Introduction

Wildfires pose a critical threat to ecosystems, infrastructure, and human life. Timely and accurate detection is essential for effective intervention and mitigation. However, developing high-performing object detection models for wildfire detection is often constrained by the lack of labeled data and the time-intensive process of manual annotation.

This project presents an end-to-end AutoML pipeline for wildfire detection using a CI/CD/CT (Continuous Integration, Continuous Deployment, and Continuous Training) architecture. The pipeline automates the entire lifecycle of a detection model. It starts from raw image collection and continues through pre-labeling, human validation, training, distillation, quantization, and deployment.

Motivation

Manual labeling of wildfire imagery is time-consuming and error-prone. In addition, models degrade over time as environmental conditions and data distributions shift. Our system aims to continuously learn from new data using a scalable, semi-supervised approach. It automates as much of the machine learning workflow as possible and involves human review only when necessary.

Key Features

  • Automated pre-labeling using YOLOv8 and Grounding DINO

  • Model matching and validation using IoU and confidence thresholds

  • Human-in-the-loop review for mismatches via Label Studio

  • Image augmentation to improve generalization

  • End-to-end training, distillation, and quantization

  • CI/CD/CT-compatible design for regular updates and retraining

Workflow Overview

  1. Data Collection
    Unlabeled wildfire images are collected from remote sensors and placed into a raw image directory.

  2. Pre-labeling (YOLO and Grounding DINO)
    Both models generate bounding boxes independently. YOLO is fast and lightweight. Grounding DINO supports natural language prompts.

  3. Matching
    Predictions from both models are matched using class name and IoU. Unmatched results are flagged for human review.

  4. Human-in-the-Loop Review
    Label Studio is used to manually verify or correct mismatched results.

  5. Augmentation
    Verified labeled images are augmented to enrich the dataset.

  6. Training
    A new YOLO model is trained on the augmented dataset.

  7. Distillation and Quantization
    The full model is distilled into a lightweight version and then quantized for deployment.

  8. Model Registry Update
    Trained models are stored in the registry and used for future pre-labeling.

This pipeline ensures scalability, adaptability, and model freshness without relying heavily on constant manual labeling. The integration of human review only when needed helps balance efficiency with accuracy.