Multi-level feature fusion networks for smoke recognition in remote sensing imagery

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Forestry and Fire Management

Background:

• Accurate smoke detection in remote sensing is crucial for forest fire monitoring, especially within Internet of Things (IoT) systems.
• Existing methods struggle with complex scenarios, including variable smoke appearances, cluttered backgrounds, and smoke-like phenomena (clouds, haze), leading to missed detections and false alarms.

Purpose of the Study:

• To develop a novel deep learning framework, the Multi-level Feature Fusion Network (MFFNet), for robust and accurate smoke detection in remote sensing images.
• To overcome the limitations of current smoke detection techniques by enhancing feature extraction and fusion for improved discrimination against background noise and similar-looking phenomena.

Main Methods:

• Utilized a pre-trained ConvNeXt model for multi-scale feature extraction from remote sensing images.
• Incorporated an Attention Feature Enhancement Module to refine multi-scale features, emphasizing discriminative smoke attributes.
• Employed a Bilinear Feature Fusion Module for feature integration, reducing background interference and contrastive feature learning for enhanced robustness.

Main Results:

• MFFNet achieved a high accuracy of 98.87% on the benchmark USTC_SmokeRS dataset.
• Demonstrated a 94.54% detection rate on the extended E_SmokeRS dataset, with a notably low false alarm rate of 3.30%.
• Outperformed existing methodologies in smoke recognition within complex remote sensing imagery.

Conclusions:

• MFFNet presents a significant advancement in smoke detection technology for forest fire monitoring.
• The proposed framework's multi-level feature fusion and contrastive learning approach effectively addresses challenges posed by complex environments and visual similarities.
• The model's high accuracy and low false alarm rate underscore its potential for practical implementation in early wildfire detection systems.