Water-Conditioned Adaptive Perceptor (WCAP)

• WCAP is an adaptive computational framework that improves underwater perception by integrating water-specific modeling and data-driven adaptation.
• It leverages techniques like Riemannian metric warping, uncertainty quantification, and active sensing to enhance object detection and state estimation in aquatic environments.
• Empirical evaluations demonstrate significant gains in segmentation accuracy, demand forecasting reduction, and visual coverage in diverse water-based applications.

A Water-Conditioned Adaptive Perceptor (WCAP) is a class of adaptive computational modules or frameworks designed to optimize perception or inference under dynamic aquatic or water-degraded conditions. WCAPs leverage both physical modeling and advanced learning techniques to explicitly compensate for the challenges posed by underwater optics, sparse measurement, or fluctuating aquatic environmental factors, and have seen deployment in domains including camouflaged object detection, digital twinning of water networks, and robotic active perception. Characteristically, a WCAP integrates local or global data-driven adaptation—such as metric tensor prediction, uncertainty quantification, or parameter suggestion—mediated by water-specific cues or models, to improve estimation fidelity, segmentation, or image acquisition in the presence of water-originated distortions.

1. Motivation and Problem Context

Water environments introduce a suite of perceptual challenges: in underwater computer vision, nonuniform light attenuation, turbidity, and backscattering degrade visual signals, impeding the detection of faint, transparent, or camouflaged targets. Traditional convolutional architectures employing rigid, Euclidean sampling grids are ill-suited for such nonstationary, physically-induced distortions (Wu et al., 3 Feb 2026). Similarly, water distribution networks (WDNs) monitored by digital twins suffer from limited sensor placement and high uncertainty, necessitating efficient and uncertainty-aware state estimation (Homaei et al., 6 Nov 2025). For autonomous underwater vehicles, the mission-critical task of visual inspection is likewise degraded by variable turbidity, requiring real-time adaptation of imaging parameters to maintain coverage and quality (Cardaillac et al., 23 Apr 2025).

The WCAP concept arises as a unified approach to these problems: it adaptively modulates the perception process—whether at the level of convolutional sampling, measurement allocation, or active sensor control—so that perception remains robust to the intrinsic variability and degradation associated with water environments.

2. Mathematical Foundations and Model Structure

2.1 Riemannian Metric Warping for Image Features

In underwater camouflaged object detection, the core WCAP framework reformulates feature sampling on a Riemannian manifold (Wu et al., 3 Feb 2026). Given a feature map $F \in \mathbb{R}^{H \times W \times C}$, a global context descriptor $g$ is derived via global average pooling (GAP), then mapped to a set of real-valued parameters $\{a,b,c\}$ via a multilayer perceptron (MLP). These parameters comprise a lower-triangular matrix $L(g) \in \mathbb{R}^{2 \times 2}$, from which a symmetric positive-definite (SPD) metric tensor $G(g) = L(g) L(g)^{T} + \epsilon I_2$ is constructed.

This tensor governs the warping of convolutional receptive fields: standard kernel offsets $\Delta p$ are stretched according to the geodesic metric $d_G(\Delta p) = \sqrt{\Delta p^T G(g) \Delta p}$, and a grid sampling operation shifts each kernel sample via the transformation $T_G(\Delta p) = G(g)^{-1/2} \Delta p$. This design allows the model to stretch sampling along directions of greatest physical image degradation, compensating for light attenuation and scattering unique to underwater scenes.

2.2 Adaptive Uncertainty-Driven Sensing

For digital twins of WDNs, WCAP operationalizes adaptation as sensor allocation driven by per-node uncertainty (Homaei et al., 6 Nov 2025). A node-specific LSTM is trained to forecast short-term demand, while a split conformal prediction (CP) layer computes empirical prediction intervals from calibration residuals. For each time step, the node-wise interval width $U_t^{(i)} = 2 \hat{Q}^{(i)}_{1-\alpha}$ serves as the uncertainty metric; monitoring resources are greedily directed to the $B$ nodes with greatest $U_t^{(i)}$, optimizing measurement utility under budget constraints.

2.3 Data-Driven Active Perception

In robotic underwater inspection, the WCAP framework employs an active perception module that predicts the optimal camera-target distance and light intensity to maximize image quality under dynamically varying water conditions (Cardaillac et al., 23 Apr 2025). A physics-based light propagation model, blended with a large synthetic dataset and a contrast-predicting MLP, enables online estimation of expected image quality metrics for candidate control settings. The optimal settings are selected via a derivative-free optimization method (Nelder–Mead simplex), balancing contrast and survey coverage as specified by mission priorities.

3. Implementation Paradigms and Integration

3.1 Convolutional Neural Network Integration

In DeepTopo-Net (Wu et al., 3 Feb 2026), WCAP is positioned between the encoder (a ViT-based masked autoencoder, MAE) and the topology-aware decoder (Abyssal-Topology Refinement Module, ATRM). The enhanced features output by WCAP are produced as follows:

• $F = \text{Encoder}(I)$
• $g = \text{GAP}(F)$
• $G = \text{Metric}(g)$
• $F_{\text{warped}} = \text{Warp}(F, G)$
• $$F_{\text{out}} = F_{\text{warped}} + \text{FreqGate}(F_{\text{warped}}, g)$$

Here, FreqGate applies a Laplacian filter to extract high-frequency cues, fused with global context via gated summation. ATRM then further processes $F_{\text{out}}$ for segmentation.

3.2 Digital Twin Sensor Allocation

The implementation for WDNs (Homaei et al., 6 Nov 2025) consists of:

• Node-wise, two-layer LSTM sequence models ($d=64$)
• On-line autoregressive forecasting
• Split conformal prediction for interval calibration and real-time per-node uncertainty scoring
• Greedy selection of uncertain nodes for measurement, with control loop running at 5–10 Hz, scalable to hundreds of nodes with sub-second latency

3.3 Underwater Active Perception Control

The robotic perception WCAP (Cardaillac et al., 23 Apr 2025) integrates:

• Physics-based rendering in Blender, with wideband absorption/scattering, Fournier–Forand phase function, and physics-derived noise models
• A 20-input MLP trained on 138,240 synthetic renders, predicting contrast as a function of environmental and control state
• Online optimization (Nelder–Mead) for (Δd, ΔI), supporting real-time vehicle control within a ROS architecture

4. Quantitative Evaluation and Empirical Performance

Extensive empirical results validate the impact of WCAP across domains:

Configuration	MAS3K mIoU	RMAS mIoU	GBU-UCOD mIoU	GBU-UCOD S_α
Baseline (MAE)	0.770	0.708	0.785	0.865
Baseline + WCAP	0.785	0.722	0.808	0.894

On the GBU-UCOD dataset, the addition of WCAP delivered a +2.3% mIoU and +2.9% S_α gain over the MAE baseline, particularly improving the segmentation of low-contrast and slender structures (Wu et al., 3 Feb 2026).

In WDN digital twins at 40% sensor coverage (Homaei et al., 6 Nov 2025):

• Demand RMSE reduced by 33–34% compared to uniform random sampling
• Empirical coverage rates tightly matched nominal 90% prediction interval targets
• Pressure RMSE and violation rates improved by ~34% and ~45%, respectively

In robotic inspection simulations employing WCAP for distance and illumination adaptation (Cardaillac et al., 23 Apr 2025):

• Visual coverage area up to 127% greater under low turbidity compared to static patterns
• Model MAE in predicting image standard deviation: 0.01 (calibrated), with 40–92% error reduction under increased turbidity
• Feature-matching inlier ratios improved (0.42 vs 0.35), signifying denser and more reliable visual features for mapping or inspection

5. Design Principles, Regularization, and Transferability

Key design choices for WCAP implementations include minimal, physically-motivated parametrization (three explicit parameters via Cholesky decomposition of the metric tensor for SPD guarantee in (Wu et al., 3 Feb 2026)), interpretable adaptive sampling anchored in explicit water-induced distortions, and end-to-end trainability under compound losses (e.g., reconstruction + segmentation).

Metric regularization is not explicitly imposed in underwater object detection; however, for other domains where excessive warping may harm interpretability or alignment, a Frobenius-norm regularizer $\|G - I\|_F^2$ can be employed to constrain deviation from the identity metric unless data compels otherwise.

The WCAP principle generalizes beyond underwater domains. Scenarios with spatially-varying distortion, such as aerial imaging under refraction or medical endoscopy, are amenable to similar adaptive metric learning; one may extend the metric to higher dimensions or modify the data-driven adaptation block as required.

6. Practical Deployment and Toolchain

WCAP modules are typically lightweight, requiring only small MLPs and grid sampling operators, and are portable across encoder architectures: adapting the descriptor dimension $d_g$ and retraining suffices for transfer.

Implementation templates include:

• Standard initialization (Xavier or He) for neural branches
• AdamW or Adam optimizers; learning rates in the range $5 \times 10^{-5}$ to $10^{-3}$
• Batches of 16–32; input resolutions from 384×384 (vision) to hundreds of nodes (WDN)
• Open-source toolchains: PyTorch and TensorFlow for learning modules, EPANET Toolkit for hydraulic solvers, ROS nodes for robotic interfacing, and Blender for photorealistic simulation (Wu et al., 3 Feb 2026, Homaei et al., 6 Nov 2025, Cardaillac et al., 23 Apr 2025)

7. Impact, Limitations, and Extensions

WCAP establishes a data-driven, domain-aware standard for adaptive perception in environments where water-induced variability fundamentally alters appearance or observability. Empirically, WCAP consistently improves segmentation, forecasting, and measurement utility in underwater and aquatic applications. A plausible implication is broader applicability in any domain where physical degradation or partial observability aligns with a parameterizable, data-driven adaptation mechanism.

Known limitations include reliance on comprehensive calibration or training data (especially in environments with unmodeled or highly time-varying water properties), and the need for integration with domain solvers (hydraulic, rendering, or physical simulators). Extensions under consideration include multi-scale warping, higher-dimensional metric adaptation, and explicit regularization for generalized or safety-critical deployment (Wu et al., 3 Feb 2026).