Gathering predictors of biodiversity change and reconstructing land cover history in Central Apennines using machine learning and remote sensing data

The selection of methods for image processing and software functionality is crucial for monitoring Earth's landscapes. This work presents the use of Machine Learning (ML) methods for remote sensing (RS) data processing. The aim is to perform cartographic analysis of land cover changes with a case of central Apennines, Italy. Technically, we present a ML-based classification method using GRASS GIS software integrated with Python library Scikit-Learn. Image processing using ML methods was investigated by employing the algorithms of GRASS GIS. The data are obtained from the United States Geological Survey (USGS) and include a time series of Landsat 8-9 OLI/TIRS satellite images. The operational workflow of image processing includes RS data processing. The images were classified into raster maps with automatically detected categories of land cover types. The approach was implemented by using a set of modules in scripting language of GRASS GIS, including for non-supervised classification used as training dataset of random pixel seeds. The ML classifiers were used to detect changes in land cover types derived from images. The results show different vegetation conditions in spring and autumn periods. Unlike the existing methods of image classification, ML considers the differences among the spectral reflectance of pixels when modelling topology of patches. Other advantages are that ML uses data on texture and spectral features to measure the similarity of neighbouring landscape patches during the process of generating random decision trees. This study demonstrated the benefits of ML for cartography, RS data processing and geoinformatics.