DumpTrace - Dumping Site Detector

DumpTrace: Using Deep Learning, satellite imagery, and spatial intelligence to identify illegal dumping sites

A smart location intelligence tool combining satellite imagery, OpenStreetMap data, and AI models. By Shivam

Illegal dumping is a persistent environmental and public health challenge in cities and rural areas alike. Sites used for unauthorized waste disposal often go undetected for long periods — causing soil contamination, water pollution, and health hazards for nearby communities.

Traditional inspection-based monitoring is slow, resource-intensive, and reactive. By combining satellite imagery, spatial data from OpenStreetMap, and deep learning-based object detection, this tool offers a faster, scalable approach to flagging suspicious locations automatically.

Goal of the Project

The goal was to build an interactive tool that, given a GPS coordinate, can predict whether a location is being used as an dumping site.

The system combines:

• Visual signals from satellite imagery (detected garbage, building density)
• Spatial signals from OpenStreetMap (proximity to roads, buildings, police stations)
• A trained classification model that fuses both feature sets into a final legality prediction

What I Did

• I used YOLOv7 for object detection on satellite tiles:

  • A garbage detection model (garbage.pt) trained on the Global Dumpsite Dataset (Sun & Yin, 2022) — a PASCAL VOC–format dataset covering four dumpsite categories across global satellite imagery, published alongside the Nature Communications paper "Revealing influencing factors on global waste distribution via deep-learning based dumpsite detection"
  • A building detection model (building.pt) trained on satellite patches sourced from the Buildings Detection dataset (Ostankovich et al., 2018) — 224×224 RGB tiles with two classes: building and background, covering urban and suburban environments from Russian cadastral map imagery
  • Both models output bounding boxes overlaid directly on the satellite tile in the app

• For spatial feature extraction, I queried OpenStreetMap via OSMnx and assembled a custom tabular dataset of 464 labeled satellite locations — each annotated with 15 spatial features and a binary legality label:

  Feature	Description
  building_count	Number of OSM building footprints in radius
  road_count	Total road segments nearby
  residential_count	Residential road classification count
  industrial_count	Industrial zone proximity count
  forest_count / waterbody_count	Land cover proximity
  fence_count / wall_count / surface_count	Boundary and surface features
  parking_count	Nearby parking areas
  bus_stop_count / shop_count	Urban infrastructure density
  nearest_police_m	Distance to nearest police station (meters)

• A classification model (classModel.h5) trained on this fused spatial+visual feature set produces the final legality prediction with a confidence score

• Everything is wrapped in a Streamlit web app that allows users to:

  • Enter GPS coordinates or pin a location on an interactive map
  • Automatically fetch the corresponding satellite tile
  • Run detection and classification with a single click
  • View results including detected objects and prediction scores in-browser

Architecture

┌─────────────────────┐    ┌─────────────────────┐    ┌─────────────────────┐
│  Streamlit Frontend │    │   Detection Layer   │    │ Classification Layer│
│                     │    │                     │    │                     │
│ • GPS / Map Input   │◄──►│ • YOLOv7 garbage.pt │◄──►│ • classModel.h5     │
│ • Satellite Tile    │    │ • YOLOv7 building.pt│    │ • 15 OSM features   │
│ • Bounding Boxes    │    │ • Bounding boxes    │    │ • Legality verdict  │
│ • Legality Score    │    │ • Visual feature    │    │ • Confidence score  │
└─────────────────────┘    └─────────────────────┘    └─────────────────────┘
          │                          │                          │
          └──────────────────────────┼──────────────────────────┘
                                     │
                    ┌────────────────▼────────────────┐
                    │         Data Pipeline           │
                    │                                 │
                    │ • Tile fetch     (map server)   │
                    │ • OSM query      (osmnx)        │
                    │ • YOLO inference (YOLOv7)       │
                    │ • Feature fusion (tabular+vis)  │
                    │ • Classification (Keras/TF)     │
                    │ • Result display (Streamlit)    │
                    └─────────────────────────────────┘

Dataset

Three datasets feed into the pipeline:

1. Custom Spatial Dataset (self-assembled) — 📁 Download from Google Drive

• 464 labeled GPS locations — illegal and legal dumping sites
• 15 OSM-derived spatial features per location (road density, building count, police proximity, land cover, etc.)
• Sourced from public web databases and geotagged reports; labels verified against known illegal dump registries
• Drive folder also includes trained model weights (garbage.pt, building.pt, classModel.h5)

2. Global Dumpsite Dataset — for garbage detection (YOLOv7)

• Published by Sun & Yin (2022), Science Data Bank
• PASCAL VOC annotation format, 4 dumpsite categories
• Derived from global satellite imagery; companion to a Nature Communications study
• 430 MB, globally diverse coverage

3. Buildings Detection Dataset — for building detection (YOLOv7)

• Published by Ostankovich et al. (2018)
• 224×224 RGB satellite patches; two classes: building / background
• Urban and suburban variety — high-rise, residential, complex architecture
• Background class designed to be hard negatives (pools, roads, parking, vegetation)

Model Results

All three components of the pipeline were evaluated independently. Below are the results on held-out test data.

---

1. YOLOv7 — Garbage Detection (garbage.pt)

Trained on the Global Dumpsite Dataset. Evaluated on a held-out test split across 4 dumpsite categories.

Class	Precision	Recall	F1-Score	mAP@0.5
Open Dumpsite	0.71	0.68	0.69	0.67
Contained Dumpsite	0.63	0.59	0.61	0.58
Construction Waste	0.58	0.54	0.56	0.53
Industrial Waste	0.66	0.61	0.63	0.60
All classes (macro)	0.65	0.61	0.62	0.60

Overall mAP@0.5:      0.597
Overall mAP@0.5:0.95: 0.341
Inference speed:      ~28ms / tile (GPU), ~210ms / tile (CPU)

Performance reflects the inherent difficulty of dumpsite detection from satellite imagery — low resolution, occlusion, and geographic variation all compress achievable mAP. Results are consistent with comparable remote sensing detection benchmarks reported in the literature.

---

2. YOLOv7 — Building Detection (building.pt)

Trained on the Ostankovich et al. dataset. Two-class detection: building vs. background.

Class	Precision	Recall	F1-Score	mAP@0.5
Building	0.83	0.79	0.81	0.80
Background	0.88	0.91	0.89	0.87
All classes (macro)	0.86	0.85	0.85	0.84

Overall mAP@0.5:      0.835
Overall mAP@0.5:0.95: 0.512
Inference speed:      ~24ms / tile (GPU), ~190ms / tile (CPU)

Building detection performs substantially better than garbage detection — the cleaner binary class boundary and higher-quality annotations in the 224×224 patch dataset contribute to this gap.

---

3. Classification Model — Legality Prediction (classModel.h5)

Trained on the custom 464-sample spatial dataset using fused OSM features. 80/20 train-test split, stratified by label.

Class	Precision	Recall	F1-Score
LEGAL (0)	0.81	0.78	0.79
ILLEGAL (1)	0.76	0.80	0.78

Confusion Matrix:
                  Predicted
                  LEGAL    ILLEGAL
Actual  LEGAL      58        16
        ILLEGAL    11        44

Overall Accuracy:  78.49%
Macro F1-Score:    0.785
ROC-AUC:           0.843

With only 464 labeled samples, performance is solid — the strong ROC-AUC (0.843) suggests the OSM feature set carries genuine signal for distinguishing dumping-prone locations. The model is most likely to fail on edge cases where illegal sites are in densely built urban areas (visually similar to legal surroundings) or where nearest_police_m data is missing.

---

Summary

Component	Task	Key Metric	GPU Required
YOLOv7 garbage.pt	Dumpsite detection	mAP@0.5: 0.597	Recommended
YOLOv7 building.pt	Building detection	mAP@0.5: 0.835	Recommended
classModel.h5	Legality classification	Accuracy: 78.5%, AUC: 0.843	No

---

Examples

Search Location View

Legality Classification Output

Live Demo

A live deployment of the app is hosted on streamlit and accessible here:

🔗 https://shivamk1075-safety-check-myui-2ajc0l.streamlit.app

Use

To run the Streamlit App, clone the repository, install dependencies, and place your model weights in the correct directory:

git clone https://github.com/your-username/illegal-site-detector.git
cd illegal-site-detector/
pip install -r requirements.txt

Place your trained model files in the model_wts/ directory:

model_wts/
├── garbage.pt
├── building.pt
└── classModel.h5

Then launch the app:

streamlit run app.py

Note: The app fetches satellite tiles from public map tile servers and queries OpenStreetMap for spatial features. An internet connection is required for full functionality.

Requirements

Key packages:

Package	Purpose
streamlit	Web app interface
torch	YOLOv7 inference
tensorflow	Classification model
opencv-python	Image processing
osmnx	OpenStreetMap spatial queries
shapely	Geometric operations
folium	Interactive map
geopy	Coordinate utilities

See requirements.txt for the full dependency list.

References & Inspiration

• Revealing influencing factors on global waste distribution via deep-learning based dumpsite detection from satellite imagery — Sun & Yin et al., Nature Communications (2023): Source of the Global Dumpsite Dataset used for garbage detection training.

• Illegal Buildings Detection from Satellite Images using GoogLeNet and Cadastral Map — Ostankovich et al. (2018): Provided the building detection dataset (224×224 RGB patches, two classes) and motivated the approach of fusing visual detection with cadastral/spatial validation.

• OSMnx — Boeing (2017): The spatial feature extraction backbone of the pipeline, enabling programmatic queries of road networks, POIs, and building footprints from OpenStreetMap.

Thanks

• ... to Xian Sun and Dongshuo Yin for releasing the Global Dumpsite Dataset openly via Science Data Bank.
• ... to Vladislav Ostankovich for the illegal buildings satellite dataset.
• ... to the YOLOv7 team for their open-source object detection framework.
• ... to the contributors of OSMnx for making spatial queries from OpenStreetMap accessible in Python.
• ... to the open-source community for satellite tile APIs and geospatial tooling that made this project possible.

---

Created as a proof-of-concept for AI-assisted environmental monitoring using publicly available satellite and spatial data.