DroneWaste dataset for waste recognition in drone imagery

DroneWaste dataset for waste recognition in drone imagery Luca Morandini aAndrea Diecidue a∗Thanos Petsanis b,c Enrico Targhini aGeorgios Karatzinis cGiacomo Boracchi a Elias B. Kosmatopoulos b,c Piero Fraternali a Athanasios Ch. Kapoutsis b,c October 7, 2025 aPolitecnico di Milano, Department of Electronics, Information, and Bioengineering, via Ponzio 34/5, Milan 20133, Italy bDemocritus University of Thrace, Department of Electrical Engineering, University in Xanthi, Greece cInformation Technologies Institute, The Centre for Research & Technology Hellas, Thessaloniki, Greece Abstract Illegal waste disposal has a negative impact on the environment and people’s quality of life. Drone imagery enables law enforcement authorities to efficiently assess the environmental impact during on-site inspections of suspicious landfill sites. Automated tools based on Deep Learning techniques can then quickly analyze aerial images to recognize several waste materials and classify their hazard level. However, large high-quality datasets are required for training and testing waste recognition models. Currently, no such datasets are publicly available, which limits the development of accurate waste identification algorithms. This paper presents DroneWaste , a dataset of aerial images extracted from orthomosaics that are reconstructed from drone-collected imagery. The dataset is a collection of 4993 images of 17 solid waste dumps that contain 20 different types of materials. The DroneWaste dataset is publicly accessible from the Zenodo repository. Technical validation proves that the dataset can be used for building object detection models able to recognize several types of waste in aerial imagery. This is a preprint version of a paper that is under submission in a peer-reviewed journal. 1 Background & Summary Illegal waste dumping is a form of environmental crime with severe impacts on ecosystems, such as increased pollution, loss of biodiversity, and risks to human health. This phenomenon is increasingly recognized to have significant adverse effects on the economy and on the ecology on a global scale [1,2]. Many illicit waste management organizations resort to illegal dumping or waste trafficking to minimize costs at the expense of environmental integrity and public health [3]. To counteract this escalating problem, Law Enforcement Agencies (LEAs) and Environmental Protection Agencies (EPAs) are required to identify potential illegal waste sites on a large-scale territory. When illegal activities are discovered, accurate forensic evidence must be collected to prosecute the involved criminals. A factor highly relevant to the definition of the criminal charge is the type of waste improperly treated. The penalties become more severe when toxic or dangerous substances are present, due to the high risks they pose to the environment and public health. The list of materials classified as extremely hazardous is specified by European Standards [4], including items such as asbestos slabs, hydrofluorocarbons, and toxic chemicals. On-site inspections of suspicious sites are expensive, require significant human effort, and can threaten the safety of the involved personnel when hazardous materials are present. Furthermore, illegal waste dumps are often located in inaccessible and remote areas far from densely populated ∗Corresponding author: [email protected] 1 places. Due to these constraints, combined with the vastness of the territory to be monitored, LEAs often conduct investigations using drones for imagery acquisition. Recent advances in Computer Vision (CV) and Deep Learning (DL) offer new opportunities for automating waste identification in images. Modern image analysis methods can be applied to satellite images to locate suspicious sites in vast regions [5]. Then, drone surveys support EPA inspectors in investigating the most critical waste dumping sites, thus enhancing the overall efficiency of the territory monitoring activity. However, modern DL models require large annotated datasets to learn the peculiar features of materials and recognize waste accumulations. In recent years, various datasets have been introduced in the waste detection literature, which utilize different data sources, including satellites, drones, and smartphones. Table 1provides a comparison of the most relevant waste detection datasets. Satellite imagery for waste identification has attracted increasing interest, mainly due to the availability of high-resolution images covering vast areas. Several publicly available datasets for solid waste detection in satellite images are mentioned in the scientific works. Solid Waste Aerial Detection (SWAD) [6] is a collection of satellite images encompassing urban, rural, and mountainous scenes, which are annotated with single-class bounding boxes delineating solid waste objects. Global Dumpsite Test Data (GDTD) dataset [7] includes numerous illegal dumps and a few regulated landfills. The employed satellite imagery has a Ground Sampling Distance (GSD) ranging from 0.3mto 1m. The relevant sites are annotated with bounding boxes labeled with one of four categories (domestic, construction, agricultural, or covered waste). The RS4SW dataset [8] employs Google Earth images (0.5mGSD) of three urban areas: Langfang (China), Faridabad (India), and Tezoyuca (Mexico). Each image is categorized as portraying solid waste (SW) or other nonsolid waste (non-SW) scenes. The waste annotated in this dataset comprises industrial materials, domestic litter in residential zones, and construction and demolition debris. The Construction Waste Landfill Dataset (CWLD) [9] includes Gaofen-2 satellite and Google Earth images of Beijing’s Changping and Daxing districts. It focuses on construction waste and supports semantic segmentation tasks, for which pixel-level annotation masks are provided. Finally, AerialWaste [5] comprises more than 10,000 aerial images with different sources: AGEA orthophotos, WorldView-3 satellite, and Google Earth images. The annotations provide binary image-level labels that specify the presence/absence of waste, and a subset of images is further tagged with 22 distinct waste materials. Remote sensing and satellite imagery are useful for scanning large areas and identifying suspicious locations. However, the provided GSD, ranging from 1.8m[6] to 20cm [5], is insufficient for the fine-grained identification of most waste materials. The adoption of Unmanned Aerial Vehicles (UAVs) [10] and the recent advances in DL models and CV techniques have provided LEAs and EPAs with effective tools for close-view investigation of the suspicious sites, once identified with satellite image analysis. The employed drones are equipped with a variety of sensors (e.g., high-resolution cameras, multispectral sensors, LiDAR, and GNSS-RTK), which support the detailed characterization of materials and the collection of forensic evidence from the inspected sites. A few datasets of UAV images depicting urban litter in the wild have been proposed in the literature. UAVVaste [11] comprises 772 low-altitude drone images of urban and natural scenes (e.g., streets, parks, lawns), annotated with 3700 instance masks and bounding boxes for a single rubbish class. However, this dataset includes samples of sparse small litter that are easily distinguishable from the surrounding environment. Therefore, they are not representative of the large accumulations of waste typical of illegal landfills. The SUIRD dataset [12] consists of 628 images collected from low-altitude UAV flights, which are augmented to create a more extensive training dataset. Accumulated or scattered garbage is annotated with single-class bounding boxes, but the waste categories are not provided. Finally, the RSD dataset [13] provides 2600 drone images annotated at the pixel level with three generic waste categories (construction, household, mixed). However, this level of detail is not relevant for training DL models to recognize a wide spectrum of waste materials. UAV images are geometrically corrected and stitched together to generate images with a GSD of 3.8 cm/px, which prevents the identification of small items. The perspective offered by low-altitude UAV flights enables the precise identification and classification of a wide range of materials. However, publicly available drone-captured datasets provide a single waste class [11,12] or a limited range of generic waste categories [13]. Therefore, these datasets are unsuitable for waste detection in landfills that include a wide range of materials captured in real-world scenarios. Besides the typical satellite and drone datasets, various studies introduced specialized data collections to recognize specific waste materials in images captured with hand-held cameras or smartphones. The TACO dataset [14] comprises 1500 high-resolution images of urban and natural scenes annotated with 4784 instance-level segmentation masks using a taxonomy of 60 fine-grained classes grouped into 28 super-categories. Early efforts such as TrashNet [15] and WaDaBa [16] provide a few thousand handheld photographs of common disposable objects. TrashNet includes 2 2527 images collected with smartphones that depict six different types of garbage (glass, paper, cardboard, plastic, metal, trash). WaDaBa includes 4000 images of plastic waste, generated by capturing 100 objects from 40 different views under controlled conditions. OpenLitterMap [17] is a continuously growing, crowd-sourced collection of over 100k geotagged images captured by smartphone cameras, containing a wide range of individual litter samples in urban environments. Finally, indoor waste recognition is addressed by the MJU-Waste dataset [18], which provides 2475 waste images captured with an RGB-Depth camera and annotated with instance-level segmentation masks. However, hand-held and smartphone cameras are ineffective for simulating the characteristics of LEA’s drone flights, which are therefore inadequate for waste identification from UAV imagery due to significant differences in perspective and resolution. Table 1: Description of datasets for waste recognition across different spatial scales, from cameralevel to UAV and satellite-based imagery. For each dataset, the supported tasks are reported (OD: Object Detection, IS: Instance Segmentation, SS: Semantic Segmentation, BC: Binary Classification, SLC: Single-Label Classification, MCL: Multi-Label Classification). Input Dataset Classes Images Task Sources Satellites SWAD [6] 1 1996 OD WorldView-2, SPOT GDTD [7] 4 2219 OD Different satellites RS4SW [8] 2 3680 BC Google Earth CWLD [9] 4 3653 SS Gaofen-2, Google Earth AerialWaste [5] 22 10,434 IS WorldView-3, Google Earth, aerial orthophotos UAV UAVVaste [11] 1 772 IS, OD UAV SUIRD [12] 1 628 OD UAV RSD [13] 3 2600 SS UAV DroneWaste (Ours) 20 4993 IS, OD UAV Ground sensors TACO [14] 60 4784 IS Smartphone camera TrashNet [15] 6 2527 SLC Smartphone camera WaDaBa [16] 6 4000 SLC Digital camera OpenLitterMap [17] 100+ 100,000+ MLC Smartphone camera MJU-Waste [18] 1 2475 IS RGB-D camera The detection and classification of waste instances is a challenging task due to the broad range of heterogeneous materials, their different shapes and textures, and the potential confusion with ordinary non-waste objects. DL algorithms require a substantial amount of high-quality annotated images to effectively recognize the distinctive features of waste materials. Currently, the lack of UAV-captured waste datasets in real-world scenarios hinders the research towards scalable and accurate waste recognition methods that can effectively assist LEAs and EPAs in their inspection activities. The contribution of this work is the DroneWaste dataset, a professionally curated set of UAV images for waste recognition, which is constructed on the following principles: 1. 4993 images are extracted from orthomosaics generated on 17 sites representing different scenarios and environments. 2. The average orthmosaic GSD in most sites is 2 cm/px, with some exceptions up to 2.8 cm/px. 3. Images are professionally annotated with 5135 waste instances that cover 20 waste types that include the majority of materials typically present in landfills. 4. Annotations are curated by professional photo-interpreters specialized in using UAV images for dump site inspection. 5. Each waste instance in an image is annotated with a polygon delimiting the boundaries and a bounding box that contains the waste object. 6. The dataset adheres to the standard COCO format [19]. An initial version of the DroneWaste dataset has been employed in our preliminary work [20] where an Object Detection (OD) network was trained to recognize 7 types of waste. Results showed that OD is effective, especially in recognizing regularly shaped waste materials (such as 3 Textile,Pallets and Tyres), with the best model achieving AP scores above 60%. The work also proposed a practical DL-based pipeline that supports the investigation processes of LEAs and EPAs in analyzing illegal waste sites. 2 Methods The DroneWaste dataset comprises images of real-world landfills and industrial waste disposal sites. The on-site UAV surveys were performed in Italy by ARPA Lombardia, a regional EPA, and in Greece by CERTH, a non-profit organization supervised by the General Secretariat for Research and Technology (GSRT) of the Hellenic Republic and one of the largest research centers in Greece. Commercial drones, described in Section 2.1, were employed to capture the images of 17 waste dumping sites. Figure 1illustrates the procedure for generating the DroneWaste dataset. For each inspected location, images acquired during a survey are processed to generate an orthomosaic, which is then manually annotated to create the site ground truth. From each orthomosaic, the dataset images are extracted and labeled by mapping the ground truth annotations previously defined on the entire orthomosaic. A final filtering step removes the annotations of partially visible waste instances. Data acquisition Site survey Mission design Orthomosaic generation Data annotation Tile extraction Tile annotation Ground truth creation Dataset preparation Image extraction Annotation mapping Annotation filtering Figure 1: The DroneWaste dataset creation workflow. Drone-acquired images are combined to generate an orthomosaic that covers the waste dumping site. Each orthomosaic is manually annotated to create the site ground truth. Images that compose the DroneWaste dataset are extracted using a sliding window approach. 2.1 Data acquisition The 17 UAV surveys that are included in the DroneWaste dataset capture real-world waste dumping activities in diverse scenarios and landscapes. The images from 6 locations show industrial sites where both authorized materials and illegal waste are present. Another 8 scenarios represent open landfills in rural areas characterized by the dumping of mixed waste or construction material. Finally, 3 locations are characterized by accumulations of waste scattered over a large area. Table 2specifies the characteristics of the sites included in the dataset. Data collection flights were conducted using two commercial drones: DJI Phantom 4 Pro and DJI Mavic 2 Enterprise Advanced. The Phantom 4 Pro drone mounts a DJI FC6310 camera that has a 1-inch CMOS sensor (20 MP, 24 mm equivalent focal length), an adjustable aperture (f/2.8–f/11), and wide ISO and shutter speed ranges. The Mavic 2 Enterprise Advanced drone is equipped with a DJI FC2453 camera, using a 1/2-inch CMOS sensor (48 MP, 24 mm equivalent focal length, f/2.8 fixed aperture). A commercial path planning software, DJI Pilot, is used for mission planning. The adopted flight paths, tailored for photogrammetry applications, are standard grid formations with −90◦pitch angle (i.e., the camera pointing down towards the ground). A 4 Table 2: Summary of the most relevant site properties. For each site, the number of images and annotations in the DroneWaste dataset is reported. The Materials column specifies the number of waste materials present at each site. Most orthomosaics are generated with a GSD of 2 cm/px. The reported GSD varies because, in those scenarios, it was not possible to collect enough images to produce an orthomosaic with higher resolution. The Area column reports the area in m2of each generated orthomosaic. Site ID Scenario Location Area GSD Images Annotations Materials Site 1 open landfill Greece 22,200 2.3 119 75 10 Site 2 open landfill Italy 18,300 2 130 100 12 Site 3 scattered waste Italy 95,500 2 698 243 14 Site 4 industrial site Italy 25,100 2 182 461 12 Site 5 industrial site Italy 30,000 2 217 649 18 Site 6 scattered waste Greece 227,400 2.8 813 252 13 Site 7 open landfill Greece 1300 2.4 6 10 2 Site 8 industrial site Italy 28,800 2 210 173 11 Site 9 open landfill Greece 11,800 2 83 526 11 Site 10 open landfill Greece 3800 2.3 17 25 5 Site 11 open landfill Italy 113,100 2 825 173 11 Site 12 industrial site Italy 26,100 2 186 540 10 Site 13 industrial site Italy 24,200 2 169 658 14 Site 14 industrial site Italy 40,400 2 279 792 14 Site 15 open landfill Greece 22,600 2.2 136 115 8 Site 16 scattered waste Italy 115,600 1.9 848 186 15 Site 17 open landfill Greece 15,900 2.4 75 157 5 standardized data acquisition protocol was used across all sites, and several photogrammetry best practices were implemented to improve the quality of the reconstructed orthomosaic. In particular, the frontlap between images is kept constant at 80%, the sidelap at 70%, and dynamic elements (e.g., moving vehicles or people) were avoided. The flight altitude ranges between 15 and 100 meters, depending on the area size and the obstacles. In smaller landfill sites, a higher number of low-altitude UAV images can be collected to cover the surveyed area. In contrast, at larger dumping sites with obstructions (e.g., buildings), low-altitude flights are forbidden due to the risk of collision. Aerial images were captured with the RGB camera at full sensor resolution. Each image embeds several metadata in EXIF format, such as GPS coordinates, camera gimbal orientation, and lens properties. This information is exploited during the orthomosaic generation process by photogrammetry software to provide a rough initialization for all camera poses, improving the reconstruction quality. 2.2 Orthomosaic generation The UAV images collected during a flight are combined using photogrammetry software to generate a georeferenced orthomosaic of the surveyed site. An orthomosaic is a large image with high detail and resolution, assembled from many smaller UAV images. A georeferenced orthomosaic is a special case where each pixel of the output map is associated with a location with known coordinates. Any photogrammetry software would be suitable for the task. Specifically, this work uses OpenDroneMap (ODM) [21], an open-source tool for drone image processing, which applies several processing phases to correct distortions in UAV images, create a 3D reconstruction, and generate a georeferenced orthomosaic covering the entire scene. The image is orthorectified, meaning that distortions and perspectives are removed from the resulting image. The DroneWaste orthomosaic maps are generated with a GSD of 2 cm/px. This value represents a trade-off between image resolution and computational requirements. High spatial resolution enables the discrimination of fine-grained textures in waste materials. Yet, an excessively high resolution increases the computational and storage requirements, making it impractical for large survey areas. Table 2reports the GSD of each site orthomosaic. For most sites, the target GSD of 2 cm/px is reached during the generation process. However, in some scenarios, it was not possible to collect enough images to produce an orthomosaic with a higher GSD, so some orthomosaics are reconstructed with a slightly lower spatial resolution. The maximum (worst) GSD is 2.8 cm/px 5 for Site 6. Figure 2illustrates examples of tiles extracted from the generated orthomosaics. The captured scenarios are very heterogeneous and reflect the range of waste disposal sites and of materials composing the DroneWaste dataset. Figure 2: Examples of tiles extracted from the DroneWaste orthomosaics, along with the colorcoded legend of the annotated waste categories. 2.3 Data annotation The produced orthomosaics cover extensive areas with high spatial resolution, resulting in a significant image size. For example, Site 6, the largest orthomosaic included in the DroneWaste dataset, has a size of 19,450×16,491 px and occupies 686 MB of disk space. Such dimensions are too large for commercial annotation tools, typically employed to label smaller images. To enable the use of standard annotation tools, a site orthomosaic is decomposed into smaller overlapping tiles that cover a fixed area of the ground. Tiles with a size of 40×40 m are extracted using a sliding window approach with an overlap of 5 m between adjacent tiles. The tiles were manually annotated using Roboflow [22], a commercial annotation tool that enables collaborative workflows. The annotation process utilizes a Smart Polygon functionality, which is a semi-automatic tool that leverages the Segment Anything Model (SAM) [23] to quickly segment individual objects in the scene. SAM enables the rapid definition of polygons around waste instances by selecting a few points on the image. The automatically created polygons can be manually corrected to better fit the shape of the waste instance and the waste class is added to each polygon by the expert annotator. Table 3lists the 20 categories considered in the DroneWaste dataset. Each waste class is associated with a European Waste Code (EWC) [4] to align the annotations with the European List of Waste (LoW) defined by the European Commission under Directive 2008/98/EC on waste. The waste categories included in the dataset have been chosen based on the presence of materials in the site surveys. The Materials column in Table 2shows that all locations contain only a subset of the 20 waste categories, resulting in a heterogeneous distribution across sites. The waste classes can be divided into pile categories, including materials typically occurring in piles or heaps with irregular boundaries (e.g., Rubble,Mixed items), and instance categories, for which individual waste elements can be recognized (e.g., Metal barrels,Pallets,Tyres). Note that the dataset includes instances of the class Asbestos, a material commonly used in the roofs of old buildings, which needs to be identified due to its high toxicity even when present in the form of roof cover and not as proper waste. The manual annotation phase was conducted by a team of 7 researchers and 2 investigators from an EPA, all experienced in recognizing waste from UAV imagery. Since multiple annotators from different organizations were involved, the procedure was standardized to ensure consistency. Guidelines were established to provide a brief description and examples of each waste category, and 6 Table 3: The list of 20 categories of waste annotated in the DroneWaste dataset. Each class is assigned an EWC code to uniquely identify the waste material. Color Category Type EWC ●Rubble pile 12.61 Soils ●Construction and demolition materials pile 12.11 Concrete, bricks and gypsum waste ●Asphalt milling pile 12.12 Waste hydrocarbonised road-surfacing material ●Excavation materials pile 12.31 Waste of naturally occurring minerals ●Appliances instance 08.21 Discarded major household equipment ●Electronic equipment instance 08.23 Other discarded electrical and electronic equipment ●Furniture instance 10.11 Household wastes ●Metal barrels instance 06.31 Mixed metallic packaging ●Plastic packaging instance 07.41 Plastic packaging wastes ●Wood pile 07.53 Other wood wastes ●Pallets instance 07.51 Wood packaging ●Scrap pile 06.11 Ferrous metal waste and scrap ●Plastic pile 07.42 Other plastic wastes ●Vehicles instance 08.12 Other discarded vehicles ●Tyres instance 07.31 Used tyres ●Paper instance 07.2 Paper and cardboard wastes ●Foundry pile 12.42 Slags and ashes from thermal treatment and combustion ●Asbestos instance 12.21 Asbestos wastes ●Textile instance 07.6 Textiles wastes ●Mixed items pile 10.2 Mixed and undifferentiated materials the annotations created by a team member were reviewed by a different annotator to minimise user bias. After the manual data labelling phase, all polygons annotated on the tiles were projected onto the orthomosaic. When a larger waste element was not fully contained within a single tile, polygons with the same category from nearby tiles were combined into a single polygon that covered the entire waste instance. Further annotation validation identified and resolved inconsistent cases, such as: •Duplicate annotations on the same tile (e.g., nested annotations of the same class). •Overlapping annotations with different categories on the same tile. •Overlapping annotations with different categories on adjacent tiles (e.g., incorrectly classified partially visible waste instances). Such conflicts were identified by an automatic script and resolved manually. The final ground truth of each site includes a series of polygons associated with a waste category and georeferenced to the orthomosaic coordinates. Figure 3shows all the orthomosaics included in the dataset. Some of them cover an extensive area while others (e.g., Site 7,Site 11) are much smaller. In many cases, the waste is scattered throughout the scene with large areas of waste-free background. 2.4 Dataset preparation The tiles that compose the DroneWaste dataset have a standard size of 640×640 px, which reflects the typical dimensions used by image analysis models. The sliding window processing step extracts tiles with a 10% overlap between rows and columns to avoid missing objects on the edge of a tile. Images near the orthomosaic border that contained more than 70% black pixels were discarded from the dataset. When using a sliding window, it may happen that only a small portion of a waste instance is visible in a tile. Such partial visibility negatively impacts the waste recognition task, as a detection model may fail to learn relevant features from a reduced portion of a waste instance. Therefore, small partial annotations were removed. The filtering logic differs between instance and pile waste categories. For instance categories, where objects have regular size and shape, samples are kept if a minimum portion of the object remains visible in an image. Category-specific 7 Figure 3: The full orthomosaics of all the sites included in the DroneWaste dataset. thresholds were defined empirically by considering the smallest image region still permitting the recognition of an object and the threshold values range from 10% to 20% of the whole object area. For pile categories, the area size of annotations varies significantly (piles can range from very small to very large) and therefore a partial heap annotation is filtered out when it represents less than 1% of the whole annotation. Ultimately, the dataset creation process outputs the ground truth samples of all the 17 sites. Table 2reports the number of images and of annotated waste elements. Each ground truth sample has a segmentation mask and a bounding box. The full collection of ground truth samples of DroneWaste includes 5135 annotations on 4993 images. 8 3 Data Record The DroneWaste dataset is published in a Zenodo repository (https://doi.org/10.5281/zenodo. 17045558) and comprises the following artifacts, which constitute the public part of the dataset: 1. Images folder: contains the images extracted from all site orthomosaics. The provided images are not georeferenced because the coordinates are considered sensitive information and have therefore been removed. 2. Dataset ground truth: a JSON file that describes the images and annotations of the dataset using the COCO format. Both segmentation masks and bounding boxes are defined for all annotated waste instances. 3. Information file: general information about the current dataset version. Not relevant for training or evaluating detection models. Figure 4illustrates the organization of the DroneWaste dataset directory structure. idronewaste/ Ddronewaste v1.0.json Ainfo.txt iimages/ Jsite1 4.png Jsite1 5.png . . . Jsite2 4.png . . . Figure 4: Directory structure of the DroneWaste dataset, showing the organization of metadata files and imagery. Figure 5illustrates the distribution of the waste categories in the dataset. Although the class distribution may appear imbalanced, several considerations are worth mentioning: 1. Some categories are inherently rare (e.g., Asphalt milling,Foundry or Paper), thus underrepresented in the DroneWaste dataset. Nevertheless, these rare instances have been annotated to enable the development and performance assessment of detection methods with few samples. 2. Categories representing individual items (e.g., Plastic packaging,Tyres, or Metal barrels) often result in many single-object small annotations. 3. Conversely, categories aggregated in piles (e.g., Rubble or Mixed items) are typically annotated as few large patches covering significant portions of an image. The difference in distribution between instance and pile classes is also evident in Figure 6, which reports the relative pixel coverage of each class in the images. Items such as Pallets and Textile, which are typically small and isolated, have a larger number of annotations but cover a smaller area compared to the other classes. On the other hand, Rubble or Mixed items are often grouped in piles and occupy a larger region of the images. Another material that covers wide areas is Asbestos because its instances most often correspond to roofs that span large portions of the image. Figure 7presents a site-wise breakdown of the dataset, showing for each location the number of UAV-captured photos compared to the annotated images extracted from the orthomosaic. Each bar is divided into two colors, representing images annotated with at least one waste element and the images that contain only background. The chart also reports the percentage of backgroundonly images to the total number of site images. Images without any visible waste material are 9 [5] Rocio Nahime Torres and Piero Fraternali. Aerialwaste dataset for landfill discovery in aerial and satellite images. Scientific Data, 10(1):63, 2023. 2,3 [6] Liming Zhou, Xiaohan Rao, Yahui Li, Xianyu Zuo, Yang Liu, Yinghao Lin, and Yong Yang. Swdet: Anchor-based object detector for solid waste detection in aerial images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16:306–320, 2022. 2,3 [7] Xian Sun, Dongshuo Yin, Fei Qin, Hongfeng Yu, Wanxuan Lu, Fanglong Yao, Qibin He, Xingliang Huang, Zhiyuan Yan, Peijin Wang, et al. Revealing influencing factors on global waste distribution via deep-learning based dumpsite detection from satellite imagery. Nature Communications, 14(1):1444, 2023. 2,3 [8] Bowen Niu, Quanlong Feng, Jianyu Yang, Boan Chen, Bingbo Gao, Jiantao Liu, Yi Li, and Jianhua Gong. Solid waste mapping based on very high resolution remote sensing imagery and a novel deep learning approach. Geocarto International, 38(1):2164361, 2023. 2,3 [9] Shaofu Lin, Lei Huang, Xiliang Liu, Guihong Chen, and Zhe Fu. A construction waste landfill dataset of two districts in beijing, china from high resolution satellite images. Scientific Data, 11(1):388, 2024. 2,3 [10] Tyrone Bright, Sarp Adali, and Cristina Trois. Systemic review and meta-analysis: The application of ai-powered drone technology with computer vision and deep learning networks in waste management. Drones, 9(8), 2025. 2 [11] Marek Kraft, Mateusz Piechocki, Bartosz Ptak, and Krzysztof Walas. Autonomous, onboard vision-based trash and litter detection in low altitude aerial images collected by an unmanned aerial vehicle. Remote Sensing, 13(5):965, 2021. 2,3 [12] Weiyang Chen, Yiyang Zhao, Tengfei You, Haifeng Wang, Yang Yang, and Kun Yang. Automatic detection of scattered garbage regions using small unmanned aerial vehicle low-altitude remote sensing images for high-altitude natural reserve environmental protection. Environmental science & technology, 55(6):3604–3611, 2021. 2,3 [13] Yang Liu, Bo Zhao, Xuepeng Zhang, Wei Nie, Peng Gou, Jiachun Liao, and Kunxin Wang. A practical deep learning architecture for large-area solid wastes monitoring based on uav imagery. Applied Sciences, 14(5), 2024. 2,3 [14] Pedro F Proen¸ca and Pedro Sim˜oes. Taco: Trash annotations in context for litter detection. arXiv preprint arXiv:2003.06975, 2020. 2,3 [15] Mindy Yang and Gary Thung. Classification of trash for recyclability status. CS229 project report, 2016(1):3, 2016. https://huggingface.co/datasets/garythung/trashnet,https: //github.com/garythung/trashnet.2,3 [16] Janusz Bobulski and Jacek Piatkowski. Pet waste classification method and plastic waste database-wadaba. In Image Processing and Communications Challenges 9: 9th International Conference, IP&C’2017 Bydgoszcz, Poland, September 2017, Proceedings, pages 57–64. Springer, 2018. 2,3 [17] Se´an Lynch. Openlittermap. com–open data on plastic pollution with blockchain rewards (littercoin). Open Geospatial Data, Software and Standards, 3(1):1–10, 2018. 3 [18] Tao Wang, Yuanzheng Cai, Lingyu Liang, and Dongyi Ye. A multi-level approach to waste object segmentation. Sensors, 20(14):3816, 2020. 3 [19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision – ECCV 2014, pages 740–755, Cham, 2014. Springer International Publishing. 3,11 [20] Morandini, Luca, Buda, Alessio, and Fraternali, Piero. Geospatial artificial intelligence for solid waste recognition from uav imagery. E3S Web Conf., 643:01004, 2025. 3 [21] OpenDroneMap. Odm—a command line toolkit to generate maps, point clouds, 3d models and dems from drone, balloon or kite images. OpenDroneMap/ODM GitHub Page, 2020. 5 16 [22] Hansen T. Dwyer B., Nelson J. Roboflow (version 1.0) [software]. https://roboflow.com, 2024. [accessed 25 January 2025]. 6 [23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Doll´ar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 6 [24] Muhammad Yaseen. What is yolov8: An in-depth exploration of the internal features of the next-generation object detector, 2024. 11 [25] Yunjie Tian, Qixiang Ye, and David Doermann. Yolov12: Attention-centric real-time object detectors, 2025. 11 [26] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137–1149, 2017. 11 [27] Eduardo Villamor Medina, Theodora Tsikrika, Piero Fraternali, Luigi Caldararu, Vasiliki Efstathiou, Sandra Balbierz, Efstathios Skarlatos, Eva Korenjak, Anastasios Karakostas, Luca Di Nuovo, Federico Benolli, Dario Bellingeri, Dries Borloo, Renato Sciunnach, Jimmy Berggren, Ovidiu Manolache, Ioannis Petropoulos, Radu Bors, Nir Haimov, Andrew Staniforth, and Jordan Thompson. PERIVALLON: Improved Intelligence Picture and Operational Capacities to Combat Organised Environmental Crime, pages 197–211. Springer Nature Switzerland, Cham, 2025. 15 Author Contributions All authors contributed to the study design, dataset development, and manuscript writing. Data preparation, visualization, and technical validation were carried out by L.M., A.D., T.P., E.T., G.K. and supervised by G.B., E.B.K., P.F., A.C.K. Competing Interests The authors declare no competing interests. Acknowledgements This work was partially funded by European Union’s Horizon Europe project PERIVALLON - Protecting the EuRopean terrItory from organised enVironmentAl crime through inteLLigent threat detectiON tools, under grant agreement no. 101073952. We acknowledge the Environmental Protection Agency ARPA Lombardia, particularly the OU CREO (Regional Center of Earth Observation), for their invaluable expertise during the data acquisition and dataset annotation phases of this research. This work was supported by the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (ID:90254). 17