Joint Hail Detection from Satellite and Radar Observations with Spatially Adaptive Alignment and Wavelet-Gated Refinement.
Saved in:
| Title: | Joint Hail Detection from Satellite and Radar Observations with Spatially Adaptive Alignment and Wavelet-Gated Refinement. |
|---|---|
| Authors: | Wang, Jiamin1 (AUTHOR), Wang, Haijiang1,2 (AUTHOR) whj@cuit.edu.cn, Li, Jieyi2,3 (AUTHOR), Liu, Tao1,4 (AUTHOR), Gu, Taofeng1,2,3 (AUTHOR), Xue, Yunheng2,4 (AUTHOR) |
| Source: | Remote Sensing. Jun2026, Vol. 18 Issue 11, p1743. 29p. |
| Subjects: | Multisensor data fusion, Hailstorms, Thunderstorms, Deep learning, Remote-sensing images, Radar antennas |
| Abstract: | Highlights: What are the main findings? The proposed framework improves joint hail detection from satellite and radar observations by learning from data how to align the two sources differently at each location, which reduces the offset between geostationary cloud-top signals and radar precipitation cores caused by parallax and storm tilt. Additional training constraints further make the fusion more stable and reliable. The results indicate that the improved hail detection comes from the complementary effects of the three modules, with scale-aware enhancement reinforcing hail signatures in each source, bidirectional deformable alignment enabling effective satellite–radar fusion, and wavelet-gated refinement preserving sharper hail boundaries. What are the implications of the main findings? For joint satellite–radar hail detection, the real bottleneck is the spatial mismatch from parallax and storm tilt, not the fusion step itself. This mismatch can be corrected through data-driven network and loss design, without analytically solving the geometry or storm dynamics. For radar–satellite fusion, network design should make better use of local information. The relationship between the two signals is better captured by spatially varying local adjustments than by a single global shift applied across the whole scene. This also helps keep detection stable as storms evolve or when one source becomes less reliable. Detecting hail from remote sensing observations remains challenging because hail develops rapidly and its signatures may appear at different levels within a storm. Ground-based radar and geostationary meteorological satellites are the two primary observing systems for this task, yet their observations are often spatially misaligned. Satellite measurements mainly characterize the thermal structure near the cloud top, whereas radar observations capture the lower-level precipitation core. This mismatch is further exacerbated by satellite parallax, namely the apparent horizontal shift of high cloud tops caused by the oblique viewing geometry of a geostationary satellite, together with the vertical tilt of convective storms. Existing joint methods generally combine satellite cloud-top information with radar precipitation information directly, without explicitly correcting the spatial displacement, which limits detection accuracy. To address this issue, we propose HailDeformer, a deep learning framework that first aligns satellite and radar features through a bidirectional deformable cross-attention module equipped with a position-wise confidence gate and optimized with smoothness, contrastive alignment, and observation-structure consistency losses, and then refines the fused representation using an inter-scale attention module and a wavelet-guided refinement module. Experiments on a four-region dataset from China show that HailDeformer consistently outperforms Direct Fusion, Manual Weighting, Cross-Attention Fusion, and Optical Flow Alignment, achieving a mean Average Precision at IoU 0.5 (mAP@0.5) of 0.916, an F1 score of 0.864, a Critical Success Index (CSI) of 0.760, and the lowest False Alarm Ratio (FAR) of 0.149. Ablation studies further confirm that all proposed modules and associated constraints contribute to the overall performance, with the alignment module providing the largest improvement. Additional evaluations demonstrate that HailDeformer remains effective throughout storm evolution and under challenging observational conditions. [ABSTRACT FROM AUTHOR] |
| Copyright of Remote Sensing is the property of MDPI and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | Highlights: What are the main findings? The proposed framework improves joint hail detection from satellite and radar observations by learning from data how to align the two sources differently at each location, which reduces the offset between geostationary cloud-top signals and radar precipitation cores caused by parallax and storm tilt. Additional training constraints further make the fusion more stable and reliable. The results indicate that the improved hail detection comes from the complementary effects of the three modules, with scale-aware enhancement reinforcing hail signatures in each source, bidirectional deformable alignment enabling effective satellite–radar fusion, and wavelet-gated refinement preserving sharper hail boundaries. What are the implications of the main findings? For joint satellite–radar hail detection, the real bottleneck is the spatial mismatch from parallax and storm tilt, not the fusion step itself. This mismatch can be corrected through data-driven network and loss design, without analytically solving the geometry or storm dynamics. For radar–satellite fusion, network design should make better use of local information. The relationship between the two signals is better captured by spatially varying local adjustments than by a single global shift applied across the whole scene. This also helps keep detection stable as storms evolve or when one source becomes less reliable. Detecting hail from remote sensing observations remains challenging because hail develops rapidly and its signatures may appear at different levels within a storm. Ground-based radar and geostationary meteorological satellites are the two primary observing systems for this task, yet their observations are often spatially misaligned. Satellite measurements mainly characterize the thermal structure near the cloud top, whereas radar observations capture the lower-level precipitation core. This mismatch is further exacerbated by satellite parallax, namely the apparent horizontal shift of high cloud tops caused by the oblique viewing geometry of a geostationary satellite, together with the vertical tilt of convective storms. Existing joint methods generally combine satellite cloud-top information with radar precipitation information directly, without explicitly correcting the spatial displacement, which limits detection accuracy. To address this issue, we propose HailDeformer, a deep learning framework that first aligns satellite and radar features through a bidirectional deformable cross-attention module equipped with a position-wise confidence gate and optimized with smoothness, contrastive alignment, and observation-structure consistency losses, and then refines the fused representation using an inter-scale attention module and a wavelet-guided refinement module. Experiments on a four-region dataset from China show that HailDeformer consistently outperforms Direct Fusion, Manual Weighting, Cross-Attention Fusion, and Optical Flow Alignment, achieving a mean Average Precision at IoU 0.5 (mAP@0.5) of 0.916, an F1 score of 0.864, a Critical Success Index (CSI) of 0.760, and the lowest False Alarm Ratio (FAR) of 0.149. Ablation studies further confirm that all proposed modules and associated constraints contribute to the overall performance, with the alignment module providing the largest improvement. Additional evaluations demonstrate that HailDeformer remains effective throughout storm evolution and under challenging observational conditions. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 20724292 |
| DOI: | 10.3390/rs18111743 |
