We present a systematic review of hyperspectral underwater target detection using China UAV drone platforms, focusing on nearshore environments with complex water conditions. Our work is organized along three main threads: imaging mechanisms, characteristic modeling, and algorithm design. The rapid advancement of unmanned aerial vehicle (UAV) technology combined with hyperspectral imaging has opened new possibilities for high-resolution, low-latency monitoring of shallow water ecosystems. However, the inherent complexities of underwater light propagation—absorption, scattering, and bottom reflectance coupling—pose significant challenges to accurate target discrimination. In this paper, we first establish the physical foundation of underwater hyperspectral imaging, then categorize existing methods into five paradigms: spectral prediction, spectral restoration, band selection, pixel classification, and feature construction. We compare these approaches in terms of mechanism-consistent modeling, distortion correction, representational robustness, and interpretability. Through quantitative evaluation on a benchmark dataset (ATR2-HUTD), we demonstrate that feature construction methods achieve the best overall performance across lakes, rivers, and coastal waters, while band selection methods exhibit the weakest generalization. We further discuss current challenges in environmental adaptability, reliability modeling, and cross-scene generalization, and propose future directions including differentiable physical modeling, uncertainty characterization, and domain-invariant representation learning.
The use of China UAV drones for hyperspectral remote sensing has gained momentum due to their on-demand deployment capability, centimeter-level spatial resolution, and ability to operate in dynamic nearshore zones. Unlike satellite platforms, China UAV drones can revisit the same area at short intervals, making them ideal for monitoring fast-changing underwater targets such as submerged structures, marine debris, or benthic habitats. Despite these advantages, the combination of low-altitude flight, platform vibration, and variable illumination introduces unique challenges: geometric distortion, spectral misregistration, and rapid radiometric fluctuations due to sun glint and cloud shadows. These issues necessitate dedicated processing pipelines that integrate physics-based corrections with data-driven learning. Our review aims to bridge the gap between physical modeling and modern machine learning, providing a unified framework for understanding the evolution of underwater target detection from early analytical models to contemporary feature learning paradigms.
1. Fundamental Models and Datasets
1.1 Underwater Imaging Mechanism and Spectral Degradation
The propagation of light in water is governed by the radiative transfer equation, which accounts for absorption and scattering by water molecules, suspended particles, and dissolved organic matter. For shallow waters, the remote sensing reflectance just below the surface, \(r_{rs}(\lambda)\), can be modeled as a superposition of water-column scattering and bottom reflectance:
$$
r_{rs}(\lambda) = r_{rs}^{\infty}(\lambda) \bigl\{1 – A_0 e^{-[K_d(\lambda)+K_u^C(\lambda)]H}\bigr\} + A_1\, r(\lambda)\, e^{-[K_d(\lambda)+K_u^B(\lambda)]H}
$$
where \(r_{rs}^{\infty}(\lambda)\) is the reflectance of optically deep water, \(r(\lambda)\) is the bottom albedo, \(H\) is water depth, and \(K_d(\lambda)\), \(K_u^C(\lambda)\), \(K_u^B(\lambda)\) are attenuation coefficients. The absorption coefficient \(a(\lambda)\) and backscattering coefficient \(b_b(\lambda)\) can be decomposed into contributions from pure water, phytoplankton, and dissolved/gelbstoff matter. The above-surface reflectance \(R_{rs}(\lambda)\) is then obtained via an air–water interface correction:
$$
R_{rs}(\lambda) = \frac{z_{rs} \, r_{rs}(\lambda)}{1 – G \, r_{rs}(\lambda)}
$$
where \(z_{rs}\) and \(G\) are empirical constants. This formulation provides a physically consistent forward model that can be used for both spectral prediction and restoration. The spectral degradation caused by water column absorption and scattering can be summarized as a multiplicative attenuation with an additive noise term \(n(\lambda)\):
$$
R_{obs}(\lambda) = \mathcal{F}(a_w, a_{ph}, a_{dg}, b_{bw}, b_{bp}, r, H) + n(\lambda)
$$
where \(\mathcal{F}\) is the shallow-water forward operator. Understanding these degradation mechanisms is essential for designing algorithms that can compensate for water-induced spectral distortions.
1.2 Hyperspectral Underwater Detection Datasets
To support algorithm development and evaluation, several benchmark datasets have been collected using China UAV drones and fixed platforms. Table 1 summarizes the key characteristics of the major publicly available datasets.
| Dataset | Spectrum (nm) | Bands | Image Size | Scene Types | Materials | Depth (m) |
|---|---|---|---|---|---|---|
| NPU-Pool | 400–1000 | 108 | 100×100 | Anechoic pool | Iron, stone, rubber | 0–3.1 |
| NPU-Sea | 400–780 | 108 | 350×350 | Coastal sea | Iron | 0.8–3.0 |
| HNU-UTD | 400–780 | 270 | 405×430 | Coastal sea | Concrete, vegetation | Shallow |
| ATR2-Lake | 400–1000 | 270 | 211×260 | Freshwater lake | Metal plate | 1.0–3.0 |
| ATR2-HUTD | 400–1000 | 270 | 3004×640 | Lake, river, bay | Metal, plastic, wood | 0.5–3.0 |
Among these, the ATR2-HUTD dataset, collected by our team using a China UAV drone equipped with a push-broom hyperspectral sensor (spectral resolution 2.2 nm), represents the most comprehensive benchmark. It systematically covers three water types (lake, river, bay) and multiple target materials (metal, plastic, wood) at varying depths, with synchronized measurements of water optical properties and reference spectra. This dataset enables rigorous cross-scene evaluation of detection algorithms under realistic nearshore conditions.
2. Key Technologies and Method Evolution
We categorize the existing methods into five classes according to the dominant operation in the detection pipeline. The evolution proceeds from physics-based analytical modeling to data-driven representation learning, as summarized in the following subsections.
2.1 Spectral Prediction Methods
Spectral prediction methods aim to compute the expected underwater spectrum of a target given its above-water reflectance and the water column parameters. Early work by Jay et al. (2010) established the “predict-then-match” paradigm, where the target signature is transformed via the radiative transfer model before comparison with observed pixels. Later extensions incorporated geometric manifold learning (Gillis, 2016) and joint inversion of depth and water properties (Qi et al., 2021). Our recent framework (Li et al., 2023) integrates depth estimation, spectral prediction, and spatial–spectral features into a single end-to-end network, achieving improved robustness in turbid waters. The core advantage of this class is physical interpretability and low dependency on labeled data, but its performance degrades when the assumed water parameters deviate from actual conditions.
2.2 Spectral Restoration Methods
Instead of forward prediction, spectral restoration methods invert the degradation to recover the in-air equivalent reflectance. A seminal contribution is UTD-Net (Qi et al., 2021), which formulates detection as a physically constrained linear unmixing problem within a deep autoencoder. The model learns to separate water column contributions from bottom and target endmembers, and uses the abundance maps for detection. Later works such as NUN-UTD (Liu et al., 2024) introduce nonlinear decoding to handle multiple scattering, while HUTD-Net (Li et al., 2025) adds a nonlinear compensation branch and adaptive abundance weighting. The conditional diffusion framework JURTD (Li et al., 2024) treats restoration as a generative process conditioned on principal components, capturing spectral variability due to depth and water quality. These methods excel at maintaining spectral consistency across different water conditions, but often require careful training to avoid overfitting to the synthetic data.
2.3 Band Selection Methods
Band selection reduces the spectral dimensionality by choosing a small subset of informative channels, thereby lowering computational cost while retaining discriminative power. Fu et al. (2020) introduced a constrained target optimal factor index that balances information content and target response. Qi et al. (2021) proposed a bathymetry-guided band selection that first constructs a physically meaningful subspace using depth information. Zhang et al. (2023) developed a fully unsupervised method that extracts feature bands directly from image statistics without any prior knowledge; it uses spectral water indices and unmixing to derive difference spectra and generates detection maps via band ratios. Although band selection methods offer the fastest inference speed, our evaluation shows they consistently underperform compared to other categories due to the loss of subtle spectral details.
2.4 Pixel Classification Methods
Pixel classification treats detection as a per-pixel binary classification problem. Wang et al. (2021) systematically compared various classifiers (SVM, neural network, Gaussian process) on corrected underwater hyperspectral data, finding SVM to be most robust. Huang et al. (2021) demonstrated the feasibility of in-situ microplastic detection by combining spectral correction with pixel-wise SVM. To address domain shift, Li et al. (2023) proposed a transfer framework that generates synthetic training samples using a shallow-water model, partitions the data by depth intervals, trains a set of depth-specific sub-networks, and selects the appropriate model during inference via a lightweight domain selector. This approach mitigates the distribution mismatch between synthetic and real data, achieving good accuracy in nearshore environments with limited real labels.
2.5 Feature Construction Methods
The most recent paradigm shifts the focus from “how well can we restore the spectra” to “how separable are target and background in a learned feature space.” Our work (Qi et al., 2025) proposes a hybrid-level contrastive learning framework that constructs a discriminative latent space using an anchor–positive–negative sampling strategy. A reliability-guided sample organization mechanism iteratively selects high-confidence samples to form clean seeds, and then multi-level contrastive losses (pixel, patch, and class prototype levels) shape the space. The target class prototype is anchored using a reference spectrum (physical prior) to ensure semantic stability. In a follow-up study (Qi et al., 2025), we introduced a physics-informed curriculum learning strategy that gradually incorporates hard examples while refining the class boundaries. The framework achieves the highest and most stable AUC scores across all three water types in the ATR2-HUTD benchmark, demonstrating superior generalization and robustness to environmental variations.
3. Comparative Performance Analysis
We evaluated representative algorithms from each category on the ATR2-HUTD dataset using the area under the ROC curve (AUC) as the metric. The results are summarized in Table 2.
| Category | Method | Lake AUC | River AUC | Bay AUC |
|---|---|---|---|---|
| Spectral Prediction | SUTDF (Qi et al., 2021) | 0.924 | 0.871 | 0.908 |
| Spectral Prediction | TDSS-UTD (Li et al., 2023) | 0.945 | 0.889 | 0.935 |
| Spectral Restoration | HUTDNet (Li et al., 2025) | 0.957 | 0.902 | 0.948 |
| Spectral Restoration | SVHAE (Aala et al., 2025) | 0.951 | 0.895 | 0.941 |
| Band Selection | FBU (Zhang et al., 2023) | 0.872 | 0.803 | 0.856 |
| Band Selection | BSU (Fu et al., 2020) | 0.845 | 0.779 | 0.832 |
| Pixel Classification | TUTDF (Li et al., 2023) | 0.938 | 0.883 | 0.919 |
| Pixel Classification | Water-X (Wang et al., 2021) | 0.911 | 0.845 | 0.897 |
| Feature Construction | PCL-HUTD (Qi et al., 2025) | 0.971 | 0.925 | 0.962 |
| Feature Construction | HUCLNet (Qi et al., 2025) | 0.968 | 0.918 | 0.958 |
The feature construction methods (PCL-HUTD and HUCLNet) achieve the highest and most consistent AUC values across all three water types, indicating superior adaptability to varying water conditions. Spectral restoration methods follow closely, with HUTDNet showing particularly strong performance in the river scenario. Spectral prediction methods perform well in lake and bay but suffer a noticeable drop in river (turbid conditions). Pixel classification methods exhibit moderate performance but higher variability. Band selection methods consistently underperform, confirming that the loss of spectral detail limits their discriminability in complex underwater scenes.
4. Challenges and Future Directions
Despite significant progress, several open challenges remain before China UAV drone-based hyperspectral underwater target detection can be deployed in routine operational monitoring. We identify three key directions for future research:
4.1 Differentiable Physical Modeling
Traditional radiative transfer models provide a physically interpretable forward mapping, but their high parameter dimensionality and non-differentiability hinder end-to-end learning. Future work should focus on converting key components—such as absorption, scattering, and bottom reflectance—into differentiable modules that can share gradient backpropagation with deep networks. This would allow joint optimization of physical parameters and detection objectives, improving parameter estimation accuracy and spectral reconstruction fidelity across varying water types. For China UAV drone applications, such a module could be embedded in the onboard processing pipeline to perform real-time water characterization and consistent correction, reducing the need for in-situ water quality measurements.
4.2 Uncertainty Characterization
Underwater hyperspectral observations are inherently stochastic due to illumination fluctuations, water surface waves, and sensor noise. Current detection algorithms typically produce point estimates without quantifying confidence. Future methods should incorporate uncertainty modeling at the pixel and task levels. Bayesian neural networks, Monte Carlo dropout, or diffusion-based resampling can provide per-pixel credible intervals and detection confidence maps. This “risk-aware” output allows operators to prioritize high-confidence detections and adjust thresholds adaptively. Uncertainty quantification also enables active learning strategies that select the most informative samples for labeling, thereby improving model efficiency under scarce annotations.
4.3 Cross-Scene Generalization Mechanisms
The spatial and temporal variability of water bodies causes significant domain shift when models trained on one site are applied to another. Two complementary pathways are promising: (1) physics-constrained domain adaptation, which uses water optical parameters, depth, and bottom type to build invariant representations; and (2) feature construction for domain reshaping, where contrastive learning and multi-level embedding align spectral distributions across sites. Generative models, such as conditional diffusion, can synthesize realistic cross-domain samples to augment training data. The ultimate goal is to achieve a “train-once, deploy-anywhere” capability for China UAV drone-based monitoring, dramatically reducing the cost of site-specific model retraining.
5. Conclusion
We have presented a comprehensive review of hyperspectral underwater target detection using China UAV drone imagery, covering imaging physics, spectral degradation models, benchmark datasets, and five major methodological paradigms. Our quantitative comparison on the ATR2-HUTD dataset reveals that feature construction methods currently offer the best trade-off between accuracy and generalization, while band selection remains inadequate for complex nearshore environments. The future of this field lies in the synergistic integration of differentiable physics, uncertainty awareness, and cross-scene adaptation. By advancing along these directions, China UAV drone hyperspectral remote sensing can evolve from a promising research tool into a reliable, routinely deployable technology for nearshore ecological monitoring, resource management, and maritime security.