Deep Learning for Wildfire Detection Using Multi-Sensor Multi-Resolution Satellite Images

QC 20241118

Abstract

In recent years, climate change and human activities have caused increasing numbers of wildfires. Earth observation data with various spatial and temporal resolutions have shown great potential in detecting and monitoring wildfires. Sensors with different spatial and temporal resolutions detect wildfires in different stages. For low spatial resolution and high temporal resolution satellites, they are mostly used in active fire detection and early-stage burned area mapping because of their frequent revisit. While these products are very useful, the existing solutions have flaws, including many false alarms due to cloud cover or buildings with roofs in high temperatures. Also, the multi-criteria threshold-based method does not leverage rich temporal information of each pixel at different timestamps and rich spatial information between neighbouring pixels. Therefore, advanced processing algorithms are needed to detect active fires. For medium spatial resolution and low temporal resolution satellites, they are often used to detect post-fire burned areas. Optical sensors like Sentinel-2 and Landsat-8/9 are commonly used but their low temporal resolution makes them difficult to monitor ongoing wildfire as they are likely to be affected by clouds and smoke. Synthetic Aperture Radar (SAR) satellites like Sentinel-1, ALOS-2 and RADARSAR Constellation Mission (RCM) can penetrate through the cloud and their spatial resolutions are around 30 meters. However, limited studies have compared the effectiveness of C-band and L-band data and investigating the usage of compact polarization on burned area mapping.

The main objective of this thesis is to develop deep learning methods for improved active fire detection, daily burned area mapping and post-fire burned area mapping utilizing multi-sensor multi-resolution earth observation images. 

 Temporal models such as Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Transformer networks are promising for effectively capturing temporal information embedded in the image time-series produced by high temporal resolution sensors. Spatial models, including ConvNet-based and Transformer-based architectures, are well-suited for leveraging the rich spatial details in images from mid-resolution sensors. Furthermore, when dealing with image time-series that contain both abundant temporal and spatial information, spatial-temporal models like 3D ConvNet-based and Transformer-based models are ideal for addressing the task. 

In this thesis, the GRU-based GOES-R early detection method consists of a 5-layer GRU network that utilizes GOES-R ABI pixel time-series and classifies the active fire pixels at each time step. For 36 study areas, the proposed method detects 26 wildfires earlier than VIIRS active fire product. Moreover, the method mitigates the problem of coarse resolution of GOES-R ABI images by upsampling and the results show more reliable early-stage active fire location and suppresses the noise compared to GOES-R active fire product.

Furthermore, the VIIRS time-series images are investigated for both active fire detection and daily burned area mapping. For active fire detection, the image time-series are tokenized into vectors of pixel time-series as the input to the proposed Transformer model. For daily burned area mapping, the 3-dimensional Swin-Transformer model is directly applied to the image time-series. The attention mechanism of the Transformer helps to find the spatial-temporal relations of the pixel. By detecting the variation of the pixel values, the proposed model classifies the pixel at different time steps as an active fire pixel or a non-fire pixel. The proposed method is tested over 18 study areas across different regions and provides a 0.804 F1-Score. It outperforms the VIIRS active fire products from NASA which has a 0.663 F1-Score. For daily burned area mapping, it also outperforms the accumulation of VIIRS active fire hotspots in the F1 Score (0.811 vs 0.730). Also, the Transformer model is proven to be superior for active fire detection to other sequential models GRU and spatial models like U-Net. Additionally, for burned area detection, the proposed AR-SwinUNETR also shows superior performance over spatial models and other baseline spatial-temporal models.

To address the limitation of optical images due to cloud cover,  C-bBand data from Sentinel-1 and RCM, as well as L-band data from ALOS-2 PALSAR-2, are evaluated for post-fire burned area detection. To assess the effectiveness of SAR at different wavelengths, the performance of the same deep learning model is cross-compared on burned areas of varying severities in broadleaf and needleleaf forests using both Sentinel-1 SAR and PALSAR-2 SAR data. The results indicate that L-band SAR is more sensitive to detecting low and medium burn severities. Overall, models using L-band data achieve superior performance, with an F1 Score of 0.840 and an IoU Score of 0.729, compared to models using C-band data, which scored 0.757 and 0.630, respectively, across 12 test wildfires. For the RCM data, which provides compact polarization (compact-pol) at C-band, the inclusion of features generated from m-$\chi$ compact polarization decomposition and the radar vegetation index, combined with the original images, further enhances performance. The results demonstrate that leveraging polarization decomposition and the radar vegetation index improves detection accuracy for baseline deep learning models compared to using compact-pol images alone.

In conclusion, this thesis demonstrates the potential of advanced deep learning methods and multi-sensor Earth observation data for improving wildfire detection and burned area mapping, achieving superior performance across various sensors and methodologies.

urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-356334