Pipeline Overview

The figure below illustrates the complete CI/CD/CT pipeline we developed for automated wildfire detection.

This pipeline is composed of the following stages:

1. Automated Data Collection – Continuously gathers unlabelled image data.

2. Pre-labelling – Uses YOLOv8 and Grounding DINO for generating bounding box predictions.

3. Matching & Filtering – Compares predictions to filter out mismatches.

4. Human-in-the-loop Review – Supports manual verification via Label Studio for unmatched samples.

5. Augmentation – Applies image transformations to enrich training data.

6. Training – Fine-tunes the YOLOv8 model using labeled and augmented data.

7. Distillation & Quantization – Optimizes the trained model for lightweight deployment.

8. CI/CD Integration – Final models are versioned and registered for deployment.