A geometric and deep learning reproducible pipeline for monitoring floating anthropogenic debris in urban rivers using in situ cameras Journal: Elsevier
Gauthier Grimmer, Romain Wenger, Clément Flint, Germain Forestier, Gilles Rixhon, Valentin Chardon · 2025
Read A geometric and deep learning reproducible pipeline for monitoring floating anthropogenic debris in urban rivers using in situ cameras Journal: Elsevier by Gauthier Grimmer, Romain Wenger, Clément Flint, Germain Forestier, Gilles Rixhon, Valentin Chardon (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-23798.txt.
Introduction
Gauthier Grimmer and five co-authors build a reproducible pipeline for watching floating trash in city rivers with fixed cameras. Plastic and other human-made debris drifting downstream harms wildlife, water quality, navigation, and recreation. Counting it continuously by hand is impractical, so the authors pair deep learning detection with a geometric model that estimates each object's real size from a flat image.
The study compares deep learning models for accuracy and speed under messy real conditions, and probes pitfalls like data leakage between training and test images. Two lessons stand out: including negative images — scenes without debris — matters greatly, and guarding against temporal leakage shapes honest scores. The geometric size estimates, corrected by regression, proved feasible. The authors frame the result as a blueprint for cheap, automated river monitoring that cities could actually deploy and maintain.
Catalog record
| CREATOR | Gauthier Grimmer, Romain Wenger, Clément Flint, Germain Forestier, Gilles Rixhon, Valentin Chardon |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 86 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-23798
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A geometric and deep learning reproducible pipeline for monitoring floating anthropogenic debris in urban rivers using in situ cameras Journal: Elsevier Gauthier Grimmer, Romain Wenger, Clément Flint, Germain Forestier, Gilles Rixhon, Valentin Chardon Abstract The proliferation of floating anthropogenic debris in rivers has emerged as a pressing environmental concern, exerting a detrimental influence on biodiversity, water quality, and human activities such as navigation and recreation. The present study proposes a novel methodological framework for the monitoring the aforementioned waste, utilising fixed, in-situ cameras. This study provides two key contributions: (i) the continuous quantification and monitoring of floating debris using deep learning and (ii) the identification of the most suitable deep learning model in terms of accuracy and inference speed under complex environmental conditions. These models are tested in a range of environmental conditions and learning configurations, including experiments on biases related to data leakage. Furthermore, a geometric model is implemented to estimate the actual size of detected objects from a 2D image. This model takes advantage of both intrinsic and extrinsic characteristics of the camera. The findings of this study underscore the significance of the dataset constitution protocol, particularly with respect to the integration of negative images and the consideration of temporal leakage. In conclusion, the feasibility of metric object estimation using projective geometry coupled with regression corrections is demonstrated. This approach paves the way for the development of robust, low-cost, automated monitoring systems for urban aquatic environments. 1 Introduction Several million tonnes (1.15–2.41 Mt) of anthropogenic debris are transported annually to the oceans via hydrosystems ( Lebreton et al. 2017 ; González-Fernández et al. 2021 ; Palmas et al. 2022 ) . Their presence in rivers results from both anthropogenic activities—such as poor waste management, illegal dumping, waste-water treatment or urbanisation ( Van Emmerik and Schwarz 2020 ; Schöneich-Argent et al. 2020 ; De Vries et al. 2021 ) —and natural processes like aerial deflation and rainfall induced surface runoff ( Van Emmerik and Schwarz 2020 ) . This accumulation can be particularly significant downstream of urban areas, due to high population density and socio-economic activities significantly contribute to various debris inputs into hydrosystems ( Rech et al. 2014 ; Van Emmerik and Schwarz 2020 ; Gómez et al. 2022 ) . Anthropogenic debris—including plastics, metals, and glass ( Hanke et al. 2013 ; Suteja et al. 2025 ) —represent a major environmental concern, especially in the context of urban river management. Plastics alone accounts for approximately 80% of these debris ( Rech et al. 2014 ; González-Fernández et al. 2021 ) , making it a pervasive pollutant affecting all ecosystems. In rivers, macroplastics ( > 5 mm; González-Fernández et al. 2021 ) degrade into microplastics ( < 5 mm; Skalska et al. 2020 ) through biochemical (i.e. photodegradation, oxidation, biodegradation) and mechanical (i.e. movement of water and sediment, etc.) fragmentation ( Liro et al. 2023 ) . Aligned with the One Health approach ( Prata et al. 2021 ; Bois and Staentzel 2025 ) , macroplastics harm wildlife (i.e. entanglement, intestinal blockage; Waring et al. 2018 ; Alabi et al. 2019 ) and promote bank overflow and flooding eventually resulting in river morphology alterations ( Van Emmerik and Schwarz 2020 ) . The average concentration of 6.7 particles of microplastics per cubic metre of water ( Ghiglione et al. 2023 ; Landebrit et al. 2024 ) represents a growing environmental and public health concern. Furthermore, microplastics interfere with natural ecosystem functions (i.e. nutrient flow, species reproduction, etc.; Prata et al. 2021 ) and are easily found in the trophic chain through the consumption of fish ( Neves et al. 2015 ) or other products containing plastics ( Kumar et al. 2022 ) . Plastics directly impact health due to additives like Bisphenol A and Persistent Organic Pollutants (POPs), associated with neurological and reproductive issues ( Alabi et al. 2019 ) . This underlines the need to monitor riverine debris to reduce macro- and microplastics input in hydrosystems. Riverine anthropogenic debris and their spatial dynamics have been studied so far through direct monitoring methods. The latter include : (i) GPS ( Tramoy et al. 2019 ) or RFID trackers ( Chardon et al. 2025 ) , (ii) debris quantification via visual observation ( Castro-Jiménez et al. 2019 ) , (iii) net sampling from bridges or boats ( Hurley et al. 2023 ) , or (iv) passing sampling (i.e. use of existing infrastructure; Van Emmerik and Schwarz 2020 ). Although these methods are efficient at the local scale, they fail, at the catchment scale, to continuously monitor debris flux and capture anthropogenic debris heterogeneity ( Hurley et al. 2023 ) . These inherent limitations call for new automated approaches enabling continuous monitoring and the production of reproducible results. In this respect, camera-acquired digital images has been recently used to tackles issues related to anthropogenic debris ( Van Lieshout et al. 2020 ; Donal et al. 2023 ; Zhang et al. 2023 ) because they allow continuous recording and thereby capturing data variability (i.e. debris diversity, water turbidity, weather conditions; Jia et al. 2023 ; Tata et al. 2021 ). In connection with riverine debris monitoring, deep learning (DL) methods are inscreasingly widespread. Van Lieshout et al. 2020 and Kataoka et al. 2024 used cameras mounted on bridges characterised by vertical views. Maharjan et al. 2022 use cameras onboard UAVs. Images captured by these methods are then processed with convolutional neural networks (CNNs)—such as Faster R-CNN ( Van Lieshout et al. 2020 ) , YOLOv5-s ( Armitage et al. 2022 ) or YOLOv8 ( Kataoka et al. 2024 ) —achieving high detection performance on riverine debris. However, both approaches are either intrisically linked to a fixed spot (i.e. cameras on bridges) or limited in time (i.e. punctual monitoring by cameras mounted on UAVs). On the one hand, becoming independent of man-made structures is thus a prerequisite when it comes to study anthropogenic debris fluxes at the catchment scale. On the other hand, the flexible implementation of cameras along streams results in issues related to heterogeneous observation conditions which can accentuate biases and negatively impact the robustness of DL models. This study thus first seeks to examine the object detection capacity of CNN models in complex recording conditions that are river environments. The second challenge after debris detection is to estimate its dimensions, taking into account the geometric distortions in the image caused by the environment hosting the cameras. The threefold aims of this study are: (i) continuous quantification and monitoring of floating debris using DL models, (ii) identifying the most suitable DL model in terms of accuracy and inference speed under complex environmental conditions, and (iii) proposing a generalisable solution by introducing a geometric modeling approach to the camera’s field of view in order to estimate the dimensions of real-world objects from image-based detections. 2 Methodology Figure 1 : Employed methodology : (1) Acquired data is manually annotated to constitute initial dataset. (2) The latter is pre-processed so that (3) several models can then be trained to detect debris in rivers. (4) Dimensions of detected objects are predicted and then (5) corrected to reduce errors 2.1 Study area The natural reserve of the Robertsau and Wantzenau forest massif is located north