Nearshore ecosystems are critical components of the marine environment, playing indispensable roles in maintaining biodiversity, regulating carbon cycles, and supporting fisheries production. The combined impacts of human activities and climate change have led to continuous degradation of nearshore ecosystem structures and functions, severely undermining their self-regulation capacity and posing realistic risks to the sustainable utilization of marine resources and regional ecological security. In this context, constructing a precise, rapid, and wide-area coverage nearshore underwater observation system has become a core scientific issue in the fields of ecological monitoring and resource management.
Underwater target detection serves as a fundamental component of nearshore ecological observation systems, aiming to localize and identify underwater targets and their characteristics to obtain critical information reflecting seabed landscape structures, ecological status, and human disturbances. This information has direct application value for marine ranching management, pollution source tracing, and maritime defense security. Therefore, the accuracy and stability of detection directly affect the decision-making quality of nearshore ecological protection and resource governance. Nearshore water bodies exhibit complex optical properties, frequent environmental disturbances, and diverse spatial scales, which continuously constrain the accuracy, efficiency, and robustness of traditional underwater target detection methods. Suspended particles, plankton, and dynamic media cause significant scattering and absorption, jointly leading to a decrease in the signal-to-noise ratio of observed signals and weakened target-background separability. Moreover, nearshore areas undergo rapid spatiotemporal changes with large variations in target scales, making it difficult for traditional underwater target detection methods to simultaneously meet the dual demands of spatial coverage and temporal resolution. Therefore, developing new underwater target detection methods that combine wide-area coverage, high timeliness, and quantitative reliability has become a critical challenge for achieving nearshore situational awareness.
Remote sensing technology has opened new pathways to address these challenges. Leveraging multi-platform, multi-modal observation systems, remote sensing can achieve high spatiotemporal resolution monitoring of large-scale nearshore waters, showing significant advantages in data acquisition, dynamic observation, and long-term series analysis. Among various remote sensing data types, visible light imagery is the most widely used, commonly applied in tasks such as environmental monitoring, landform recognition, and water quality assessment. Such images can effectively characterize spatial features like texture, shape, and color. However, in nearshore underwater target detection scenarios, their performance is limited by the optical absorption properties of water. Studies indicate that chlorophyll in water exhibits the most significant absorption around 0.45 μm (blue light) and 0.65 μm (red light), leading to markedly reduced reflectance and weakened effective radiation information. Consequently, spatial texture features are susceptible to energy loss and noise interference, diminishing target-background separability. In contrast, hyperspectral imaging, relying on hundreds of continuous narrow bands for fine sampling, provides rich spectral information within the visible to short-wave infrared range. This data not only reveals the reflection behavior of materials at different wavelengths but also enables fine discrimination of composition and state using subtle differences in spectral curves. Benefiting from this, hyperspectral technology has distinct advantages in nearshore environments with complex optical conditions and prominent background interference. Compared to traditional images that mainly rely on spatial information, hyperspectral imagery can construct a higher-dimensional representation space in the spectral domain, enabling effective discrimination of targets that are visually similar but have significantly different spectral properties. Furthermore, its wide band coverage can mitigate degradation effects caused by water absorption and scattering to some extent, providing key data support for fine modeling of underwater targets and backgrounds.
With the rapid advancement of aerospace and unmanned systems technology, hyperspectral remote sensing has formed a collaborative observation system comprising multiple platforms such as satellites, traditional airborne systems, and China UAV drones. Different platforms exhibit significant differences in various aspects. Satellite platforms offer stable imaging and wide coverage, suitable for large-scale, long-term background monitoring and trend assessment. However, their spatial resolution and revisit cycles are often constrained by factors like orbit and cloud cover, and data downlink and ground processing typically introduce non-negligible delays. In contrast, China UAV drones can take off on demand in mission areas and quickly acquire nearshore hyperspectral data, offering higher spatial resolution and lower acquisition latency. This makes them particularly suitable for refined detection scenarios involving small-scale, rapidly changing targets. In the multi-platform system, China UAV drones demonstrate unique potential for nearshore hyperspectral underwater observation due to their high maneuverability, low cost, and real-time capabilities. The near-ground flight characteristics of UAVs make data acquisition relatively less affected by atmospheric attenuation, resulting in higher image signal-to-noise ratios compared to high-altitude platforms. Furthermore, China UAV drone systems can flexibly adapt to tidal changes and meteorological disturbances, enabling continuous and efficient target detection in dynamic and complex nearshore environments. However, compared to satellite or traditional airborne platforms, the low-altitude, high-maneuver observation of China UAV drones also introduces a series of more challenging platform-level issues. Flight attitude jitter and trajectory fluctuations can easily cause geometric distortion, band curvature, and inter-frame registration errors in hyperspectral push-broom imaging, affecting target boundaries and spectral consistency. Additionally, under low-altitude conditions, illumination can fluctuate rapidly due to cloud shadows, changes in solar elevation angle, and sea surface specular reflection, leading to radiometric inconsistencies and exacerbating water and sea-surface background interference. Furthermore, the high-volume nature of high-resolution hyperspectral data, combined with limited onboard computing resources, makes near-real-time processing, lightweight deployment, and edge-cloud collaboration rigid requirements for engineering implementation. These differences mean that directly migrating satellite or airborne algorithms to China UAV drone data often fails to achieve stable performance. There is an urgent need to develop attitude-radiation joint calibration, robust spectral-spatial feature modeling, and low-latency detection algorithms tailored to the UAV imaging chain. Therefore, to meet the implementation requirements of nearshore underwater target detection, hyperspectral observation based on China UAV drones is becoming an important development direction, providing a new technological breakthrough opportunity for achieving high-precision intelligent observation of underwater environments. To facilitate comparison, Table 1 summarizes the typical differences between satellite and UAV platforms in relevant indicators for nearshore underwater target detection.
| Comparison Dimension | Satellite Platform | UAV Platform |
|---|---|---|
| Deployment Flexibility and Revisit Cycle | Low: Constrained by orbital period (days/weeks), weak emergency response capability, difficult to acquire on demand | Very High: On-demand capability, enabling minute-level deployment and high-frequency re-measurement |
| Spatial Resolution | Coarse/Medium: Typically 10–30 m, difficult to identify small targets or fine textures | Ultra-High: Can reach centimeter to decimeter level, capable of capturing sub-pixel spectral mixing details |
| Imaging Geometry and Attitude Stability | Very High: Stable platform, geometric distortion mainly due to Earth curvature and terrain, easy to correct | Low: Affected by low-altitude airflow and rotor vibration, significant attitude jitter, requiring high-frequency gimbal and complex geometric correction algorithms |
| Radiometric Consistency and Environmental Sensitivity | Atmospheric influence dominant: Long atmospheric path, requiring strict atmospheric correction; relatively uniform illumination conditions | Sensitive to illumination fluctuations: Drastic effects from cloud shadows, water surface glint, and surface reflection at low altitude; rapid illumination changes over short periods, making radiometric correction challenging |
| Payload Capacity and Signal-to-Noise Ratio | High: Equipped with large cooling sensors, long integration time, excellent SNT | Limited: Constrained by size, weight, and power; mostly uses lightweight uncooled equipment, susceptible to noise interference |
| Data Processing and Computational Constraints | Ground-based post-processing: Significant data downlink delay, relies on ground high-performance clusters, poor real-time capability | Edge-side constrained computing: Limited onboard computation (embedded GPU/FPGA), but high demands for edge computing and real-time decision-making |
| Nearshore/Water Body Detection Applicability | Suitable for macroscopic water color remote sensing (e.g., chlorophyll, suspended sediment distribution), difficult to monitor dynamic small targets | Suitable for refined target detection (e.g., submerged reefs, camouflaged targets near discharge outlets), needs to overcome strong dynamic water interference |
1. Research Foundation
The study of hyperspectral underwater target detection is inseparable from a deep understanding of the imaging physical process and a systematic analysis of data characteristics. Underwater environments exhibit significant optical complexity. Incident radiation undergoes multiple absorption and scattering events in the water column, causing severe attenuation and distortion before the signal reaches the sensor. This leads to spectral feature shifts, energy loss, and reduced target separability. Therefore, establishing an imaging mechanism model that reflects the laws of light propagation and environmental influences is a theoretical prerequisite for achieving accurate underwater target detection. Based on this, by analyzing the main degradation factors in the imaging process, a physically interpretable degradation model can be derived, providing necessary constraints and key priors for subsequent algorithm design and inversion solutions. Simultaneously, data is the core support for underwater target detection research. Influenced by imaging conditions, equipment performance, and field sampling difficulties, existing hyperspectral underwater datasets differ significantly in band coverage, scene diversity, and annotation accuracy. Systematically reviewing the composition, acquisition methods, and applicable scope of these datasets helps clarify experimental conditions and comparability boundaries, providing an objective baseline for model validation and performance evaluation.
1.1 Imaging Mechanism and Degradation Model
1.1.1 Overview of Imaging Mechanism
Hyperspectral underwater imaging mechanism describes the propagation, scattering, and absorption process of incident radiation in the water medium. The spatial distribution and concentration gradient of seawater collectively determine the spatial attenuation and spectral distortion of radiation energy during propagation, thereby affecting path radiation, backscattering, and target-background spectral separability. From the perspective of radiative transfer, the radiance L within a unit solid angle satisfies the radiative transfer equation:
$$
\mu \frac{dL(\lambda, z, \mu, \phi)}{dz} = -[a(\lambda) + b(\lambda)] L(\lambda, z, \mu, \phi) + \frac{b(\lambda)}{4\pi} \int_{4\pi} L(\lambda, z, \mu’, \phi’) P(\mu’, \phi’; \mu, \phi) d\Omega’
$$
where \( a(\lambda) \) is the absorption coefficient describing energy loss per unit path length, \( b(\lambda) \) is the scattering coefficient indicating the intensity of energy transferred from the original direction to other directions, and \( P(\mu’, \phi’; \mu, \phi) \) is the phase function. The radiance received by the sensor originates from two sources: water column scattering and bottom reflection, represented as:
$$
r_{rs}(\lambda) = r_{rs}^{\infty}(\lambda) \{ 1 – A_0 e^{-[K_d(\lambda) + K_C^u(\lambda)]H} \} + A_1 r(\lambda) e^{-[K_d(\lambda) + K_B^u(\lambda)]H}
$$
where \( r_{rs}^{\infty}(\lambda) \) is the subsurface reflectance under “optically deep water” conditions, \( r(\lambda) \) is the bottom albedo, \( H \) is the water depth, \( K_d(\lambda) \) is the downwelling attenuation coefficient, and other terms represent upwelling attenuation contributions. This equation reflects the dual-path superposition characteristic of the sensor signal. To link subsurface reflectance with the measured reflectance above the atmosphere, a water-surface conversion relationship is typically used:
$$
R_{rs}(\lambda) = \frac{z_{rs} r_{rs}(\lambda)}{1 – G r_{rs}(\lambda)}
$$
1.1.2 Spectral Degradation Mechanism Analysis
During hyperspectral imaging, incident radiation undergoes multiple absorption and scattering processes, causing significant distortion in spectral domain features. The main mechanisms leading to spectral degradation include energy attenuation, bottom reflection and water depth coupling, multi-path scattering and path radiation interference, as well as sensor response and instrument noise. The observed reflectance model can be expressed as:
$$
R_{obs}(\lambda) = F(a_w, a_{ph}, a_{dg}, b_{bw}, b_{bp}, r, H) + n(\lambda)
$$
where the function \( F(\cdot) \) is the shallow water forward operator based on the radiative transfer equation, and \( n(\lambda) \) represents imaging system and environmental noise. To reduce ill-posedness in numerical computation and inversion, the absorption coefficient is decomposed into contributions from pure water, phytoplankton, and dissolved organic matter, while the backscattering coefficient is decomposed into contributions from pure water and particles.
1.2 Hyperspectral Underwater Target Detection Datasets
Research on hyperspectral underwater target detection under complex environments has gradually formed a multi-level data system, progressing from controlled experiments to real sea areas, from single materials to multiple target types, and from small-scale image patches to large-scale imagery. Early studies were mostly conducted in anechoic pools and nearshore shallow waters using fixed supports or handheld hyperspectral devices, focusing on mechanism validation and detectability assessment with limited scene complexity. Subsequently, the application of China UAV drones equipped with hyperspectral imaging systems has matured, achieving centimeter-level spatial resolution and spectral sampling with hundreds of narrow bands, significantly expanding the coverage of water types (e.g., freshwater lakes, estuaries, nearshore bays) and target categories (e.g., metal, wood, plastic, composite substrates). This has simultaneously enhanced scene diversity and physical realism. However, due to the high cost of field experiments and the difficulty of annotation, publicly available, complete hyperspectral underwater benchmark data remains relatively scarce, especially in terms of a unified annotation system and cross-scene comparability under coupled conditions of water depth, optical parameters, and bottom albedo. Table 2 provides key information on the main current datasets.
| Dataset | Spectral Range (nm) | Number of Bands | Average Image Size | Scene Type | Materials | Depth (m) |
|---|---|---|---|---|---|---|
| NPU-Pool | 400–1000 | 108 | 100 × 100 pixels | Anechoic Pool | Iron, Stone, Rubber | 0–3.1 |
| NPU-Sea | 400–780 | 108 | 350 × 350 pixels | Nearshore Sea | Iron | 0.8–3.0 |
| HNU-UTD | 400–780 | 270 | 405 × 430 pixels | Nearshore Sea | Wave Breakers, Cement, Vegetation | – |
| ATR2-Lake | 400–1000 | 270 | 211 × 260 pixels | Freshwater Lake | Metal Plate | 1.0–3.0 |
| ATR2-HUTD | 400–1000 | 270 | 3004 × 640 pixels | Lake, River, Bay | Metal Plate, Wood, Plastic | 0.5–3.0 |
2. Key Technologies and Methodological Progress
Hyperspectral underwater target detection has evolved from an early stage dominated by physical modeling into a multi-level system that combines mechanism constraints with learned representations. This evolution was not simply a linear progression over time but was driven by escalating application requirements and methodological bottlenecks. As nearshore scenes become increasingly complex and tasks shift from qualitative identification to quantitative inversion and online monitoring, relying solely on analytical models makes it difficult to simultaneously handle complex water conditions, cross-regional generalization capabilities, and computational overhead. This forces methods to continuously introduce data-driven and feature-learning mechanisms while maintaining physical interpretability. The methods have undergone a systematic evolution from mechanism modeling based on spectral prediction or restoration, to feature reduction based on band selection, to supervised discrimination based on pixel classification, and finally to latent space shaping based on feature construction. Although different stages have different focuses, their common goal is to maintain target-background spectral separability in complex underwater environments and achieve robust detection in an interpretable and verifiable manner. The categorization of these methods is based on the dominant role in the detection pipeline, focusing on spectral inference, spectral restoration, input band selection, pixel-level direct discrimination, and discriminative feature construction.
2.1 Underwater Target Detection Based on Spectral Prediction
This approach uses shallow-water radiative transfer mechanisms to map terrestrial target spectra into “subaqueous response spectra” consistent with specific water conditions and depths, followed by spectral matching or statistical discrimination. The process forms a closed loop of “imaging conditions, spectral mapping, and discriminant inference,” making the detection process usable and verifiable in sample-scarce scenarios. Key techniques include establishing a quantitative mapping from ground reflectance to underwater observed reflectance under specified water and depth conditions, introducing spectral domain geometric modeling for nonlinear compensation, and integrating depth estimation with spectral restoration and target discrimination within a joint optimization framework. The main advantage lies in its clear physical interpretability, friendly data requirements, and good extensibility. However, its most prominent limitation is the amplification effect of environmental description errors, where deviations in inherent optical parameters, angular scattering distribution, water surface and bottom interface reflection properties, and incident lighting conditions can be progressively amplified during the spectral mapping process, thereby degrading detection stability. Early schemes mainly focused on shallow water and single materials, lacking robustness across water bodies, seasons, and multiple substrates. Even after integrating deep learning, models can still suffer from out-of-distribution degradation if the environmental coverage of training samples is insufficient. Furthermore, the serial implementation of equivalent spectrum generation and discrimination makes error propagation difficult to avoid when environmental estimation and target detection are coupled. The future breakthrough depends on the synergistic advancement of differentiable mechanisms, uncertainty constraints, spatiotemporal coupling, and self-supervised learning to achieve robust generalization across scenes and temporal phases while maintaining physical credibility.
2.2 Underwater Target Detection Based on Spectral Restoration
The central idea of this paradigm is to reconstruct or separate the intrinsic response of the target from distorted observations and establish a verifiable correspondence with ideal priors. The process follows a “restoration, prior alignment, discrimination” pipeline. Current practices present two complementary technical paths: separation-based modeling using spectral unmixing, where underwater observations are represented as weighted combinations of endmembers like water, substrate, and target, with detection metrics constructed from abundance or reconstruction residuals; and joint modeling using generative restoration, where conditional generative models learn “distorted-to-intrinsic” spectral mappings, coupling restoration and detection optimization through multi-task loss. The main advantage of this route is its core constraint of physical laws, incorporating the causal relationship between target spectral shape and water environment into an interpretable modeling framework. By reconstructing the equivalent underwater spectrum consistent with the actual scenario under the radiative transfer mechanism, the detection process traces back from the observation layer to the physical cause layer, ensuring the traceability of spectral-level evidence and the mechanism consistency of results. Due to this “restoration for detection” design, models can maintain stable performance in sample-scarce or unknown environments, exhibiting good transferability and testability. Additionally, spectral restoration methods are structurally easy to expand synergistically with modules like unmixing, depth estimation, and water quality inversion, enabling closed-loop optimization from spectral restoration to target discrimination within a unified physical framework. This provides high theoretical completeness and engineering applicability while balancing interpretability, robustness, and generalization ability in complex nearshore environments.
2.3 Underwater Target Detection Based on Band Selection
This approach transforms the conventional “full-spectrum participation” into “supporting detection with fewer and more accurate bands.” It first selects a band subset with high target-background discriminability and strong robustness against complex water background interference from the candidate bands, then constructs a lightweight classifier based on this subset. The research presents two main ideas: mechanism-driven, which uses imaging elements like water depth and attenuation intensity to establish a physical evaluation criterion for bands, prioritizing bands that can stably characterize the “equivalent response of the target in a given water body”; and data-driven, which directly measures the discriminability and redundancy of each band in the observation domain regarding the characterization differences between targets and complex backgrounds. Band selection can serve as an independent front-end for simplifying the spectral dimension of existing detection methods or be co-designed with the back-end classifier to optimize the trade-off between discriminability and robustness. This approach explicitly prioritizes band selection, replacing full-spectrum dependence with a smaller, more physically and task-relevant set of bands, achieving a better accuracy-efficiency balance in real nearshore scenes and providing more reliable spectral input for downstream tasks.
2.4 Underwater Target Detection Based on Pixel Classification
This paradigm converts underwater target detection into a per-pixel binary classification task. It typically proceeds in two steps: first, correcting imaging and environmental influences to reduce water effects; second, using supervised classifiers to directly distinguish “target present” from “target absent” at the pixel level. The basic architecture follows “observation correction, feature modeling, and discriminant learning.” The observation correction step aims to restore or enhance material-related discriminant information, reducing systematic biases caused by depth and water quality. Feature modeling can directly use raw spectral information or combine local spatial consistency for improved robustness under noise. Discriminant learning paths include “corrected traditional classification” using methods like support vector machines and neural networks, or “synthetic training, domain selection, and rapid inference,” which uses water depth partitioning and domain adaptation modules to mitigate performance degradation caused by distribution mismatches between real and synthetic data. Methodologically, this route is characterized by clear decision objectives and clear engineering workflows. Pixel-level direct discrimination avoids the need for comprehensive inversion of water parameters and is easily decoupled from front-end modules like spectral correction and band selection. Its focus on “learnable decision boundaries” is suitable for scenarios with available data but difficult to fully model mechanism. It provides a practical discriminative path complementary to mechanism modeling, extending its applicability to real complex underwater environments through synthetic training and domain selection.
2.5 Underwater Target Detection Based on Feature Construction
This paradigm rephrases the traditional chain process of “first spectral or physical transformation, then discrimination” into “first learning a discriminative feature space, then performing measurement and detection within that space.” Its core idea is to learn a semantically clear and geometrically stable feature space centered around target-background separability. The detection task is to achieve robust separation of target and background in this latent space using spectral metrics. In this approach, reference spectra are no longer simply used for template matching but are endowed with the role of “semantic anchors,” providing a stable reference position for the target class in the latent space and guiding target samples to cluster around it. This process can be understood as “semantic anchoring.” The focus is on constructing a discrimination-friendly target representation that can accommodate morphological drift caused by underwater propagation while maintaining robust separation from the background. This “separability-first” modeling stance avoids the serial process that amplifies environmental errors step-by-step and does not reframe the detection problem as restoration or pixel classification. Recent explorations further show that feature construction does not exclude physical constraints but absorbs mechanism information in a weaker, more flexible manner. The overall approach forms a unified paradigm highly compatible with complex nearshore situations: it no longer revolves around “how accurately the spectrum is restored,” but focuses on “how clear the separation is,” achieving stable target-background separability without explicitly solving underwater imaging details.
2.6 Comparative Analysis of Different Methods
To verify the actual performance of different algorithm types, this section selects representative and advanced methods from each category and conducts a performance analysis on the real nearshore hyperspectral underwater target detection dataset ATR2-HUTD, using the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve as the performance validation metric. The performance comparison of different types of methods on the ATR2-HUTD dataset is shown in Table 3. The results indicate that different categories of methods exhibit distinct performance characteristics across the three water types: lake, river, and sea water. Feature construction methods perform the best, maintaining the highest levels and strongest stability overall. They effectively enhance target-background separability under complex water conditions and reduce the impact of environmental disturbances on detection results. These methods show relatively small performance fluctuations across different water bodies, further demonstrating their superior cross-scene transferability and adaptation to complex conditions. Spectral restoration methods generally perform next best, showing good balance in most scenes, especially in river scenes where they still achieve high detection accuracy. This suggests that restoring the effective target spectrum affected by water transmission distortion can mitigate target information loss and distortion in complex environments. Spectral prediction methods perform well in lake and sea scenes, sometimes approaching optimal detection results. However, their performance degrades noticeably in river scenes, reflecting strong dependence on prior consistency, environmental stability, and prediction accuracy. Pixel classification methods can achieve high accuracy in some scenes, showing strong direct discriminative ability. Nevertheless, their cross-scene fluctuation is large, particularly in river scenes, indicating a strong dependence on training-test distribution consistency. Band selection methods show the weakest overall performance, with significantly lower detection accuracy in most scenes compared to other categories. This indicates that relying solely on a few sensitive bands struggles to stably support underwater target detection tasks in complex environments, as discriminant information is often distributed across multiple bands and their combinations.
| Method Type | Lake Scene | River Scene | Sea Scene |
|---|---|---|---|
| Spectral Prediction | 0.93 | 0.85 | 0.92 |
| Spectral Restoration | 0.94 | 0.90 | 0.93 |
| Band Selection | 0.85 | 0.78 | 0.82 |
| Pixel Classification | 0.92 | 0.84 | 0.90 |
| Feature Construction | 0.96 | 0.93 | 0.95 |
3. Summary and Prospects
This review systematically examines the research progress in nearshore underwater target detection using China UAV drone-based hyperspectral remote sensing. It focuses on the main threads of imaging mechanisms, characteristic modeling, and algorithm design. Starting from the shallow-water radiative transfer mechanism, it summarizes the core ideas and evolutionary logic of five categories of methods: spectral prediction, spectral restoration, band selection, pixel classification, and feature construction. The study shows that the development of underwater target detection has evolved from a physical process-driven stage dominated by analytical modeling to a new stage of mechanism and data synergy. The former emphasizes the interpretable mapping of imaging conditions and water parameters, while the latter uses learnable representation spaces to achieve environmental adaptability and feature robustness. Benefiting from this, the accuracy, generalization, and adaptability of detection algorithms to complex sea areas have been significantly improved. The key trends for future research can be summarized in three directions:
1) Differentiable Physical Modeling. The core of differentiable physical modeling is to transform analytical expressions such as radiative transfer equations and endmember mixing into differentiable modules that share gradient channels with deep networks. This allows for synchronous “model-data” constraints during training. Such mechanisms can improve the accuracy of parameter identification and spectral domain reconstruction while preserving physical interpretability, providing a unified physical prior guidance for target detectability across different depths and water qualities. From an application perspective, differentiable physical modules can serve as unified calibration and constraint units in onboard or ground processing chains. By jointly estimating key water parameters and imaging conditions, they provide a more consistent input distribution and interpretable priors for subsequent detection, thereby reducing systematic radiometric biases between different sea areas and data batches. This is expected to enhance the stability of detection under cross-water quality and cross-depth conditions, reduce reliance on field synchronous measurements or repeated calibrations, and maintain a more controllable performance degradation range under limited annotation conditions.
2) Uncertainty Characterization. Illumination perturbations, substrate heterogeneity, and instrument noise in complex sea environments can cause the statistical distribution of detection results to deviate from expectations. Future models need to explicitly incorporate uncertainty modeling mechanisms, using Bayesian estimation, confidence prediction, or diffusion-based resampling to characterize pixel-level spectral confidence intervals and task-level confidence intervals. On one hand, uncertainty quantification can support the construction of “risk-aware” detection strategies, enabling result credibility ranking and adaptive threshold adjustment. On the other hand, it provides a quantifiable basis for subsequent active learning and task evaluation, ensuring the model output not only includes discriminative results but also possesses reliability expression. This is crucial for decision-making in safety-critical applications like maritime surveillance and environmental management.
3) Cross-Scene Generalization Mechanisms. The spatiotemporal heterogeneity of underwater environments causes significant fluctuations in model performance when transferring between different water bodies. Future research on cross-scene generalization should proceed along two paths simultaneously. One path is mechanism-constrained domain adaptation, which uses water optical parameters, depth information, and bottom reflection properties to construct domain-invariant representations. The other path is domain reshaping based on feature construction, using contrastive learning, multi-level embedding, and spectral domain alignment strategies to maintain stable target-background separability under distribution shifts. Generative modeling and diffusion-based learning offer new tools for virtual reconstruction and augmentation of cross-domain samples, supporting wider environmental generalization under limited measured data conditions. From an application perspective, the level of generalization determines whether a model can be cost-effectively reused across multiple sea areas, time periods, and different payload platforms. If stable decision boundaries can be maintained under distribution shifts, it can significantly reduce the re-training and re-annotation burden during cross-regional deployment and improve the efficiency of task initiation and iterative updates. Progress in this direction is expected to manifest as a higher lower bound on performance after transfer and smaller fluctuations, as well as maintaining usable detection quality under sample-limited conditions.
In summary, the future of China UAV drone-based hyperspectral nearshore underwater target detection will no longer focus solely on improving individual performance indicators but will aim for the overall optimization of “mechanism, uncertainty, and generalization.” Differentiable physical modeling provides a learnable framework under theoretical constraints, uncertainty characterization constructs an expression channel for model credibility, and cross-scene generalization mechanisms establish a robust foundation for real-world applications. Only when these three aspects are integrated within the task pipeline, such as moving from physical consistency constraints to risk-aware outputs and then to continuous cross-domain adaptation, can the performance gains be stably and evaluatively translated into operational benefits. The synergistic development of these three directions is expected to advance underwater hyperspectral detection from a state of “feasibility” to a new stage of “verifiability, reproducibility, and generalizability,” providing more reliable theoretical and technical support for target identification and ecological monitoring in complex nearshore environments.