Assessing the performance of machine learning algorithms for soil salinity mapping in Google Earth Engine platform using Sentinel-2A and Landsat-8 OLI data. Issue 2 (15th January 2022)

• Record Type: Journal Article
• Authors: Aksoy, Samet; Yildirim, Aylin; Gorji, Taha; Hamzehpour, Nikou; Tanik, Aysegul; Sertel, Elif
• Is Part Of: Advances in space research. Volume 69:Issue 2(2022)
• Page Start: 1072
• Page End: 1086
• Publication Date: 2022-01-15
• Subjects: Soil salinity -- Electrical conductivity -- Machine learning -- Cross-validation -- Google Earth Engine; Space sciences -- Periodicals; Astronautics -- Periodicals; Geophysics -- Periodicals; 500.505
• DOI: 10.1016/j.asr.2021.10.024
• Languages: English
• ISSNs: 0273-1177

Abstract

Highlights: Earth Engine platform facilitated the analysis of multi-sensor data for soil salinity mapping. Landsat- 8 and Sentinel-2A derived soil salinity maps were compared. Machine learning regression methods provided good results to map soil salinity. Abstract: Soil salinization caused by natural and anthropogenic factors is an important environmental hazard especially in arid and semi-arid regions of the world. Accumulation of salts in the soil is a major threat to crop production and global agriculture; therefore, rapid and precise detection of salt-affected lands is highly critical for preserving soil sustainability and supporting food production. Advancement in remote sensing techniques and machine-learning algorithms has started to contribute to fast and large-scale monitoring and mapping of soil salinization throughout the world. This paper aims to analyze the performance of three different machine-learning algorithms to map soil salinity using Landsat-8 OLI, Sentinel-2A satellite images, and ground-based electrical conductivity (EC) measurements with the aid of Google Earth Engine (GEE) platform. Classification and regression trees (CART), random forest (RF), and support vector regression (SVR) methods are implemented to create a correlation between ground measurements and satellite-derived environmental variables or spectral indices. After selecting the optimum five variables including wetness band, three soil salinity indices, and one vegetation index, soil salinity maps are generated in three machine-learning algorithms. The output soil salinity map in RF algorithm demonstrated the most reliable spatial distribution of various soil salinity classes in the selected study area. Despite CART provided slightly better prediction of soil salinity with R-squared (R 2 ) of 0.98 for Sentinel-2A data, and 0.96 for Landsat-8 OLI data in comparison with accuracy results of RF technique with R 2 of 0.96 for Sentinel-2A data and 0.94 for Landsat-8 OLI data, the output map of RF model estimated more reliable salinity levels in salt crusts, agricultural lands, drainage areas, and swamps. The corresponding result highly matched with visual interpretation. Soil salinity maps derived from SVR algorithms by using various combinations of input variables displayed relatively poor estimation of soil EC values compared to the other two methods.