HydroVision: AI for Water Systems Forecasting

• HydroVision is a family of AI systems that fuse heterogeneous water data with physical priors to forecast and monitor rivers, lakes, and urban pipelines.
• It employs multimodal approaches—using LiDAR for terrain-guided hydrometric forecasting and RGB imagery for water-quality parameter estimation—to enhance prediction accuracy.
• The framework extends to urban water infrastructure by integrating sparse sensing, hydraulic simulation, and graph neural networks for network-wide monitoring and anomaly detection.

HydroVision is a name used in recent water-informatics literature for AI systems that infer hydrologic state from heterogeneous observations. In the cited corpus, the term refers directly to a LiDAR-guided framework for multivariate, multi-station hydrometric forecasting and to a deep-learning system that estimates optically active surface-water quality parameters from RGB imagery, while a broader HydroVision framing is also used for urban water-system intelligence based on sparse sensing, hydraulic simulation, and graph neural networks. This suggests that HydroVision is best understood less as a single algorithm than as a family of data-driven, physics-aware approaches for forecasting, monitoring, and situational assessment across rivers, lakes, and underground pipeline networks (Roudbari et al., 2024, Deshmukh et al., 2 Sep 2025, Guo et al., 8 Oct 2025).

1. Scope of the term in current literature

The term has been applied to distinct, though methodologically related, water-domain problems. One line of work studies daily hydrometric forecasting over interconnected stations using terrain-derived graphs from LiDAR and recurrent graph learning. Another line studies RGB-scene estimation of optically active water-quality variables such as Chlorophyll-$\alpha$, CDOM, Suspended Sediments, and Turbidity. A broader HydroVision framing also appears in urban wastewater monitoring, where sparse remote sensing, hydraulic simulation, and directed edge-aware graph learning are combined to infer network-wide states from limited direct instrumentation (Roudbari et al., 2024, Deshmukh et al., 2 Sep 2025, Guo et al., 8 Oct 2025).

Variant	Primary inputs	Primary task
HydroVision (Roudbari et al., 2024)	Daily water levels from six stations; LiDAR DTM at 1 m resolution	Multi-station water-level forecasting up to 12 days ahead
HydroVision (Deshmukh et al., 2 Sep 2025)	RGB images from HIVIS; NWIS water-quality records	Estimation of six optically active parameters
HydroVision framing (Guo et al., 8 Oct 2025)	Sparse SmartCover sensing; PCSWMM simulation; pipe attributes	Network-wide hydraulic forecasting and deviation-based anomaly support

A recurring property across these uses is the fusion of direct observations with structured priors that encode hydrologic or hydraulic connectivity. In the LiDAR-guided forecasting system, terrain elevation supplies the spatial prior. In the RGB water-quality system, segmentation and transfer learning convert public image streams into water-quality surrogates. In the wastewater-network framing, topology, pipe attributes, and calibrated simulation compensate for sparse sensing. This suggests that HydroVision denotes a broader methodological tendency toward physically informed AI for partially observed water systems rather than a single canonical software stack.

2. Terrain-guided hydrometric forecasting

In "HydroVision: LiDAR-Guided Hydrometric Prediction with Vision Transformers and Hybrid Graph Learning" (Roudbari et al., 2024), HydroVision is formulated as a multivariate, multi-station hydrometric forecasting model. The data consist of daily water-level measurements from six stations in Quebec over 1981–2021 and LiDAR Digital Terrain Model data from the Quebec MRNF at 1 m spatial resolution. The forecasting task uses 12 historical time steps to predict up to 12 days ahead, with reported horizons of 3, 6, 9, and 12 days. The paper formalizes two functions: $f:L \rightarrow S$, which extracts spatial structure from LiDAR, and $g:(S,T) \rightarrow \hat{T}$, which maps that structure together with hydrometric time series $T \in \mathbb{R}^{n \times m}$ to future predictions.

The architecture has two coupled components. LiDAR elevation is encoded with a Vision Transformer by dividing the terrain image into non-overlapping $16 \times 16$ patches, embedding them, adding sinusoidal positional encodings, and passing the sequence through a transformer encoder to produce a terrain-derived representation later used as a static graph. In parallel, temporal sequences are processed with Graph Convolutional Recurrent Networks. The graph structure is hybrid: a static terrain-derived graph $A_{\text{elevation}}$ is combined with an adaptive graph learned from trainable node embeddings,

$$A_{\text{adaptive}} = \text{softmax}(\text{ReLU}(E1 \cdot E2^T)),$$

and the fusion is

$$\hat{A}= \alpha A_{\text{adaptive}} + (1 - \alpha) A_{\text{elevation}}.$$

Graph convolution then uses

$$Z^{(l)} = \sigma (\tilde{A}Z^{(l-1)}W^{(l)} + b^l),$$

while temporal dependence is modeled by GRU gates in which the affine maps are replaced by graph-convolutional operators: $$r^{t} = \sigma (\mathcal{F}(\tilde{A}, [X^t, h^{t-1}]) + c_r),$$

$f:L \rightarrow S$0

$f:L \rightarrow S$1

$f:L \rightarrow S$2

Training uses a 70% / 10% / 20% split, batch size 64, Adam, MAE loss, maximum 300 epochs, early stopping after 20 validation failures, learning-rate decay ratio 0.1, eight attention heads, PyTorch 1.7.1, and an NVIDIA GeForce RTX 2080 Ti with 11 GB memory. HydroVision is compared against AGCRN, Informer, DCGCN, STtransformer, GTS, STAWnet, and MTGNN. It achieves the best MAE and RMSE at all reported horizons: MAE $f:L \rightarrow S$3 and RMSE $f:L \rightarrow S$4 for 3, 6, 9, and 12 days, respectively. The abstract summarizes this as an average 10% reduction in prediction error across all days, with greater improvements for longer forecasting horizons. An ablation comparing the full hybrid graph against adaptive-only graph learning reports MAE $f:L \rightarrow S$5 versus $f:L \rightarrow S$6, indicating that terrain-informed static structure improves longer-horizon forecasting.

A notable limitation is that several critical implementation details are omitted. The exact construction of $f:L \rightarrow S$7 from transformer output is not specified; the value or learning strategy for $f:L \rightarrow S$8 is not given; the dynamic graph is described in prose as adapting to temporal changes, but the provided mathematics define a learned adaptive adjacency rather than a time-indexed $f:L \rightarrow S$9. The experiments are also limited to six stations in one Quebec region. These omissions do not negate the reported forecasting gains, but they constrain exact reproducibility and make the method best interpreted as a terrain-aware hybrid-graph design pattern rather than a fully specified canonical recipe.

3. RGB-based prediction of optically active water-quality parameters

In "HydroVision: Predicting Optically Active Parameters in Surface Water Using Computer Vision" (Deshmukh et al., 2 Sep 2025), HydroVision denotes a scene-classification/regression framework that estimates six optically active parameters from standard RGB images of natural surface-water scenes: Chlorophyll-$g:(S,T) \rightarrow \hat{T}$0, Chlorophylls, Colored Dissolved Organic Matter (CDOM), Phycocyanins, Suspended Sediments, and Turbidity. The image source is the United States Geological Survey Hydrologic Imagery Visualization and Information System, while labels come from the National Water Information System. The study identifies 111 monitoring sites that measure at least one relevant parameter and collects over 500,000 seasonally varied images between February 2022 and November 2024. After preprocessing and filtering, the parameter-specific labeled datasets contain 28,061 images for Chlorophyll-$g:(S,T) \rightarrow \hat{T}$1, 12,681 for Chlorophylls, 17,328 for CDOM, 18,785 for Phycocyanins, 21,535 for Suspended Sediments, and 480,125 for Turbidity.

The preprocessing pipeline is central to the system. A pre-trained U-Net-based segmentation model isolates the water region by masking non-water pixels to RGB zero. Nighttime images are removed using astronomical sunrise and sunset calculations from the astral library, and only images with at least 20% segmented water pixels are retained. Image metadata and NWIS measurements are then linked by site code, state, and day/night status, with nearest-in-time alignment implemented through pandas.merge_asof. The segmentation model is evaluated indirectly against surrogate reference masks produced by classical methods, with retained segmentation quality reported as IoU mean $g:(S,T) \rightarrow \hat{T}$2, Dice mean $g:(S,T) \rightarrow \hat{T}$3, Precision mean $g:(S,T) \rightarrow \hat{T}$4, Recall mean $g:(S,T) \rightarrow \hat{T}$5, Accuracy mean $g:(S,T) \rightarrow \hat{T}$6, and Specificity mean $g:(S,T) \rightarrow \hat{T}$7. These values indicate high water-pixel recall but also notable inclusion of non-water regions such as sky-like or reflective areas.

The predictive stage trains six independent regression models rather than a single multi-task model. Four CNN backbones—VGG16, ResNet50, MobileNetV2, and DenseNet121—are compared with a Vision Transformer. Inputs are resized to $g:(S,T) \rightarrow \hat{T}$8; CNNs use ImageNet-initialized transfer learning with a regression head consisting of Global Average Pooling, a dense layer with either 512 or 1024 neurons, ReLU, L2 regularization, and a linear scalar output, while the last 10 layers are unfrozen for fine-tuning. The ViT uses vit-base-patch16-224, the [CLS] token, dropout, a dense ReLU layer, and a linear output, with the last three transformer blocks unfrozen. For the five smaller tasks, Bayesian optimization searches dropout rate, L2 strength, learning rate, dense width, and optimizer; turbidity uses a distinct large-data training setup with fixed hyperparameters, TensorFlow XLA compilation, batch size 8, cosine decay, and no horizontal flipping.

The best validation result is DenseNet121 on CDOM with $g:(S,T) \rightarrow \hat{T}$9, MSE $T \in \mathbb{R}^{n \times m}$0, MAE $T \in \mathbb{R}^{n \times m}$1, RMSE $T \in \mathbb{R}^{n \times m}$2, and sMAPE $T \in \mathbb{R}^{n \times m}$3. DenseNet121 is also best among the CNNs on Chlorophyll-$T \in \mathbb{R}^{n \times m}$4 ($T \in \mathbb{R}^{n \times m}$5), Chlorophylls ($T \in \mathbb{R}^{n \times m}$6), Phycocyanin ($T \in \mathbb{R}^{n \times m}$7), and Turbidity ($T \in \mathbb{R}^{n \times m}$8). A later temporal validation on data from November 5, 2024 to February 17, 2025 is more conservative: the strongest result remains CDOM, where ResNet50 reaches Pearson $T \in \mathbb{R}^{n \times m}$9 and concordance correlation coefficient $16 \times 16$0, while Chlorophyll-$16 \times 16$1 is only moderately correlated and Chlorophylls, Phycocyanin, Suspended Sediments, and Turbidity show weak transfer to unseen future data. The paper also reports a classification-style analysis using Low/Medium/High quantile bins; CDOM shows clear diagonal dominance, whereas the other variables exhibit heavier confusion and a tendency to underpredict higher classes.

The publication therefore supports two simultaneous conclusions. First, ordinary RGB imagery contains enough information to estimate at least some optically active parameters in uncontrolled outdoor scenes, with CDOM as the clearest success case. Second, HydroVision is not yet a uniform solution across all six targets. The authors identify limitations in well-lit-scene dependence, imperfect segmentation, site-specific bias, asynchronous image–measurement matching, class imbalance, and the limited spectral expressiveness of RGB for parameters whose signatures are weak or ambiguous in the visible range.

4. HydroVision as a broader framework for urban water-system intelligence

A broader HydroVision framing appears in "AI-Driven Forecasting and Monitoring of Urban Water System" (Guo et al., 8 Oct 2025), whose details explicitly describe the work as highly relevant to a HydroVision-style water-system intelligence platform. The problem is network-wide hydraulic inference in urban underground water and wastewater infrastructure when only a small fraction of buried assets can be instrumented directly. The proposed system uses three integrated stages: sparse remote sensing at selected manholes, augmentation with hydraulic simulation for full-network coverage, and a dedicated graph neural network called HydroNet for forecasting normal hydraulic behavior and supporting anomaly detection through deviations from expected patterns.

The sensing layer is deliberately sparse. SmartCover sensors are mounted inside manholes rather than in pipes, measure only flow velocity / flow rate and water depth, transmit data via antenna, and are battery-powered for roughly two years of operation. In the study network, only 5 manholes are directly instrumented, the cadence is every 10 minutes, and the observation window runs from October 1, 2023 to January 31, 2024. The campus wastewater network contains 22 vitrified clay pipes and 23 nodes total, consisting of 22 manholes plus 1 outlet. The calibrated PCSWMM hydraulic model supplies time series for the remaining 18 uninstrumented nodes at the same temporal resolution, creating a digital-twin-assisted supervisory dataset.

HydroNet represents the network as a directed graph $16 \times 16$2, with node features consisting of time-varying water depth and flow rate and edge features consisting of nine static pipe attributes: Length, Roughness, Diameter (Geom1), Slope, GIS Length, Max Flow, Max Velocity, Max / Full Flow, and Max / Full Depth. The architecture comprises two ST-MPNN blocks followed by an output layer. Each ST-MPNN block includes a temporal module using gated temporal convolutions over a lookback window $16 \times 16$3, corresponding to two hours at 10-minute sampling, and a spatial directed message-passing module that embeds edge attributes. The temporal encoding is

$16 \times 16$4

and message passing along edge $16 \times 16$5 uses

$16 \times 16$6

followed by the update

$16 \times 16$7

The forecasting protocol uses 12 steps of history to predict the next 12 steps for both depth and flow rate.

Experiments use a 7:1:2 train/validation/test split, early stopping on validation loss, and NVIDIA A6000 GPUs with 48 GB VRAM. Baselines are CaST, GMAN, ST-SSL, STG-MAMBA, and STGCN. HydroNet reports depth forecasting performance of MAE $16 \times 16$8 ft, RMSE $16 \times 16$9, MAPE $A_{\text{elevation}}$0, and flow forecasting performance of MAE $A_{\text{elevation}}$1 cfs, RMSE $A_{\text{elevation}}$2, MAPE $A_{\text{elevation}}$3. The strongest baseline on depth MAE is STGCN at 0.0123, and the strongest baseline on flow MAE is STGCN at 0.0066, so the reported gains are about 31% and roughly 42%, respectively. The paper also notes strong physical correlations among edge attributes, such as slope and diameter $A_{\text{elevation}}$4, max flow and slope $A_{\text{elevation}}$5, and max flow and velocity $A_{\text{elevation}}$6, supporting the value of edge-aware propagation.

This line of work matters for HydroVision because it extends the term from river or lake sensing toward subterranean urban infrastructure. At the same time, the paper is careful about what it does not define. Although anomaly detection is repeatedly described as being enabled by deviation from predicted normal hydraulic behavior, no explicit anomaly score, threshold, or leak-classification loss is provided. Similarly, hydraulic simulation enters as a calibrated data source rather than as an explicit differentiable physics layer. The result is best read as an architectural template for sparse observability in infrastructure networks rather than a fully formalized anomaly-detection framework.

5. Shared methodological themes and adjacent directions

Taken together, the HydroVision literature exhibits several recurring design patterns. First, connectivity is rarely taken as fixed and fully observed. In the hydrometric forecasting study, terrain elevation is used to derive a static graph and is fused with an adaptive learned graph. In the wastewater-network study, directionality is imposed by downstream hydraulic flow and edge attributes modulate message passing. In the RGB water-quality study, connectivity is less explicit, but segmentation and site-linked image–measurement alignment serve an analogous role by restricting the model to the physically relevant water scene. This suggests a common HydroVision principle: observability is partial, so spatial structure must be inferred or constrained by terrain, topology, or interface geometry.

Second, HydroVision research repeatedly fuses AI with physically meaningful priors rather than relying on purely generic sequence learning. Terrain elevation, pipe slope and diameter, hydraulic simulation, sunrise–sunset filtering, and surface-water segmentation all serve as physically grounded constraints. A related hydrologic forecasting direction appears in "HydroFusion-LMF: Semi-Supervised Multi-Network Fusion with Large-Model Adaptation for Long-Term Daily Runoff Forecasting" (Fan et al., 4 Oct 2025), which is not named HydroVision but extends the same design logic. It decomposes runoff as

$A_{\text{elevation}}$7

routes residuals through heterogeneous experts, fuses them by a hydrologic context-aware gate, and augments supervision with masked reconstruction, contrastive alignment, augmentation consistency, and pseudo-labeling. On its reported benchmark, HydroFusion-LMF attains test MSE $A_{\text{elevation}}$8 and MAE $A_{\text{elevation}}$9, improving the strongest baseline DLinear by 10.2% and 10.3%, respectively. This suggests that regime adaptivity and label-efficient learning are likely to become important adjacent directions for HydroVision-style systems as they move from proof-of-concept forecasting toward broader operational deployment.

Third, HydroVision research increasingly crosses sensing modalities. The 2024 HydroVision model combines LiDAR with station time series; the 2025 HydroVision model combines RGB camera imagery with water-quality telemetry; the HydroVision framing in urban infrastructure combines sparse remote sensing with hydraulic simulation and asset metadata. A plausible implication is that future HydroVision systems will be multimodal by design, with terrain, imagery, telemetry, and digital twins acting as complementary rather than competing information sources.

6. Limitations, misconceptions, and open questions

A common misconception is that HydroVision refers to a single standardized platform. The literature does not support that reading. One HydroVision system forecasts water levels across six Quebec stations from LiDAR and hydrometric time series; another predicts optically active parameters from RGB images at 111 USGS sites; a broader HydroVision framing is used for sparse-sensor urban wastewater intelligence. The shared theme is methodological, not institutional: AI-mediated hydrologic or hydraulic inference under partial observability.

The limitations are correspondingly domain-specific. In the LiDAR-guided forecasting model, the exact conversion from transformer output to the terrain-derived adjacency $$A_{\text{adaptive}} = \text{softmax}(\text{ReLU}(E1 \cdot E2^T)),$$0, the value of $$A_{\text{adaptive}} = \text{softmax}(\text{ReLU}(E1 \cdot E2^T)),$$1, and the station-to-LiDAR alignment procedure are not specified, and the experiments are limited to six stations in one Quebec region (Roudbari et al., 2024). In the RGB water-quality model, segmentation sometimes misclassifies sky as water, nighttime scenes are discarded, well-lit imagery is required, and the strongest robust evidence is concentrated in CDOM rather than all six targets (Deshmukh et al., 2 Sep 2025). In the wastewater-network framing, performance is demonstrated only on a campus-scale network; the approach assumes accurate pipeline metadata and a calibrated PCSWMM model, while broader generalization to larger urban systems, incomplete GIS, or extreme weather is left unresolved (Guo et al., 8 Oct 2025).

Open problems follow directly from these limitations. More explicit uncertainty estimation, stronger temporal generalization, richer multimodal fusion, principled anomaly scoring, and reproducible definitions of learned spatial structure remain incompletely solved. The corpus also points to a scaling question: whether HydroVision systems can move from localized proof-of-concept settings—six stations, 111 image-linked sites, 23-node campus sewer networks—to operational regional or city-scale deployments without losing physical fidelity or interpretability. The literature to date suggests that this is feasible in principle, but not yet established as a mature, standardized practice.