Automated delineation of salt-affected lands and their progress in coastal India using Google Earth Engine and machine learning techniques

Abstract

Assessment of salt-affected land (SAL) is still a major challenging task worldwide, especially in developing nations. The advancement of remotely sensed digital satellite images of different spectral bands has enabled the assessment of soil salinity. Sentinel-2 and Landsat 8 and 5 images of 2020, 2015 and 2009 and Shuttle Radar Topographical Mission data of 2014 were obtained from the Google Earth Engine data catalogue. Twenty spectral indices have been used which include four vegetation indices, twelve soil salinity indices, four topographical characteristics and their spectral bands. The Random Forest model was used to detect SAL. A total of 593 soil samples were used in the model. Of the electrical conductivity values of samples collected in the field, 70% of the soil samples were used for the model training, and the remaining 30% were used for validation. Also, fivefold cross-validation was carried out to validate the model prediction. The predicted SAL extent identified during 2020 was 134.4 sq. km with an overall accuracy of 93% using fivefold cross-validation. In 2015 and 2009, the total SAL was 128.42 and 120.41 sq. km, respectively. The total SAL has increased by 11.6% during the study period. The present study demonstrated the strength of remote sensing techniques to assess the SAL, which will help quantify the unproductive lands at the state or national level for reclamation or other productive use.