BEGIN:VCALENDAR VERSION:2.0 PRODID:-//pretalx//talks.osgeo.org//foss4g-2026//speaker//FSNJRZ BEGIN:VTIMEZONE TZID:JST BEGIN:STANDARD DTSTART:20000101T000000 RRULE:FREQ=YEARLY;BYMONTH=1 TZNAME:JST TZOFFSETFROM:+0900 TZOFFSETTO:+0900 END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:pretalx-foss4g-2026-ZFRGQZ@talks.osgeo.org DTSTART;TZID=JST:20260903T130000 DTEND;TZID=JST:20260903T133000 DESCRIPTION:Bare-earth Digital Elevation Models (DEMs) or Digital Terrain M odels (DTMs) are fundamental to geospatial applications\, from flood model ling and landslide assessment to infrastructure planning and environmental management. However\, original publicly accessible global elevation produ cts\, such as SRTM (Shuttle Radar Topography Mission)\, ASTER GDEM (Advanc ed Spaceborne Thermal Emission and Reflection Radiometer Global Digital El evation Model)\, and Copernicus DEM\, represent Digital Surface Models (DS Ms). DSMs are elevation models that include canopy heights and building st ructures\, rather than true bare-earth topography. Their vertical accuraci es (how closely elevations match the real ground height) typically range f rom 4–15 m RMSE (root mean square error)\, and their spatial resolutions (size of the smallest discernible detail) are constrained to 30 m or coar ser. Airborne Light Detection and Ranging (LiDAR) derived DTMs achieve sub -meter vertical accuracy and spatial resolution because lasers/photons pen etrate vegetation to measure ground elevation. However\, high-cost LiDAR s urveys lead to fragmented coverage\, even in developed countries. For exam ple\, about 16% of New Zealand’s land surface still lacks airborne LiDAR mapping\, resulting in critical data gaps in the very remote\, rugged\, a nd densely vegetated terrain where elevation information is most important .\n\nRecent deep learning approaches have the potential to enhance the ver tical accuracy and spatial resolution of global DEMs through super-resolut ion techniques. For example\, JSPSR (Joint Spatial Propagation Super-Resol ution) networks improve Copernicus GLO-30 DSM from 30 m to 3 m spatial res olution by utilising high-resolution remote sensing imagery\, reducing ele vation RMSE by over 70% across diverse sites. However\, their performance drops in dense forest canopies. Optical sensors cannot penetrate vegetatio n\, so models must infer ground elevation from indirect cues (such as esti mating canopy height or shape). Spaceborne LiDAR missions\, such as ICESat -2’s ATL08 product\, provide global terrain measurements that can penetr ate vegetation. However\, incorporating these measurements poses two main challenges. First\, after rigorous data quality filters\, ATL08 photons ca n cover as little as 0.1–0.2% of dense-forest mountain areas. Second\, s paceborne LiDAR provides sparse point data\, whereas optical imagery and D EMs are dense raster grids\, resulting in a data-geometry mismatch that mo st conventional network architectures cannot inherently accommodate.\n\nTh is study describes an open-source deep neural network framework that tackl es these two challenges using a triple-branch multi-modal fusion network. It processes three complementary data streams: high-resolution remote sens ing imagery via a Swin Transformer encoder\, interpolated Copernicus GLO-3 0 DEMs via a parallel Swin encoder\, and serialised ATL08 along-track data via a sparse encoder that preserves measurement features without rasteris ation or interpolation. All code and trained models will be released under an open-source license to support open-science principles.\n\nThe main in novation is a multi-scale deformable cross-attention mechanism that enable s effective fusion of sparse LiDAR measurements with dense raster data. At each scale\, individual ATL08 photons independently query the surrounding image and DEM features through learned deformable sampling patterns\, all owing each photon to adaptively sample contextual information most relevan t to elevation prediction at its location. Inspired by deformable DETR\, t he deformable cross-attention mechanism implements bidirectional informati on flow: image and DEM features inform the interpretation of photon measur ements during downscaling\, while photon-derived elevation features are in jected back into the feature maps during upscaling. This design provides t hat sparse yet accurate LiDAR measurements guide feature extraction\, whil e dense image context enriches photon representations\, addressing the key challenge of multi-modal feature fusion in open geospatial data science.\ n\nUltimately\, the Spatial Propagation Network (SPN) transforms sparse DS M grids into dense predictions by conditioning content-adaptive kernels on fused multi-modal features. Through multiple propagation iterations\, cor rections propagate along paths guided by image content\, following ridgeli nes\, thalwegs\, or areas with similar vegetation characteristics\, while re-injecting precise photon measurements at each iteration to retain accur acy at known locations. \n\nSparse-to-dense progressive supervision comput es elevation loss only at ATL08-confirmed locations (typically fewer than 64 per 256×256 training patch)\, whereas multi-scale deep supervision hea ds propagate gradient information throughout the network\, even in areas w ithout direct elevation constraints. This strategy prevents learning inval id correlations in unobserved areas while retaining end-to-end differentia bility.\n\nWe evaluate our approach using the DFC30 dataset\, which we aug ment with spatially matching ATL08 measurements. The training set contains 12\,728 image-DSM-photon tuples\, and the test set contains 3\,196\, both with airborne LiDAR ground truth. Our results show considerable improveme nts versus the baseline method\, JSPSR. Across all test sites at 3 m spati al resolution\, elevation accuracy (RMSE) improves by 8% in open terrain ( from 1.1 m to 1.01 m) and 28% in dense forest canopies (from 6.7 m to 4.8 m). Overall\, we achieve a vertical accuracy of 3.2 m for all vegetation c lasses. The largest improvements occur where optical-only methods perform weakest. In indigenous forests with dense understory and complex terrain\, our method decreases systematic bias from 15.1 m to 8.8 m.\n\nAll code\, trained weights\, and preprocessing tools will be public under an open-sou rce license. This ensures complete repeatability and allows community adap tation. The trained model enables end-to-end 3 m DTM generation for unmapp ed areas of New Zealand using global DEMs. It will produce a seamless bare -earth elevation product at a national scale\, covering about 268\,000 squ are kilometres. This supports applications such as landslide mapping\, hyd rological analysis for freshwater management\, carbon stock assessment in forests\, and infrastructure planning in rural areas.\n\nFor the FOSS4G co mmunity\, this work makes three key contributions. First\, it delivers a p ractical\, open-source solution for generating high-resolution bare-earth DTMs by fusing global datasets (Copernicus DEM\, remote sensing imagery\, and ICESat-2 ATL08) using reproducible methods. Second\, it provides an ar chitectural strategy that respects the geometry of different data modaliti es\, e.g.\, constructed tables\, sparse point clouds and dense rasters\, b y using a template for multi-modal fusion in open geospatial science. Thir d\, it shows that open data and software can solve real-world data gaps to produce operational products that benefit communities\, environmental man agement\, and disaster resilience. More broadly\, our work supports ongoin g efforts in the open geospatial community to improve global terrain chara cterisation by fusing diverse Earth observation assets\, making high-quali ty elevation data accessible to all. DTSTAMP:20260717T225808Z LOCATION:Cosmos2 SUMMARY:Learning with Spaceborne LiDAR for Enhancement of Bare-Earth Digita l Elevation Models from Global Data - Xiandong Cai URL:https://talks.osgeo.org/foss4g-2026/talk/ZFRGQZ/ END:VEVENT END:VCALENDAR