Automated classification of heat sources detected using SWIR remote sensing

Abstract

The potential of shortwave infrared (SWIR) remote sensing to detect hotspots has been investigated using satellite data for decades. The hotspots detected by satellite SWIR sensors include very high-temperature heat sources such as wildfires, volcanoes, industrial activity, or open burning. This study proposes an automated classification method of heat source detected utilizing Landsat 8 and Sentinel-2 data. We created training data of heat sources via visual inspection of hotspots detected by Landsat 8. A scheme to classify heat sources for daytime data was developed by combining classification methods based on a Convolutional Neural Network (CNN) algorithm utilizing spatial features and a decision tree algorithm based on thematic land-cover information and our time series detection record. Validation work using 10,959 classification results corresponding to hotspots acquired from May 2017 to July 2019 indicated that the two classification results were in 79.7% agreement. For hotspots where the two classification schemes agreed, the classification was 97.9% accurate. Even when the results of the two classification schemes conflicted, either was correct in 73% of the samples. To improve the accuracy, the heat source category was re-allocated to the most probable category corresponding to the combination of the results from the two methods. Integrating the two approaches achieved an overall accuracy of 92.8%. In contrast, the overall accuracy for heat source classification during nighttime reached 79.3% because only the decision tree-based classification was applicable to limited available data. Comparison with the Visible Infrared Imaging Radiometer Suite (VIIRS) fire product revealed that, despite the limited data acquisition frequency of Landsat 8, regional tendencies in hotspot occurrence were qualitatively appropriate for an annual period on a global scale.