Automated sewer defect detection has advanced through deep learning, particularly supervised methods using CCTV images, but based on large annotated datasets. This paper proposes a semi-supervised learning (SSL) approach to reduce labeling demands. The method comprises self-supervised pre-training on unlabeled images using SwAV (Swapping Assignments between multiple Views) followed by fine-tuning for multi-label classification. Experiments on the Sewer-ML dataset demonstrate that the SSL approach, trained on only 35k labeled images, achieves an F1-score of 69.11%, and F2_CIW of 54.22%, surpassing the fully supervised baseline trained from scratch on 1.04 million images. Increasing the unlabeled pre-training data further enhances performance, while ImageNet initialization consistently outperforms training from scratch. Self-supervised learning also helps mitigate the effects of mislabeled data, which is observed to be present even in the Sewer-ML ground truth. Overall, self-supervised learning provides an accurate, scalable, and cost-effective alternative to fully supervised approaches, particularly in data-scarce or imperfectly labeled scenarios.

Flood hazard maps are essential for protection and emergency plans, yet their probabilistic application is constrained by the computational cost of numerical models. Deep learning surrogates can provide orders of magnitude faster predictions, but their use for uncertainty quantification in realistic settings and their ability to incorporate hydraulic structures remain largely unexplored. Studying deep learning surrogates for probabilistic flood maps is non-trivial because of the lack of reference ground-truth data that might lead to misleading confidence in predictions. Moreover, hydraulic structures are challenging to include due to their generally unidimensional nature. In this work, we investigate the use of deep learning surrogates for realistic, large-scale flood simulations in case studies with hydraulic structures under diverse boundary conditions. To this end, we employ the multi-scale shallow-water-equations graph neural network (mSWE-GNN) that enjoys transferability to different boundary conditions and locations and whose graph-based architecture allows to represent structures such as canals, underpasses, and elevated elements as inputs. To address the lack of reference ground-truth data, we further introduce the average relative mass error (ARME), a mass-conservation-based criterion that helps identify physically plausible simulations. We applied the model on dike ring 41 in the Netherlands, generating probabilistic flood maps that account for uncertainties in breach location and breach outflow hydrographs. The model was trained on 30 simulations, generated with Delft3D, and evaluated against unseen benchmark simulations from the Dutch national flood catalogue, achieving a critical success index (CSI) of 73.6 % while running 10 000 times faster than the numerical simulator. The proposed ARME is negatively correlated with the CSI, with a Spearman correlation coefficient of -0.7, making it a useful indicator of simulation plausibility when evaluating unseen case studies. We obtained probabilistic flood maps by running 10 000 different flooding scenarios on a computational mesh of 180 000 cells in approximately 10 h, with about half of the simulations classified as plausible based on the mass-conservation check. This framework offers a practical tool for rapid probabilistic flood hazard assessment and a way to prioritize detailed physical simulations, supporting more efficient and robust flood risk management.

Simulating urban drainage hydraulics is computationally demanding, limiting its application in tasks that require real-time or repeated simulations. Graph Neural Networks (GNNs) are promising metamodels, but the effect of their internal components and transferability potential remain underexplored. This study addresses these gaps through two main contributions: (1) a systematic evaluation of key architectural components, including graph layer type, processor depth, and prediction window with links to physical transport dynamics; and (2) transferability experiments across domains (across two distinct drainage networks) and tasks (from head to flow prediction). As case studies, we selected two combined sewer networks in The Netherlands that differ in their hydraulic dynamics. We find that metamodels with moderate depth and a ten-step prediction window achieve high accuracy (RMSE of 2–5 cm for hydraulic heads and 0.02 m³/s for flowrates). They also reach speed-ups of up to four orders of magnitude higher compared to the physics-based model, SWMM, when executing parallel simulations in GPU. Based on our two case studies, we find that pre-trained metamodels with full fine-tuning effectively adapt to a new task within the same domain, whereas cross-domain transfer requires appropriate normalization and fine-tuning. Furthermore, joint training on both case studies enables the metamodel to capture representations of both systems, suggesting potential for more general applicability. These findings demonstrate that metamodel architecture can reflect physical system behavior and offer practical guidance for building fast, accurate, and generalizable GNN-based metamodels—establishing a foundation for their use in applications such as uncertainty analysis, design optimization, and nowcasting.

The steady state of a water distribution system abides by the laws of mass and energy conservation. Hydraulic solvers, such as the one used by EPANET approach the simulation for a given topology with a Newton-Raphson algorithm. However, iterative approximation involves a matrix inversion which acts as a computational bottleneck and may significantly slow down the process. In this work, we propose to rethink the current approach for steady state estimation to leverage the recent advancements in Graphics Processing Unit (GPU) hardware. Modern GPUs enhance matrix multiplication and enable memory-efficient sparse matrix operations, allowing for massive parallelization. Such features are particularly beneficial for state estimation in infrastructure networks, which are characterized by sparse connectivity between system elements. To realize this approach and tap into the potential of GPU-enhanced parallelization, we reformulate the problem as a diffusion process on the edges of a graph. Edge-based diffusion is inherently related to conservation laws governing a water distribution system. Using a numerical approximation scheme, the diffusion leads to a state of the system that satisfies mass and energy conservation principles. Using existing benchmark water distribution systems, we show that the proposed method allows parallelizing thousands of hydraulic simulations simultaneously with very high accuracy.

Supervised deep learning methods have been widely employed to detect floating macroplastic litter (>5 mm) in (fresh)water bodies. However, few studies used them to quantify floating litter fluxes in rivers with wide cross-sections, that is important for pollution assessment. Additionally, commonly used supervised learning (SL) models rely on extensive labeled data, that is time-consuming and expensive to obtain. Moreover, regardless of the model type, current deep learning models for litter detection usually fail to correctly identify small litter items. To overcome these issues, we propose a semi-supervised learning (SSL)-based framework combined with Slicing Aided Hyper Inference (SAHI) for quantifying cross-sectional floating litter fluxes in rivers. The framework includes four steps: (a) collecting camera images of river surfaces from multiple locations across the river, (b) developing a robust litter detection model using SSL, (c) applying this model with SAHI to detect litter items in images, and (d) post-processing the detection results to quantify fluxes. The SSL method involves: (i) self-supervised pre-training of a ResNet50 on a large amount of unlabeled data, and (ii) supervised fine-tuning of a Faster R-CNN with the ResNet50 backbone on a limited amount of labeled data. We evaluated the in-domain detection performance of SSL models with varying pre-training epochs and pre-training dataset sizes, using images from waterways of The Netherlands, Indonesia and Vietnam, that were used for model pre-training and fine-tuning. Additionally, we assessed the zero-shot out-of-domain detection performance of SSL models and litter flux quantification performance of the proposed framework on a Vietnam case study, that was not used for model development. We benchmarked our results against the SL methods and human visual counting. The results show that SSL models benefit from longer pre-training time and larger pre-training dataset, achieving an in-domain F1-score increase of 0.2 and a zero-shot out-of-domain increase of up to 0.14, over baseline SL benchmarks. Furthermore, the SAHI method correctly identifies 45 additional small litter items (areas < 1,000 cm2), improving the F1-score by up to 0.19, compared to the results obtained without SAHI. The flux measurement results indicate that the SSL-based framework substantially underestimates fluxes by a factor of 3–4 compared to human measurements, due to missed detections of transparent litter items and items entrapped in water hyacinths. However, it estimates nearly twice the fluxes of the baseline SL-based framework, aligning more closely with human measurements. These findings highlight the potential of SSL-based framework to enhance litter flux measurement. Scaling it with broader datasets could significantly advance global-scale litter monitoring systems.

Current research on riverine macrolitter does not yet provide a theoretic framework on the dynamics behind its accumulation and distribution along riverbanks. In an attempt to better understand these dynamics a detailed field survey of three months was conducted in which location of macrolitter items within a single groyne field along the Waal riverbanks was tracked. The data provided insight into the daily changing patterns of spatial item distribution with respect to the waterline. Furthermore, the rates of item uptake and deposition were monitored and related to hydrologic fluctuations. Uptake was initiated by rising water levels and was generally higher when the water level increased faster. Deposition occurred continuously, despite hydrologic fluctuations. This caused the riverbank macrolitter budget to be positive during stable or dropping water levels and negative during rising water levels. Although the results show clear patterns an extended monitoring duration is required to fully understand the fate of plastic objects.

Researchers and practitioners have extensively utilized supervised Deep Learning methods to quantify floating litter in rivers and canals. These methods require the availability of large amount of labeled data for training. The labeling work is expensive and laborious, resulting in small open datasets available in the field compared to the comprehensive datasets for computer vision, e.g., ImageNet. Fine-tuning models pre-trained on these larger datasets helps improve litter detection performances and reduces data requirements. Yet, the effectiveness of using features learned from generic datasets is limited in large-scale monitoring, where automated detection must adapt across different locations, environmental conditions, and sensor settings. To address this issue, we propose a two-stage semi-supervised learning method to detect floating litter based on the Swapping Assignments between multiple Views of the same image (SwAV). SwAV is a self-supervised learning approach that learns the underlying feature representation from unlabeled data. In the first stage, we used SwAV to pre-train a ResNet50 backbone architecture on about 100k unlabeled images. In the second stage, we added new layers to the pre-trained ResNet50 to create a Faster R-CNN architecture, and fine-tuned it with a limited number of labeled images (≈1.8k images with 2.6k annotated litter items). We developed and validated our semi-supervised floating litter detection methodology for images collected in canals and waterways of Delft (the Netherlands) and Jakarta (Indonesia). We tested for out-of-domain generalization performances in a zero-shot fashion using additional data from Ho Chi Minh City (Vietnam), Amsterdam and Groningen (the Netherlands). We benchmarked our results against the same Faster R-CNN architecture trained via supervised learning alone by fine-tuning ImageNet pre-trained weights. The findings indicate that the semi-supervised learning method matches or surpasses the supervised learning benchmark when tested on new images from the same training locations. We measured better performances when little data (≈200 images with about 300 annotated litter items) is available for fine-tuning and with respect to reducing false positive predictions. More importantly, the proposed approach demonstrates clear superiority for generalization on the unseen locations, with improvements in average precision of up to 12.7%. We attribute this superior performance to the more effective high-level feature extraction from SwAV pre-training from relevant unlabeled images. Our findings highlight a promising direction to leverage semi-supervised learning for developing foundational models, which have revolutionized artificial intelligence applications in most fields. By scaling our proposed approach with more data and compute, we can make significant strides in monitoring to address the global challenge of litter pollution in water bodies.

Data-driven metamodels reproduce the input-output mapping of physics-based models while significantly reducing simulation times. Such techniques are widely used in the design, control, and optimization of water distribution systems. Recent research highlights the potential of metamodels based on Graph Neural Networks as they efficiently leverage graph-structured characteristics of water distribution systems. Furthermore, these metamodels possess inductive biases that facilitate generalization to unseen topologies. Transferable metamodels are particularly advantageous for problems that require an efficient evaluation of many alternative layouts or when training data is scarce. However, the transferability of metamodels based on GNNs remains limited, due to the lack of representation of physical processes that occur on edge level, i.e. pipes. To address this limitation, our work introduces Edge-Based Graph Neural Networks, which extend the set of inductive biases and represent link-level processes in more detail than traditional Graph Neural Networks. Such an architecture is theoretically related to the constraints of mass conservation at the junctions. To verify our approach, we test the suitability of the edge-based network to estimate pipe flowrates and nodal pressures emulating steady-state EPANET simulations. We first compare the effectiveness of the metamodels on several benchmark water distribution systems against Graph Neural Networks. Then, we explore transferability by evaluating the performance on unseen systems. For each configuration, we calculate model performance metrics, such as coefficient of determination and speed-up with respect to the original numerical model. Our results show that the proposed method captures the pipe-level physical processes more accurately than node-based models. When tested on unseen water networks with a similar distribution of demands, our model retains a good generalization performance with a coefficient of determination of up to 0.98 for flowrates and up to 0.95 for predicted heads. Further developments could include simultaneous derivation of pressures and flowrates.

The operation of water distribution systems is based on reliable knowledge about the steady state of the system. This involves sensors to measure flow, facilitating a comprehensive overview of the system’s performance. Given the costs associated with sensor installation and operation, it is important to be strategic with sensor allocation. Recently developed Gaussian Processes with topological kernels can efficiently model mass and energy conservative flows and provide uncertainty bounds. Our work proposes a novel method of state estimation and a greedy search algorithm for water flow meter placement based on the uncertainty bounds provided by a Gaussian Process.

Computational models for water resources often experience slow execution times, limiting their application. Metamodels, especially those based on machine learning, offer a promising alternative. Our research extends a prior Graph Neural Network (GNN) metamodel for the Storm Water Management Model (SWMM), which efficiently learns with less data and generalizes to new UDS sections via transfer learning. We extend the metamodel’s functioning by adding flowrate prediction, crucial for assessing water quality and flooding risks. Using an Encoder–Processor–Decoder architecture, the metamodel displays high accuracy on the simulated time series. Future work is aimed at incorporating more physical principles and testing further transferability.

Large Multimodal Models are emerging general AI models capable of processing and analyzing diverse data streams, including text, imagery, and sequential data. This paper explores the possibility of exploiting multimodality to develop more interpretable AI-based predictive tools for the water sector, with a first application for sewer defect detection from CCTV imagery. To this aim, we test the zero-shot generalization performance of three generalist large language-vision models for binary sewer defect detection on a subset of the SewerML dataset. We compared the LMMs against a state-of-the-art unimodal Deep Learning approach which has been trained and validated on >1 million SewerML images. Unsurprisingly, the chosen benchmark showcases the best performances, with an overall F1 Score of 0.80. Nonetheless, OpenAI GPT4-V demonstrates relatively good performances with an overall F1 Score of 0.61, displaying equal or better results than the benchmark for some defect classes. Furthermore, GPT4-V often provides text descriptions aligned with the provided prediction, accurately describing the rationale behind a certain decision. Similarly, GPT4-V displays interesting emerging behaviors for trustworthiness, such as refusing to classify images that are too blurred or unclear. Despite the significantly lower performance from the open-source models CogVLM and LLaVA, some preliminary successes suggest good potential for enhancement through fine-tuning, agentic workflows, or retrieval-augmented generation.