footprint guidance at feature and reconstruction levels. Under city-disjoint evaluation, TSONet reduces building RMSE and MAE by 3.64% and 4.49%, respectively, relative to the most accurate comparison method originally designed for building-height estimation, while maintaining similar height-derived footprint IoU. Ablations support the combined design for height accuracy and footprint consistency. PhiSat-2 reduces building RMSE by 3.73% and 6.05% relative to the evaluated Sentinel-2 and Landsat 8 configurations, respectively, with all seven bands providing the lowest building-height errors among its tested band subsets. LiDAR-based testing supports PHDataset-style supervision in a held-out city, while selected patch comparisons show finer urban height patterns than coarser global products. These findings highlight the value of PhiSat-2 spatial-spectral information for urban building height mapping, with PHDataset and TSONet providing a foundation for further development.