• Journal Article

Learning Class-specific Spectral Patterns to Improve Deep Learning-based Scene-level Fire Smoke Detection from Multi-spectral Satellite Imagery

L. Zhao; J. Liu; S. Peters; J. Li; N. Mueller; S. Oliver


This paper introduces the Input Amplification (InAmp) module, a deep learning (DL) innovation designed to enable DL models to automatically learn class-specific spectral patterns for accurate fire smoke detection using satellite imagery. Traditional fire smoke detection methods depend on threshold values manually derived from satellite spectral bands, based on domain knowledge. These thresholds are challenging to generalise and integrate into existing DL models, which are not inherently equipped to learn such spectral patterns. InAmp addresses this gap by explicitly and automatically learning class-specific spectral patterns from multi-spectral satellite imagery. Functioning as an input pre-processing block, InAmp facilitates seamless integration with various DL architectures. The effectiveness of InAmp is demonstrated through its integration with Convolutional Neural Network (CNN) architectures and testing on two datasets: USTC_SmokeRS (MODIS, three bands) and Landsat_Smk (Landsat 5/8, six bands). Our results show improved fire smoke detection accuracy, and we also present how InAmp successfully extracts class-specific spectral patterns, showcasing its potential in advancing fire smoke detection based on satellite imagery.

Read full Publication