A Comprehensive Hyperspectral Image Dataset for Forest Fire Detection and Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Forestry

Background:

• Forest fires pose significant environmental and economic threats.
• Accurate and timely detection of forest fires is crucial for mitigation efforts.
• Existing hyperspectral datasets may lack the scale, quality, or diversity needed for advanced fire detection models.

Purpose of the Study:

• Introduce OHID-FF, a novel large-scale hyperspectral satellite image dataset for forest fire detection and classification.
• Provide a comprehensive resource for training and evaluating deep learning models in this domain.
• Establish a new benchmark for hyperspectral image classification tasks related to fire events.

Main Methods:

• Compilation of 1,197 hyperspectral images across 22 diverse Australian scenarios.
• Images feature 32 spectral bands with 10-meter spatial resolution.
• Detailed dataset preparation, including data sourcing, tiling, and annotation procedures.

Main Results:

• OHID-FF offers superior data volume and imaging quality compared to existing fire datasets.
• Benchmark experiments reveal challenges for current methods in classifying OHID-FF data.
• Demonstrated the potential of deep learning models for fire detection and classification using the dataset.

Conclusions:

• The OHID-FF dataset is a valuable resource for advancing forest fire detection research.
• The dataset highlights limitations in existing methods and sets a new benchmark for hyperspectral classification.
• Future work can leverage OHID-FF for developing more robust and accurate fire management technologies.