Automated Selection of Sentinel-2 Spectral Bands for Fire Detection

This work investigates the occurrence, parameters, and consequences of fires in satellite images that can be directly exploited by several combinations of different multispectral image bands.

When we want to understand the semantics of a recorded digital image, we can cut it into smaller-size image patches and routinely classify these image patches via common unsupervised or supervised image classification techniques. In addition, when we include some clever interactive learning steps to attach semantic labels to the hitherto mathematically classified image patches, this should allow for a highly automated and powerful image understanding procedure.

On the other hand, starting with simple examples, the application-oriented analysis and exploitation of Sentinel-2 images can combine and display selected colour bands and their combinations. This has already been discussed in many (mostly GIS-oriented) publications ranging from the straightforward assignment of directly available pseudo-RGB colour bands up to advanced machine learning approaches for the extraction of content-related information (such as image feature descriptors or indices) [1-4]. Further, we will also refer to a few recently published advanced information extraction tools [5-10].

As an alternative to these (mostly conventional) image classifications, we describe a powerful semantic image classification technique that starts with the generation of topics (instead of classes) that was originally described by [11].Here, the resulting topic maps can be further combined and be used for colour band displays and their interpretation. When we combine the properties and capabilities of Sentinel-2 images with topic interpretation techniques, the most interesting question is whether a semantic interpretation based on topic maps outperforms common feature-based approaches.

To this end, we selected several Sentinel-2 multi-band images comprising different geographical areas affected by fires. This presentation shows the actual impact of various band combinations of Sentinel-2 channels and illustrates the band-dependent appearance of Fires, Smoke, Clouds, and other specific categories linked to the investigated continental areas. The basic algorithm being used for this investigation is Latent Dirichlet Allocation that has been applied as a data mining tool to discover patterns in the data, combined with automated band selection approaches.

The combination of automated image classification and multi-colour visualization seems to be an interesting alternative to Deep Learning.

[1] https://gisgeograpy,com/sentinel-2-bands-combinations

[2] https://worldofittech.com/sentinel2-bands-and-combinations

[3] https://giscrack.com/list-of-band-combinations-in-sentinel-2a

[4] https://eo4geocourses/github.io/IGIK_Sentinel2-Data-and-Vegetation-Indices

[5] A. Revill et al., The Value of Sentinel-2 Spectral Bands for the Assessment of Winter Wheat Growth and Development, Remote Sensing, 11(17), 2018.

[6] K. Kowalski et al., A generalized framework for drought monitoring across Central European grassland gradients with Sentinel-2 time series, Remote Sensing of Environment, 286, 2023.

[7] M.K. Vanderhoof et al., High-frequency Time Series Comparison of Sentinel-1 and Sentinel-2 Satellites for Mapping Open and Vegetated Water Across the United States, Remote Sensing of Environment, 288, 2023.

[8] E.C. Rodriguez-Garlito et al., Mapping Invasive Aquatic Plants in Sentinel-2 Images Using Convolutional Neural Networks Trained with Spectral Indices, JSTARS, 16, pp.2889-2899, 2023

[9] Z. Chen, et al., Mapping Mangrove Using a Red-Edge Mangrove Index (REMI) Based on Sentinel-2 Multispectral Images, TGRS, 61, pp.1-11, 2023.

[10] A. Temenos, Interpretable Deep Learning, GRSL, 20, pp.1-5, 2023.

[11] D.M. Blei, et al., Latent Dirichlet Allocation, Journal of Machine Learning Research, 3, pp.993-1022, 2003.