Research map / 3D vision / Depth and stereo
Dive into Aerial Remote Sensing Underwater Depth Estimation with Hyperspectral Imagery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds ATR-HUDE, a ~500-pair UAV hyperspectral-LiDAR underwater depth dataset, and proposes HUDEMamba, combining a parameter-free physics-inspired image translator with a Mamba-based context-aware depth estimator using superpixel-pixel selective scanning.
Problem
Underwater depth estimation from aerial remote sensing is poorly served by visible-spectrum images (three bands, water absorption), and hyperspectral underwater depth estimation (HUDE) lacks public datasets and methods that exploit spectral information and spatial correlation.
Approach
Translate the hyperspectral image into a depth-related image using a parameter-free physics-based unmixing step (endmember with max abundance as water body, abundance giving attenuation), then estimate depth with a Mamba (VMamba-T) encoder using superpixel-guided coarse-to-fine scanning and adaptive bin prediction.
Key results
- On full-scene testing, HUDEMamba achieves SIlog 0.201, A.Rel 0.020, RMSE 0.037 versus the best baseline SUTDF at 0.324, 0.039, 0.065.
- On cross-scene testing, HUDEMamba reaches average SIlog 2.439 and RMSElog 0.321, improving SIlog by 0.608 over the best competing method (SUTDF, 3.047).
- Removing the physics-inspired image translator raises full-scene RMSElog from 0.025 to 0.063 and drops delta1 from 1.000 to 0.984.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the physics-inspired image translator (PIT) (ablation) | HUDEMamba · ATR-HUDE full-scene testing | RMSElog: 0.063 -> 0.025 | Helped |
| add the physics-inspired image translator (PIT) (ablation) | HUDEMamba · ATR-HUDE full-scene testing | δ1: 0.984 -> 1.000 | Helped |
| add the physics-inspired image translator (PIT) (ablation) | HUDEMamba · ATR-HUDE cross-scene testing | RMSElog: 0.394 -> 0.321 | Helped |
| add the physics-inspired image translator (PIT) (ablation) | HUDEMamba · ATR-HUDE cross-scene testing | δ1: 0.282 -> 0.400 | Helped |
| use the context-aware depth estimator instead of a Transformer estimator (ablation) | HUDEMamba with PIT · ATR-HUDE full-scene testing | RMSElog: 0.042 -> 0.025 | Helped |
| use the context-aware depth estimator instead of a Transformer estimator (ablation) | HUDEMamba with PIT · ATR-HUDE cross-scene testing | δ1: 0.342 -> 0.400 | Helped |
| use the context-aware depth estimator instead of vanilla Mamba (ablation) | HUDEMamba with PIT · ATR-HUDE full-scene testing | RMSElog: 0.034 -> 0.025 | Helped |
| use the context-aware depth estimator instead of vanilla Mamba (ablation) | HUDEMamba with PIT · ATR-HUDE cross-scene testing | δ1: 0.367 -> 0.400 | Helped |
| use HUDEMamba instead of SUTDF | ATR-HUDE full-scene testing · ATR-HUDE full-scene testing | RMSElog: 0.045 -> 0.025 | Helped |
| use HUDEMamba instead of SUTDF | ATR-HUDE cross-scene testing · ATR-HUDE cross-scene average | SIlog: +0.608 in SIlog | Helped |
Limitations
- Dataset is small (500 pairs from three coastal scenes in Sanya, China).
- Cross-scene performance remains far from full-scene performance (SIlog about 2.4 vs 0.2).
Experimental setup
Datasets: ATR-HUDE
Benchmarks: ATR-HUDE
Metrics: SIlog RMSE RMSElog A.Rel S.Rel δ1 δ2 δ3
Baselines: UWNet UDepth Joint-ID WaterMono SUTDF PIT+Transformer PIT+vanilla Mamba
Builds on: Mamba VMamba AdaBins physical image formation model SLIC superpixels
Closest papers by meaning
- Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion (CVPR 2024 · similarity 0.83)
- HIR-Diff: Unsupervised Hyperspectral Image Restoration Via Improved Diffusion Models (CVPR 2024 · similarity 0.81)
- Automatic Spectral Calibration of Hyperspectral Images: Method, Dataset and Benchmark (CVPR 2025 · similarity 0.81)
- Sp3ctralMamba: Physics-Driven Joint State Space Model for Hyperspectral Image Reconstruction (AAAI 2025 · similarity 0.81)
- Unmixing Before Fusion: A Generalized Paradigm for Multi-Source-based Hyperspectral Image Synthesis (CVPR 2024 · similarity 0.81)
- Adaptive Dual-domain Learning for Underwater Image Enhancement (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.