Land cover classification using Land Parcel Identification System (LPIS) data and open source Eo-Learn library

Abstract

Depending on the development of remote sensing technologies, data at various spatial, temporal and spectral resolutions are obtained and play a key role in Earth investigation. Land cover/use maps are also created using these data. These data play an important role as a base for determining land cover and land use classes, detecting changes, monitoring forest, agriculture and wetlands, as well as updating the LPIS data used by countries. As well as being reference data in the LPIS update, current land cover and land use classification results with satellite images also provide statistical information about which class has changed over the years and how much the class or classes need to be updated, and will also provide the planning of the update to be made on a national scale. In this study, land cover classification for 2020 covering Çivril-Baklan Plain, which is between the borders of Denizli Province (Turkey), Baklan, Çal and Çivril Districts, was performed. The LPIS classes of the study area drawn in 2015 and the land cover classes as a result of the 2020 classification were compared as a result of the classification process, and particularly the areas with change were referenced in the LPIS update. In the classification process, the open source Eo-Learn library, which uses machine learning and deep learning algorithms in remote sensing studies and Sentinel-2 images that can be accessed directly in this library were used. The open source Eo-Learn library has been preferred because it provides great convenience to users in the classification process with ready-made machine learning models and easy access to satellite images, image processing steps (cloud masking, calculate index, feature extraction, temporal interpolation) are carried out in a certain workflow. Instead of land cover classification from a single image, 17 different time (multi-temporal) images between 01.02.2020 and 31.11.2020 were used. In the classification process, 12 different features were used for the image taken on each date, including NDVI, NDWI, NDBI indices and Tasseled Cap transformations, as well as (Blue) B02, (Green) B03, (Red) B04, (NIR) B08, (SWIR-1) B11, (SWIR-2) B12 bands. In the classification process, the Light Gradient Boosting Machines (LightGBM) algorithm in the Eo-Learn library and the physical blocks produced within the scope of the Land Parcel Identification System (LPIS) project were used as ground truth data; arable land, bare land, forest, artificial surface, shrubland, tree crops and water classes were created. The classification results were evaluated using the K fold (k = 5) cross validation method, with F1 score, recall, and precision values calculated for each class, and the overall accuracy %92.2. When the classification result is compared with the LPIS data, it is seen that the arable land, shrub land, and water classes have changed. This study determined that these classes, in particular, need to be updated.