Multi_Source_Swimming_Pool_Detection_Dataset

# SPDD — Swimming Pool Detection Dataset > A large-scale, multi-source benchmark for detecting outdoor swimming pools in high-resolution Earth observation imagery, with a Florida-specific component for statewide mapping. ## Dataset Description ### Dataset Overview **SPDD (Swimming Pool Detection Dataset)** is a publicly released object-detection benchmark built for reliable inventory mapping of outdoor swimming pools from sub-meter remote sensing imagery. It combines broad scene diversity from six open datasets across the Americas, Europe, and Asia (**SPDD-A**, 22,705 tiles / 89,668 instances) with a Florida-specific branch constructed from OpenStreetMap (OSM) polygons, model-assisted screening, and manual verification (**SPDD-B**, 14,478 tiles / 168,644 instances). The full dataset contains **37,183 image tiles and 258,312 annotated swimming-pool instances**, all in a unified horizontal bounding-box format. SPDD is designed to support small-object detectors and vision-foundation-model research, and serves as the training/evaluation basis for statewide pool mapping in Florida. **Motivation.** Outdoor swimming pools are small, morphologically variable artificial water features whose abundance and spatial distribution inform water-demand assessment, mosquito-borne disease control, property-related spatial analysis, and emergency planning. Public records (e.g., OSM) are incomplete and unevenly updated, and classical water indices (NDWI/MNDWI) are unreliable at pool scale. Generic aerial benchmarks such as DOTA and AI-TOD include a pool class but with limited, geographically biased instances. SPDD was constructed to provide (1) broad visual diversity from multiple regions and (2) local domain adaptation to dense, occluded, and screen-enclosure-rich Florida residential settings. | Item | Value | |---|---| | Total image tiles | 37,183 | | Total annotated instances | 258,312 | | Object classes | 1 (swimming pool) | | Annotation format | Horizontal bounding box, normalized YOLO (class x_center y_center w h) | | Tile sizes | 512×512, 640×640, 800×800 px | | Imagery | RGB, high-resolution (Florida branch at 0.6 m GSD) | | Geographic coverage | North/South America, Europe, Asia + Florida (USA) | | Train / Val / Test | 29,852 / 3,683 / 3,648 | **Supported usage scenarios.** Small-object detection in high-resolution remote sensing imagery; benchmarking one-stage / transformer / foundation-model-enhanced detectors; swimming-pool and artificial-water-body inventory mapping; domain-adaptation and spatial-transfer studies (e.g., from generic to Florida residential scenes); downstream water-demand, public-health, and urban-form research (with proper covariate linkage). ### Tasks Supported by the Dataset The primary supported task is **object detection** (horizontal bounding-box localization of outdoor swimming pools in RGB remote-sensing tiles). Reference benchmark results on SPDD: Using SPDD, the foundation-model-enhanced detector **D-FMSE** (frozen DINOv3 features injected into YOLOv12-X via gated residual fusion) achieves **AP₅₀ = 96.84%** on the held-out test set, +2.20 points over the YOLOv12-X baseline (94.64%). Across ten spatially independent Florida zones, micro-averaged precision/recall are 98.18% / 91.61%. (Reported for reference; users should train their own baselines.) ## Dataset Format and Structure ### Data Format All annotations use the **normalized YOLO format** (one `.txt` per image): ``` <class_id> <x_center> <y_center> <width> <height> ``` - `class_id` = `0` (single class: swimming pool) - All coordinates normalized to [0, 1] relative to the image dimensions - Horizontal bounding boxes (HBB); original OBB / mask annotations were converted to HBB, which may increase labeled background area for irregular/rotated pools Each image tile (`images/<split>/*.jpg`) has a co-located label file (`labels/<split>/<same_basename>.txt`). A `spdd.yaml` data config is provided defining the single class `swimming_pool` and the split directories. **Example label file (`labels/train/00009_slice_2_0.txt`):** ``` 0 0.828192 0.114764 0.031932 0.014581 ``` ### Dataset Loading Method The dataset is organized as paired `images/` and `labels/` folders and works with any YOLO-compatible pipeline. Below are two recommended ways. **1. Ultralytics / PyTorch (training):** ```python from ultralytics import YOLO model = YOLO("yolov12x.yaml") model.train(data="spdd.yaml", imgsz=800, epochs=300, batch=24) ``` **2. ModelScope SDK (download the repository):** ```python from modelscope.hub.snapshot_download import snapshot_download dataset_dir = snapshot_download("YOUR_USERNAME/SPDD") # dataset_dir contains images/ and labels/ split into train/ val/ test/, # plus spdd.yaml for YOLO pipelines. ``` ### Data Splits SPDD is split into `train` / `validation` (val) / `test` subsets, with consistent scale composition across splits. The `configs` block in the YAML header maps each split to its image directory; labels are co-located under `labels/<split>/`. | Subset | Images | Instances | Train | Val | Test | |---|---|---|---|---|---| | SPDD-A | 22,705 | 89,668 | 18,222 | 2,247 | 2,236 | | SPDD-B | 14,478 | 168,644 | 11,630 | 1,436 | 1,412 | | **Total** | **37,183** | **258,312** | **29,852** | **3,683** | **3,648** | On-disk layout: ``` SwimmingPool-OpenSource_OSM-20251229/ ├── images/ │ ├── train/ (29,852 .jpg) │ ├── val/ ( 3,683 .jpg) │ └── test/ ( 3,648 .jpg) ├── labels/ │ ├── train/ (29,852 .txt) │ ├── val/ ( 3,683 .txt) │ └── test/ ( 3,648 .txt) ├── spdd.yaml └── README.md ``` ## Dataset Generation Related Information ### Original Data SPDD is assembled from two complementary branches. **SPDD-A — Multi-source open datasets.** Six publicly available remote-sensing datasets are extracted and standardized into normalized YOLO horizontal boxes (original annotations include HBB, OBB, and instance masks). Tiles are cropped to 512/640/800 px depending on source characteristics. | Source | Region / imagery | Tile size | Images | Instances | |---|---|---|---|---| | DOTA-v2.0 | Multi-region aerial | 640×640 | 1,738 | 5,624 | | AI-TOD-v2 | Aerial tiny-object | 800×800 | 185 | 610 | | SODA-A | Aerial small-object | 800×800 | 9,153 | 61,207 | | BH-Dataset | Belo Horizonte (Brazil), Google Earth | 800×800 | 3,239 | 9,510 | | OSM-Swimming-Pools | Spain & Portugal, Mapbox | 512×512 | 7,166 | 9,523 | | CANNES-Dataset | Cannes (France) large-area | 512×512 | 1,224 | 3,194 | | **SPDD-A total** | Mixed | Mixed | **22,705** | **89,668** | **SPDD-B — Florida-specific samples.** Starting from 34,266 OSM polygons tagged `leisure=swimming_pool` in Florida, an 800×800 tile is extracted around each polygon centroid registered to 0.6 m orthorectified RGB imagery. A YOLOv12-X model pre-trained on SPDD-A prioritizes candidate tiles (confidence > 0.9 retained), followed by manual review to keep only genuine pools with acceptable box alignment. Additional visible pools inside selected tiles are also annotated, expanding the instance count well beyond the original OSM polygon count. | Branch | Images | Instances | |---|---|---| | SPDD-A | 22,705 | 89,668 | | SPDD-B | 14,478 | 168,644 | | **SPDD total** | **37,183** | **258,312** | ### Dataset Annotation **Annotation process.** Source annotations (HBB, OBB, instance masks) are converted to a unified normalized YOLO horizontal-box format. For SPDD-B, candidate tiles are screened by a YOLOv12-X detector pre-trained on SPDD-A (confidence > 0.9), then manually verified and refined; all visible pools within retained tiles are additionally boxed. **Annotators.** SPDD-A derives from the original public-dataset annotations (credited to their respective authors). SPDD-B boxes are produced via model-assisted pre-labeling followed by manual verification by the authors. **Data statistics.** - **Scale composition (COCO-style):** small 74.34%, medium 24.87%, large 0.79%. Bounding-box area ranges from ~12 to 460,000 px² (strong long-tail). - **Instance density per image:** median 3, mean 6.9; 56.2% of images contain ≤3 instances, 11.6% contain >15 instances (balanced sparse vs. dense scenes). - **Scene types:** villa, apartment, community-facility, residential, suburban, and coastal settings. ## Dataset Copyright Information SPDD is a **composite dataset**. Each source component retains its original license and terms of use. Because the release unites multiple licenses (including ODbL from OpenStreetMap-derived components), the repository-level `license` in the YAML header is declared as **`other`**; the effective terms are the **union** of the following: - **DOTA-v2.0** — see DOTA license / user agreement. - **AI-TOD-v2** — see AI-TOD license. - **SODA-A** — see SODA-A license. - **BH-Dataset** — Roboflow / original authors' terms. - **OSM-Swimming-Pools** — derived from OpenStreetMap data, © OpenStreetMap contributors (ODbL). - **CANNES-Dataset** — see original authors' terms. - **SPDD-B (Florida)** — derived from OSM polygons (ODbL) and 0.6 m orthorectified imagery; **redistribution of the underlying imagery must follow the imagery provider's terms** — annotations and reproducible download scripts are provided where imagery cannot be redistributed. ## Citation Format If you use SPDD, please cite the accompanying paper: ```bibtex @article{lu2026statewide, title = {Statewide mapping of outdoor swimming pools in Florida from high-resolution Earth observation imagery: A foundation-model-enhanced approach}, author = {Lu, Wanjie and Liu, Yuxuan and Hu, Zhihua and Xu, Nan and Zhang, Ruiqian}, journal = {International Journal of Applied Earth Observation and Geoinformation (JAG)}, year = {2026}, note = {Under review} } ``` *(Update the citation with the final venue, DOI, and dataset DOI/repository link before release.)* ## Other Relevant Information - **Personal / sensitive information:** the dataset covers residential rooftops and backyards at 0.6–~1 m GSD. No personal identifiers are included, but tiles may capture private property. Use is restricted to aggregate, non-individual-level research (water demand, public health, urban form) and must comply with the imagery provider's terms and applicable privacy law. - **Ethics & limitations:** boxes are HBB and may over-enclose irregular/rotated pools; OSM-derived pools may under-represent unmapped or private pools; Florida branch geometry is specific to local enclosure styles and may not transfer globally without adaptation. - **Contact:** for questions about the dataset, splits, or the Florida detection layer, contact the authors (see paper correspondence). --- *This dataset card is provided for the ModelScope community release.*