Municipal solid waste (MSW) is a growing global challenge, with 3 billion people lacking access to safe disposal and global waste generation expected to rise from 2.01 to 3.40 billion tonnes by 2050. Poor waste management, especially open dumping, contributes to environmental degradation, disease transmission, and emerging health risks from microplastics. Despite the UN’s call to end unsafe waste practices, effective and inclusive management systems remain critically underdeveloped—particularly in Sub-Saharan Africa, where less than 50% of MSW is collected and about 70% is openly dumped. In East Africa, Tanzania—and especially its commercial capital, Dar es Salaam—stands out as a hotspot for mismanaged plastic waste. The city generates approximately 4,600 tonnes of waste daily, yet less than 40% is collected, with the remainder being illegaly burned or dumped to later accumulate in streets, drains, and rivers, including the heavily polluted Msimbazi River, a major contributor of plastic waste to the Indian Ocean.

To address the lack of reliable data on urban waste accumulation, we present a city-scale, open-data-based approach to map visible MSW in Dar es Salaam using deep learning and street-view imagery (SVI). In close collaboration with OpenMap Development Tanzania and HeiGIT, we developed an open-source end-to-end workflow that involves collecting, annotating, and processing 360-degree SVI across 41 wards, training a deep learning classification model to detect solid waste, and generating a high-resolution MSW pollution indicator. This approach provides actionable spatial insights to optimize urban cleaning strategies, reduce health risks linked to water contamination, and support evidence-based planning for more sustainable waste management.

The Street-view imagery was collected between throughout 2024 by OMDTZ using a Bajaj tricycle equipped with a GoPro Max 360° camera. This data collection effort covered major roads across Dar es Salaam. The imagery was made publicly available on Mapillary to enable open access and collaborative research.

To identify visible waste in street-view images, a two-step image annotation process was implemented using the open-source tool MapSwipe4Web for. Additionaly to manual labeling, we opted for Crowdsourced Pre-labeling: The MapSwipe community reviewed street-level panoramic images and flagged those containing visible waste. All positively labeled images were manually reviewed by the authors to ensure quality. Further manual labeling was conducted, particularly for underrepresented classes.

Images were annotated into the following three main classes: Waste Piles: Defined as accumulations of waste with horizontal and vertical extent, often indicative of illegal dumping or unmanaged MSW. Trash Bags: Properly packaged waste (typically in plastic or paper bags) that appears to be ready for collection. No Trash: Images with no visible waste.

To ensure spatial representativeness and reduce redundancy, images were sampled approximately every 10 meters along one-third of the road segments in each ward. In total, over 40,000 360° street-view panoramas were annotated. Due to the computational and geometric challenges associated with high-resolution equirectangular imagery, each panorama was segmented into 12 overlapping patches. This approach reduced computational load, minimized distortion effects, and maintained spatial context for classification. Further labelling on the image patches resulted in a final training dataset size of 7378 (with augmented 29605) waste images and 7640 (27410) no waste images for the trashbag model and 1818 (7325) waste images and 1920 (6965) no waste images for the waste model. A classification-based approach utilizing the YOLOv11 architecture was employed to detect visible waste, as it significantly reduces annotation effort by eliminating the need for bounding box labeling. This method was used because conventional object detection offers limited practical benefits and introduces unnecessary complexity to the annotation workflow.. The resulting models demonstrated strong performance, achieving precision scores of 0.87 for trash bags and 0.98 for waste piles, and recall scores of 0.86 and 0.98, respectively.

Nevertheless, some limitation need to be considered. For instance, the model occasionally struggled to distinguish between visually similar objects like trash bags and plastic buckets or plastic tarpaulins. For waste piles, materials like rubble or patterned floor mosaics sometimes led to misclassifications..

We deployed the trained models across the entire city, processing a total of 500,000 panoramic images covering 1300 kilometers of road network. The results were aggregated onto a hexagonal grid to produce a spatially explicit indicator of visible waste pollution. This analysis was further complemented with UAV-derived observations of waste piles along the Msimbazi River—a region largely inaccessible to ground-based street-view imaging—yielding a comprehensive waste pollution map for the entire study area.

Using this workflow we can show that there are several hotspots of waste pollution in Dar es Salaam mostly located in densly populated informal settlements, while the wards along the coastline show only little waste pollution. Combining the created data with data from OSM, it was recognised that there is a lot of open waste in the immediate vicinity of the drainage system, which can lead to a higher risk of flooding due to clogging.

Our results show the value of street view imagery for open mapping in OpenStreetMap and beyond. The extensive SVI of the entire street-network provides possibilities for other future usa-cases to enhance OSM like building material, blockage of drainage canals, road surface, etc.

The source data, code, and resulting datasets are publicly accessible, and the entire workflow is transferable to other geographic regions. The required resources are relatively low-cost, and open-source crowdsourcing tools, such as MapSwipe, can be leveraged to efficiently generate comprehensive datasets for model training.