DL_XVIEW • Object Detection in Satellite Imagery

Deep Learning–based Object Detection system for satellite imagery using YOLOv8. This project applies state-of-the-art computer vision models to detect objects such as airplanes, ships, vehicles, bridges, and more from aerial or satellite images. Data sources include xView and DOTA datasets.

Goal: Detect and classify multiple objects in high-resolution satellite imagery Model: YOLOv8 (axis-aligned detection; DOTA's oriented boxes are converted to horizontal boxes)

---

Features

• Object detection using YOLOv8 (Ultralytics)
• Dataset conversion utilities for xView / DOTA → YOLO format (convert_all_to_yolo.py)
• Interactive web-based prediction UI (ui/index.html, served by FastAPI)
• Training, validation and metric-plotting scripts
• Device selection follows CUDA → MPS → CPU, so it runs on NVIDIA, Apple Silicon and CPU

Project Structure

dl_xview_yolo/
├── scripts/
│   ├── config.py                  # Shared paths + device selection
│   ├── convert_all_to_yolo.py     # Converts xView/DOTA to YOLO format
│   ├── train_yolo.py              # Training script
│   ├── val_yolo.py                # Validation / mAP evaluation
│   ├── predict_yolo.py            # Inference (CLI) + FastAPI UI
│   └── plot_training_metrics.py   # Plots results.csv into PNG charts
│
├── ui/
│   └── index.html                 # Front-end for predictions
│
├── requirements.txt
├── LICENSE
└── README.md

Data (data/), converted labels (yolo_data/) and training outputs (runs/) are generated locally and excluded via .gitignore.

---

Installation & Setup

Recommended: Python 3.10 or higher

git clone https://github.com/Yigtwxx/dl_xview_yolo.git
cd dl_xview_yolo

python -m venv venv
source venv/bin/activate      # macOS / Linux
# venv\Scripts\activate       # Windows

pip install -r requirements.txt

Where the data lives

By default every script reads and writes inside the repository (data/, yolo_data/, runs/). To keep datasets elsewhere, set one environment variable:

export DL_XVIEW_ROOT=/mnt/data/dl_xview     # macOS / Linux
# set DL_XVIEW_ROOT=D:\dl_xview             # Windows

---

Dataset Preparation

Supported Datasets

• xView Dataset
• DOTA Dataset

Expected raw layout under data/:

data/
├── xview_raw/
│   ├── images/train/           # or train_images/
│   └── xView_train.geojson     # .geojson or .geojson.gz
└── dota_raw/
    ├── images/train/
    └── labelTxt-v1.5/train/

Convert to YOLO Format

python scripts/convert_all_to_yolo.py

Useful options:

# Custom source/output directories
python scripts/convert_all_to_yolo.py --src data --out yolo_data

# Convert a single dataset
python scripts/convert_all_to_yolo.py --datasets dota

# Regenerate data.yaml over an existing one
python scripts/convert_all_to_yolo.py --overwrite-yaml

What it does:

• Parses xView GeoJSON properties (bounds_imcoords, bbox, xmin/ymin/...)
• Converts DOTA's 8-point oriented boxes into axis-aligned YOLO boxes
• Normalizes 16-bit / single-channel GeoTIFFs into 8-bit JPEG
• Writes images/{train,val}, labels/{train,val} and data.yaml

Class layout: xView uses ids 0–59 (type_id - 1) and DOTA's 15 classes are offset to 60–74, so both datasets can share one labels/ directory without colliding. Total nc: 75.

Split: ~20% validation, chosen by an md5 hash of the filename so the split is identical on every run.

---

Training

python scripts/train_yolo.py

Options:

python scripts/train_yolo.py --epochs 100 --imgsz 1024 --batch 4 --name exp1
python scripts/train_yolo.py --weights runs/train/weights/last.pt --resume

Defaults: yolov8m.pt, AdamW at lr0=1e-4, cosine LR, imgsz=1024, batch=4, mosaic closed for the final 30 epochs, early stopping at patience=30. Checkpoints land in runs/train/<name>/weights/.

---

Validation

python scripts/val_yolo.py                       # auto-finds the latest best.pt
python scripts/val_yolo.py --weights runs/train/exp1/weights/best.pt --split val

Plot training metrics

python scripts/plot_training_metrics.py                        # auto-find results.csv
python scripts/plot_training_metrics.py --run-dir runs/train/exp1

Charts are written to <run-dir>/plots_matplotlib/.

---

Inference

Web UI (default)

python scripts/predict_yolo.py

Opens http://127.0.0.1:7860/ with a drag-and-drop panel that returns the annotated image. Weights are auto-discovered from runs/train/**/weights/best.pt.

Command line

python scripts/predict_yolo.py --no-ui --source data/test_images --conf 0.25

---

Example Results

Metric	Value
[email protected]	0.54
[email protected]	0.36
Precision	0.67
Recall	0.71

Measured on the validation split of an xView-only run. Metrics vary with the dataset mix and training configuration; see runs/train/**/results.png.

---

Model Weights

Trained weights are not tracked in Git (*.pt is in .gitignore). After training, best.pt lives at:

runs/train/<name>/weights/best.pt

---

Requirements

• Python ≥ 3.10
• PyTorch ≥ 2.0
• Ultralytics YOLOv8
• OpenCV, Pillow, NumPy, tqdm
• Matplotlib
• FastAPI + Uvicorn (only for the prediction UI)

---

License

This project is licensed under the MIT License. See LICENSE for details.

---

Author

Yiğit Erdoğan

Focus Areas: Deep Learning • Computer Vision • Data Science

• Mail: [email protected]
• LinkedIn: yigit-erdogan0
• GitHub: @Yigtwxx

---

Support

If you find this project useful, please consider giving it a ⭐ on GitHub!