Land use/land cover classification with Landsat-8 and Landsat-9 satellite images: a comparative analysis between forest- and agriculture-dominated landscapes using different machine learning methods

• Published in: Acta geodaetica et geophysica 2022, Vol.57 (4), p.695-716
• Main Authors: Saralioglu, Ekrem, Vatandaslar, Can
• Format: Article
• Language: English
• Subjects: Agriculture; Artificial neural networks; Classification; Comparative analysis; Earth and Environmental Science; Earth Sciences; Farming systems; Forests; Geophysics/Geodesy; Image classification; Land cover; Land use; Landsat; Landsat satellites; Landscape; Machine learning; Mapping; Methods; Neural networks; Original Study; Remote sensing; Satellite imagery; Satellites; Spatial analysis; Support vector machines
• ISSN: 2213-5812; 2213-5820
• DOI: 10.1007/s40328-022-00400-9

The Landsat program, which started in 1972 with Landsat-1, continues today with its newest satellite, Landsat-9, launched on 27 October 2021. The Landsat-9 data have been freely distributed since 10 February 2022 on the Earth Explorer platform. However, no scientific study on Landsat-9 for land use/land cover (LULC) mapping has yet been published, focusing on specific eco-systems. Therefore, the present study investigates the potential of Landsat-9 images for LULC classification in forest and agricultural systems. To achieve this, we selected two study areas, i.e. Kaynarca (forest-dominated) and Hocalar (agriculture-dominated), from different ecoregions of Turkey. Then, we mapped their LULCs using Landsat-8 and Landsat-9 data with the Support Vector Machine, K-Nearest Neighbors (K-NN), Light Gradient Boosting Machine (LightGBM), and 3D Convolutional Neural Network (3D-CNN) methods. The classification accuracies were assessed with the F1-score, taking the stand-types maps of the case areas as reference. It was seen that the best maps were generated by the 3D-CNN method with accuracy rates of 88.0% for Kaynarca (Landsat-8) and 87.4% for Hocalar (Landsat-9) at the landscape level. Unlike other methods, 3D-CNN removed the “salt-and-pepper effect” on the maps providing better spatial structure for further analyses. Regardless of the satellite missions, the mapping accuracies for the “productive forest” and “agriculture” classes were > 90% for Kaynarca and Hocalar, respectively. The comparative results suggest that Landsat-9 offers satisfactory LULC maps with similar classification accuracies as Landsat-8 and can be effectively used as a freely available remote sensing resource in monitoring and mapping forest- and agriculture-dominated landscapes.