Mapping burnt areas using very high-resolution imagery and deep learning algorithms - a case study in Bandipur, India

Area of Science:

• Earth and Environmental Sciences
• Computer Science
• Remote Sensing

Background:

• Accurate burnt area (BA) mapping is vital for wildfire impact assessment, carbon emission estimation, and biodiversity loss analysis.
• Effective BA data aids in guiding restoration efforts and refining fire management strategies.

Purpose of the Study:

• To design and evaluate two deep learning architectures, Custom UNET and UNET-Gated Recurrent Unit (GRU), for burnt area classification.
• To compare the performance of the novel UNET-GRU hybrid model against the Custom UNET using PlanetScope data.

Main Methods:

• Development of two deep learning models: a Custom UNET and a UNET-GRU hybrid.
• Classification of burnt and unburnt areas using PlanetScope satellite data over Bandipur, India.
• Evaluation of model performance using metrics like Precision, Recall, F1-Score, Accuracy, IoU, Dice Coefficient, and ROC curves.

Main Results:

• Both models achieved high accuracy in burnt area classification.
• The UNET-GRU hybrid model consistently outperformed the Custom UNET, especially in Recall and spatial overlap metrics (IoU, Dice Coefficient).
• The UNET-GRU achieved a higher Area Under the Curve (AUC) of 0.98 compared to the Custom UNET's 0.96, indicating superior classification performance.

Conclusions:

• The UNET-GRU model demonstrates enhanced capacity for precise burnt area classification and spatial accuracy, making it a robust choice for mapping using very high-resolution data.
• Integrating GRU into the UNET architecture significantly improves the model's ability to capture spatial and contextual features.
• The novel UNET-GRU presents a promising approach for advanced burnt area mapping applications.