The logic behind the integrated approach is straightforward but demanding. In-situ measurements remain the gold standard for identifying what is actually floating in the water, but point samples cannot capture the highly complex spatial and temporal pathways that carry litter from urban and agricultural sources through rivers, along coasts and eventually into sinks such as sediments, vegetation or the open sea. Remote sensing, in turn, offers wide coverage but cannot chemically confirm what it sees. Numerical models can forecast transport, but only if they are fed realistic inputs about how much plastic enters the water, when, and with what physical properties. The Po Delta project was designed to make each tool compensate for the weaknesses of the others.

At the source end, the team installed a low-cost Raspberry Pi camera system on bridges along the Po River, capturing one image every two seconds, with a GoPro Hero5 alongside for quality comparison. From 18,848 images, 3,526 were labelled, containing 2,914 pieces of litter ranging from multicoloured fragments and wood to lids, sheets and Styrofoam. A Faster Region-based Convolutional Neural Network, pre-trained on the COCO dataset and fine-tuned with both a proxy dataset from earlier river-monitoring work and the new Po River images, was then tasked with detecting floating items automatically. Performance was evaluated using mean Average Precision at an Intersection-over-Union threshold of 0.5, averaged across ten cross-validation folds.

The results revealed both the promise and the fragility of automated river monitoring. The COCO pre-trained configuration reached a mean mAP50 of 0.60, outperforming a model trained from scratch at 0.56, and combining all available training data beat using the Po River dataset alone at 0.57. More striking were the diagnostic tests. On imagery containing floating organic matter such as leaves and twigs, detection performance collapsed to an mAP50 of 0.25, compared with 0.83 on scenes free of natural debris. Yet when the same high-organic scene was captured with a 300 dpi Nikon DSLR instead of the 96 dpi Raspberry Pi, performance rose to 0.46—suggesting that much of the apparent confusion with organic matter stems from limited ground sampling distance rather than the neural network itself. For system designers, the lesson is that improving camera resolution or mounting positions may pay off more than tweaking algorithms.

At the coastal sink end, a DJI Matrice 600 drone equipped with a 20-megapixel Zenmuse X5 camera and a MicaSense RedEdge-3 multispectral sensor flew over Po Delta beaches in October 2021, generating georeferenced orthomosaics that were analysed with the LITTERDRONE software originally developed as a European Commission Blue-Labs activity. The software detects and classifies beached objects larger than 2.5 centimetres based on colour, shape and background, with an expert validating outputs in a human-in-the-loop step. North Pila beach emerged as the hotspot, with 118 items detected including 41 plastic fragments of 2.5 to 50 centimetres, 28 drink bottles and 12 plastic caps. Validation against physical sampling showed how sensitive the technique is to flight altitude: at 25 metres over Barricata beach, only 1 of 49 plastic items found on the ground was detected in the imagery, whereas at 15 metres over North Pila the software recovered 115 of 182 items. Driftwood whose colour closely matched the sand also masked litter on some beaches.

Satellite observations extended the picture offshore. A task-requested WorldView-2 image of the delta, acquired on 24 March 2021 at 1.6-metre resolution, was processed with two complementary approaches: an object identification method fused with edge detection on the panchromatic band, and a spectral band anomaly algorithm that subtracts an averaged water background from Rayleigh-corrected reflectance and computes a near-infrared-minus-green proxy. Both flagged suspected fishing boats, windrows—linear aggregations of floating material—and fishing buoys, with the automated maps consistent with visual inspection of true-colour composites. In parallel, operational Sentinel-2 and Sentinel-3 data were used to map total suspended matter as a tracer of river plumes, the likely hotspots where plastics aggregate, and a customised web interface delivered near-real-time daily views to guide the fieldwork.

Ground truth came from two research campaigns in 2021. From 14 to 27 March, the coastal northern Adriatic from Venice to Ancona was surveyed aboard the research vessel G. Dallaporta across 12 transects and 85 stations, with Secchi disk readings showing water transparency increasing offshore from 1 to 19 metres, 79 CTD casts, and a 330-micrometre Manta net towed at 45 locations. A second campaign in October sampled four coastal sites and five river branches. Back in the laboratory, putative microplastics larger than 300 micrometres were sorted under a stereomicroscope and identified by Attenuated Total Reflectance Fourier Transform Infrared spectroscopy against a 128-record polymer library. Polyethylene, polypropylene and polystyrene dominated, consistent with other regional studies, and most particles were white, transparent or blue fragments—likely weathered pieces of single-use packaging.

The transect data revealed sharp spatial structure. Along transect T2 near the Adige River mouth, nearshore waters were essentially clean, but concentrations spiked to 1.98 pieces per cubic metre before declining seaward to 0.02. Closer to the Po Delta mouth, transect T4 showed far higher values ranging from 0.02 to 54.26 pieces per cubic metre, following a sinusoidal pattern with a maximum halfway along the transect. Polymer diversity tracked abundance, with polyethylene present at more than 50 percent at five of six stations. The offshore maxima in both transects align with the Po River plume forming an accumulation front where freshwater meets seawater, while numerical trajectories indicate that particles near the Adige are transported southwards, leaving nearshore waters there largely plastic-free.

The modelling component used HYDROMOD-3D, a hydrodynamic coastal model running on 100-metre horizontal grids, with tracer modules modified to account for particle size, wind exposure above the water surface, and sinking versus non-sinking behaviour. Simulations driven by 2016 discharge, wind and water-level data tested two characteristic wind regimes of the northern Adriatic: the north-easterly Bora and the south-easterly Scirocco. The contrast was dramatic. Under Bora conditions, 200-millimetre particles travelled in a narrow band straight along the coast, whereas Scirocco winds drove them into a circular gyre generated by a topographically induced counter-current. Larger 500-millimetre objects were more strongly affected by wind drag. The simulations produced probabilistic accumulation zones rather than exact landing points, but those forecasts could be checked by drone surveys of recurrent beaching locations—closing the loop between prediction and observation.

The authors are candid about the limitations. Two field campaigns provide only a narrow spatio-temporal snapshot compared with multiyear surveys of Italian waters, and the modular system was assembled to assess technology readiness rather than rigorously cross-validated across all platforms, which they identify as the essential next step. Still, the proof of concept carries real weight for policy: physiochemical descriptors of the litter feed directly into instruments such as Italy’s Salvamare Law, the Marine Strategy Framework Directive and the EU Single-Use Plastics Directive, and model forecasts of accumulation hotspots could optimise clean-up operations at sea and along shorelines. If the remaining gaps—harmonised protocols, balanced training datasets, and better fusion of high-resolution imagery with operational satellite data—can be closed, the Po Delta experiment suggests that an affordable, multi-tool observing system capable of tracking plastic from source to sink is no longer a distant ambition but an engineering problem within reach.

Subject of Research: Integrated source-to-sink monitoring of plastic litter leakage in the Po Delta using remote sensing, machine learning, field sampling and hydrodynamic modelling

Article Title: A case study on the Po Delta in Italy on advancing the synergy of monitoring strategies for leakage litter from source-to-sink

Article References: Franke, J., Garaba, S. P., Brand, A. K., Duwe, K., Davison, S., Falcieri, F. M., Laforsch, C., Löder, M. G., López-Samaniego, E., Mantas, V., & Pérez-Gómez, J. P. (2026). A case study on the Po Delta in Italy on advancing the synergy of monitoring strategies for leakage litter from source-to-sink. Environmental Challenges, 25, Article 101651. https://doi.org/10.1016/j.envc.2026.101651

Image Credits: AI Generated

DOI: 10.1016/j.envc.2026.101651

Keywords: plastic pollution, Po Delta, marine litter, remote sensing, machine learning, hydrodynamic modelling, microplastics, Adriatic Sea, drone monitoring, satellite imagery, riverine plastic, Environmental Challenges