Dive into Aerial Remote Sensing Underwater Depth Estimation with Hyperspectral Imagery

Jiahao Qi, Xingyue Liu, Chen Chen, Dehui Zhu, Kangcheng Bin, Ping ZhongAAAI 2025 · DOI

Dive into Aerial Remote Sensing Underwater Depth Estimation with Hyperspectral Imagery

Introduction: The Challenge of Seeing Beneath the Surface

Estimating seafloor depth from aerial platforms—known as Underwater Depth Estimation (UDE)—is a critical task for coastal management, navigation, and environmental monitoring. However, standard RGB cameras often fail in this domain. Because water absorbs light selectively across the spectrum, RGB's three broad bands (Red, Green, Blue) frequently lose the "depth signal." This leads to a phenomenon where underwater areas with significantly different depths appear visually identical in standard imagery.

Comparison of Depth, RGB, and HSI data. While RGB images (b) show nearly identical visual appearance for different regions, HSI (c) reveals distinct spectral signatures that correspond to the actual depth differences seen in (a).

The HUDEMamba Framework

To overcome these limitations, researchers have introduced HUDEMamba, a novel framework leveraging Hyperspectral Imagery (HSI). Unlike RGB, HSI captures hundreds of narrow, contiguous bands, providing a rich "spectral fingerprint" of the water column. HUDEMamba employs a two-stage approach:

  1. Physics-inspired Image Translator (PIT): This stage uses "Hyperspectral Unmixing" to isolate the spectral signatures of water and the seafloor. By transforming raw HSI data into a simplified, depth-related representation, the model becomes more robust and generalizes better across different coastal environments without needing to "re-learn" basic optical physics.
  2. Context-aware Depth Estimator (CDE): Utilizing the state-of-the-art Mamba architecture (a State Space Model), the CDE handles long-range dependencies efficiently. A specialized Superpixel-pixel Selective Scanning strategy ensures the model maintains spatial correlations, recognizing that neighboring seafloor points typically share similar depths.

The ATR-HUDE Dataset

A major bottleneck in UDE research has been the lack of high-quality, open-access data. The authors address this by releasing the ATR-HUDE dataset, the first of its kind. Collected via a UAV-borne system in Sanya, China, the dataset includes 500 synchronized pairs of HSI data and LiDAR-derived ground truth depth maps. By timing collections with tidal cycles (LiDAR at low tide, HSI at high tide), the researchers ensured sub-centimeter alignment between imagery and depth measurements.

Key Results and Insights

HUDEMamba significantly outperforms existing RGB-based and HSI-based methods:

  • High Precision: The model achieved an RMSE of 0.037 in full-scene tests, nearly doubling the accuracy of previous state-of-the-art methods.
  • Superior Generalization: In cross-scene evaluations—where the model is tested on entirely new coastal regions—HUDEMamba maintained a significant lead, proving its effectiveness as a "task-agnostic" tool for remote sensing.
  • Spectral Resilience: The study confirms that HSI can differentiate between "dark green" patches that look identical to the human eye but exist at different depths, thanks to subtle variations in light absorption captured in the hyperspectral cube.

Future Directions

While HUDEMamba sets a new benchmark, future work will focus on expanding the ATR-HUDE dataset to include more diverse water conditions (such as turbid or inland waters) and optimizing the Mamba architecture for real-time processing on edge devices for autonomous drone navigation.

Summary updated

BibTeX
@inproceedings{qi2025dive,
  author = {Jiahao Qi and Xingyue Liu and Chen Chen and Dehui Zhu and Kangcheng Bin and Ping Zhong},
  title = {{Dive into Aerial Remote Sensing Underwater Depth Estimation with Hyperspectral Imagery}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i6.32701},
}
Figure