Automation of image processing through ML algorithms of GRASS GIS using embedded Scikit-Learn library of Python

• Published in: Examples and counterexamples Vol. 7; p. 100180
• Main Author: Lemenkova, Polina
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2025; Elsevier
• Subjects: Cartography; Computer Science; Earth Sciences; Environmental Sciences; Environmental studies; Geoinformation; Geospatial analysis; Humanities and Social Sciences; Image Processing; Informatics; Machine Learning; Programming Languages; Remote sensing; Satellite image; Sciences of the Universe; Signal and Image Processing; Statistics; Cartography; Machine learning; Satellite image; 86-08; 86A04; Geoinformation; Geospatial analysis; 86A99; Informatics; 86A30; Remote sensing; Machine Learning
• ISSN: 2666-657X; 2666-657X
• DOI: 10.1016/j.exco.2025.100180

Image processing using Machine Learning (ML) and Artificial Neural Network (ANN) methods was investigated by employing the algorithms of Geographic Resources Analysis Support System (GRASS) Geographic Information System GIS with embedded Scikit-Learn library of Python language. The data are obtained from the United States Geological Survey (USGS) and include the Landsat 8 Operational Land Imager/Thermal Infrared Sensor (OLI/TIRS) multispectral satellite images. The images were collectedon 2013 and 2023 to evaluate land cover categories in each of the year. The study area covers the region of Nile Delta and the Faiyum Oasis, Egypt. A series of modules for raster image processing was applied using scripting language of GRASS GIS to process the remote sensing data. The satellite images were classified into raster maps presenting the land cover types. These include ‘i.cluster’ and ‘i.maxlik’ for non-supervised classification used as training dataset of random pixel seeds, ‘r.random’, ‘r.learn.train’, ‘r.learn.predict’ and ‘r.category’ for ML part of image processing. The consequences of various ML parameters on the cartographic outputs are analysed, such as speed and accuracy, randomness of nodes, analytical determination of the output weights, and dependence distribution of pixels for each algorithm. Supervised learning models of GRASS GIS were tested and compared including the Gaussian Naive Bayes (GaussianNB), Multi-layer Perceptron classifier (MLPClassifier), Support Vector Machines (SVM) Classifier, and Random Forest Classifier (RF). Though each algorithms was developed to serve different objectives of ML applications in RS data processing, their technical implementation and practical purposes present valuable approaches to cartographic data processing and image analysis. The results shown that the most time-consuming algorithms was noted as SVM classification, while the fastest results were achieved by the GaussianNB approach to image processing and the best results are achieved by RF Classifier.