Skip to content
Back to Portfolio
DeepVision Wildfire - Real-time wildfire detection system

DeepVision Wildfire

Deep learning wildfire detection system for real-time smoke detection in Siberian forests

Client: Incom Group · Tomsk · Russia 🇷🇺
TensorFlowPythonJavascikit-learnSuperviselyOpenCV

Highlights

Situation

Catastrophic wildfires in Siberia required an automated system to detect and localize fires in vast, remote forest regions using real-time video feeds.

Task

Develop a deep learning-powered wildfire detection system leveraging object detection and temporal modeling across multi-source video datasets.

Action

Built a modular pipeline combining EfficientDet-D1 for spatial detection and a CNN-RNN hybrid classifier (InceptionV3 + LSTM) to track smoke evolution over time.

Result

Achieved 95.6% classification accuracy and 94% F-score, outperforming existing systems with reliable quasi-real-time inference at ~9 FPS.

Core Team

Viacheslav Danilov

Viacheslav Danilov

Research Supervisor

Tomsk Polytechnic University

Tomsk · Russia 🇷🇺

Nikita Laptev

Nikita Laptev

ML Engineer

Tomsk Polytechnic University

Tomsk · Russia 🇷🇺

Vladislav Laptev

Vladislav Laptev

Data Scientist

Tomsk Polytechnic University

Tomsk · Russia 🇷🇺

Andrey Kravchenko

Andrey Kravchenko

Software Developer

Tomsk Polytechnic University

Tomsk · Russia 🇷🇺

Olga Gerget

Olga Gerget

Professor

Tomsk Polytechnic University

Tomsk · Russia 🇷🇺

Dmitriy Sonkin

Dmitriy Sonkin

Director

Incom Group

Tomsk · Russia 🇷🇺

Overview

In response to the devastating 2019 Siberian wildfires, this project developed a high-accuracy wildfire detection system based on object detection and neural networks. The goal was to detect smoke in video footage and alert forest protection services in real-time.

The project combined EfficientDet object detectors with custom pre-processing and a novel CNN-RNN hybrid classifier. Not only did it outperform existing systems in classification and localization accuracy, but it also significantly reduced false alarms. The final model achieved a classification accuracy of 95.6% and an F1-score of 94%, making it a powerful tool for both civil protection and environmental monitoring.

Data

The system was trained on a diverse dataset of 550 video recordings (350 fire, 200 non-fire), sourced from multiple international wildfire monitoring centers:

From each video, frames were extracted, preprocessed, and annotated using Supervisely. Preprocessing involved grayscale conversion, dynamic feature extraction using frame differencing, and noise reduction via adaptive Gaussian filtering. This increased signal clarity and reduced false positives (Figure 1).

Figure 1. Sample wildfire detection from Nevada Seismological Laboratory video feed showing smoke plume identification.

Methods

The detection pipeline used a three-stage architecture designed for robust smoke detection across varied terrain and weather conditions:

  • Dynamic Feature Extraction: Frame differencing and adaptive thresholding to isolate motion patterns typical of smoke.
  • Object Detection: Multiple models were evaluated (EfficientDet-D0/D1/D2, Faster R-CNN, SSD MobileNet), with EfficientDet-D1 chosen for best localization and speed.
  • Hybrid Classification: A CNN-RNN model combining InceptionV3 for spatial features and LSTM for temporal analysis classified sequences of fire-related frames.
  • Clustering Algorithm: A unique clustering algorithm aggregated detection results across frames using intersection-over-union (IoU) similarity, improving localization by 6% and reducing false positives.

The Siberian Forest Protection Service deployment scenario demonstrates the system's capability to detect smoke in challenging conditions (Figure 2).

Figure 2. Wildfire detection in Siberian forest region demonstrating the system's real-time smoke identification capabilities.

Results

The wildfire detection system demonstrated robust performance across multiple evaluation metrics and test scenarios. The combination of advanced object detection, custom pre-processing, and temporal classification yielded highly reliable and accurate results.

  • Classification Accuracy: 95.6%, reflecting the system's overall correctness in identifying both fire and non-fire instances.
  • F1-score: 94%, indicating a strong balance between precision and recall – critical for minimizing missed fires and false alarms.
  • Recall: 97.2%, showing the system's excellent capability to detect almost all actual fire events.
  • Precision: 93%, indicating that the majority of alerts corresponded to real fire-related events.
  • Localization Accuracy (IoU): 62%, measuring how well the system identified the precise region of the fire.
  • Processing Speed: Approximately 9 FPS on a GeForce RTX 2080 Ti GPU, ensuring quasi-real-time deployment.

Notably, the application of dynamic feature extraction and temporal sequence modeling contributed to significant performance improvements:

  • •The clustering algorithm for prediction aggregation improved localization accuracy by 6%.
  • •The hybrid CNN-RNN classifier increased the F1-score by 9% over traditional static image classifiers.
  • •Dynamic smoke pattern analysis via frame differencing led to a 31% boost in recall and 21% gain in precision.

These results place the system ahead of comparable wildfire detection technologies such as Fuego and SmokeyNet, both in terms of detection accuracy and robustness under real-world video conditions.

Conclusion

This project showcases a scalable and robust wildfire detection system, outperforming current methods in both accuracy and speed. Its modular pipeline, spanning image pre-processing, real-time object detection, and temporal classification, can be adapted to various dynamic-object detection tasks, from medical diagnostics to robotics.

Future improvements could include edge deployment, drone integration, and domain adaptation for wildfire detection in different geographies and climates. The project has been implemented in commercial applications and presented at top academic conferences.

Keep Exploring

Industry AI/ML

Synthetic video with exact depth, alpha masks, optical flow and layered bokeh for training depth-aware video bokeh models, with a public dataset and a web demo.

November 2026
Medical Imaging

Ray-casting segmentation algorithm for convex anatomical structures in MRI, achieving up to 91.8% Dice score with millisecond-level runtime.

May 2019