Spatial-spectral feature extraction using multispectral linear unmixing for land use land cover change detection

Abstract

Multispectral imagery is essential in remote sensing for the Land Use Land Cover (LULC) classification task. However, open-access platforms like Landsat and Sentinel offer low spatial resolution, causing mixed pixel issues that affect classification accuracy. This paper proposes a representation space that integrates spectral and spatial information using linear unmixing. The representation space is used to train a supervised classifier, and then classification maps are employed in LULC change detection and quantification. The proposed framework for LULC change detection allows the use of satellite images with different spatial and spectral resolutions. The representation space is tested on multispectral versions of benchmark datasets (Salinas, Indian Pines, Kennedy Space Center, and Botswana), obtaining classification accuracies up to 96 %. The proposed representation space significantly improves classification accuracy (up to 11 %) and outperforms existing multispectral classification approaches in the literature. Finally, the LULC change detection framework is tested on a time series of Landsat 8 and Sentinel-2 images corresponding to a drainage basin located in the eastern region of Antioquia, Colombia. The main advantage of this framework is that it does not modify the spatial resolution of the images in the dataset, favoring the integration of images from multiple sensors and proving useful for change detection in regions with limited data. The proposed framework for LULC change detection achieved an accuracy of 90 %.