Introduction

Soil salinization is a major ecological and socioeconomic problem that adversely affects global food productivity and projected to be more intense due to the impacts of climate change in coastal countries (Mukhopadhyay et al. 2021). Apart from global warming and associated sea-level rise, poor water management is an important factor in soil salinization in arid areas (Li et al. 2018). Researchers have indicated the area under salt-affected land (SAL) for different regions and also increasing globally, but yet there is a lack of information on the global SAL extent. As per FAO (2021) report on salt-affected soil, almost 73% of global land is affected due to soil salinity, that includes 424 mha of topsoil between 0 and 30 cm, and 833mha of subsoil between 30 and 100 cm. Of the different continents, Asia has the largest SAL followed by East and North Africa, and Europe has the least extent of SAL.

Soil salinization emerges as a key environmental constraint hampering soil productivity, agricultural sustainability, and food security, particularly in arid and semi-arid regions of the world (Cuevas et al. 2019). Soil fertility is defined by the ability to supply plant nutrients in sufficient levels for the agricultural crops, which in turn is affected by soil salinization. Annual losses of the agricultural sector as 27.3 million US dollars, either due to gradually reduced productivity of salinized lands or their disuse because of degraded soil quality (Qadir et al. 2006). The level of soil salinity will indicate the health of the soil, thus allow us to plan for the management measures, alternative crops, and systems. The conventional measure of soil salinity has been carried out by collecting the soil samples from the field and measuring through laboratory analysis in terms of electrical conductivity (Rhoades et al. 1999). It is not only time-consuming but does not permit us to assess the salinity at a larger scale at a periodical level. Large-scale studies on soil salinization are very much required to assess the soil quality at an administrative unit level for planning and devising policy framework.

Remote sensing data and methods are extensively being used to map soil salinity. Extensive research in soil salinity mapping has been carried out over the last three decades using satellite imagery. Advances in the availability of satellite data and analytical capabilities have paved the way for precise and timely assessment of soil salinity across different places and times. Numerous spatial models were tested to assess saline soils based on topographic information, climatic conditions, land use information, etc. (Aksoy et al. 2022; Wei et al. 2020). Several spectral indices of vegetation e.g. Normalized Differential Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), General Differential Vegetation Index (GDVI); salinity indices such as Normalized Difference Salinity Index (NDSI), Salinity Index (SI), SI 1, SI 2, SI 3, SI-I, Canopy Salinity Index (CRSI) (Aksoy et al. 2022; Taillie et al. 2019).

Remote sensing data with high and medium spatial resolution recently has become available. Various studies have found high potential of Sentinel-2 images for SAL modelling (Davis et al. 2019). The Sentinel-2 satellite provides high spatial, spectral, and temporal resolution images with multi-spectral bands. These images offer several aspects that can help to identify soil salinity more precisely (Bannari et al. 2018). For SAL modelling and mapping, researchers have also commonly used Landsat satellite series images (Scudiero et al. 2014; Nguyen et al. 2020). In both methods (Sentinel-2 and Landsat satellite images), SAL was identified and mapped using soil salinity indices based on various band reflectance values (McBratney et al. 2003). As a result of its smoothness, soil that has a surface salt layer has high reflectance values in the visible and near-infrared bands (Elnaggar and Noller 2009; Meier et al. 2018). The spectral reflectance of the image bands can be influenced by several variables, including surface roughness, soil colour, wetness, and physical and chemical characteristics (Asfaw et al. 2018). The spectral reflectance of saline soil with a puffy black surface crust is lower (Davis et al. 2019).

Different modeling techniques used in these studies include partial least squares regression (PLSR) (Fan et al. 2015; Bai et al. 2018), Multiple Linear Regression (MLR) (Li et al. 2019, 2021), Support Vector Regression (SVR) (Wu et al. 2018), Artificial Neural Networks (ANN) (Farifteh et al. 2007), Random Forest (RF) (Nabiollahi et al. 2021). A large number of studies have shown that machine learning methods (ANN, SVM, RF) achieve high predictive accuracy compared to other methods, especially RF (Wu et al. 2018; Wang et al. 2021; Aksoy et al. 2022). The advantages of RF application in remote sensing are it works effectively over a large number of datasets, it can control the large number of variables without deletion, it evaluates the importance of each variable in the classification, and it produces an accurate internal generalization error estimate (OOB error), calculates distances between pairs that can be used to find outliers, it is reasonably resistant to noise and outliers, and it requires less computing power than other tree ensemble approaches (e.g. Boosting).

India has 2,956 Mha of saline soil and 3.77 Mha of sodic soil (Arora and Sharma 2017). 16.84 million tonnes of agricultural production annually lose due to soil salinity in India (Mandal et al. 2018). The projected population growth shows India will need about 311 and 350 million tonnes of grains in 2030 and 2050, respectively, to feed about 1.43 and 1.8 billion people, respectively (Kumar et al. 2016). Estimations show that additionally each year 10% of the land becomes salinized and around 50% of the arable land will become saline by 2050 (Kumar and Sharma 2020). Climate change may enhance the level of soil salinity especially in densely populated nations due to sea-level rise and make it more vulnerable.

Soil salinity has been assessed by Indian researchers using both traditional and remote sensing data (Narjary et al. 2019; Bhardwaj et al. 2019; Kumar et al. 2015), but these studies were in bits and pieces, mainly focused on spectral indices, followed by regression analysis. Few studies has focused on machine learning application in detecting soil salinity using different spatial resolution satellite images. Several studies on the comparative assessment of different spatial resolution satellite images such as Landsat TM (30 m), SPOT (5 m), IKONOS (1 m) in assessing soil salinity showed that IKONOS data provide much better prediction accuracy than SPOT and Landsat TM (Dwivedi et al. 2008; Eldeiry and Garcia 2008).

Growing SAL is a major environmental issue in developing and densely populated regions like India, as land availability is the limiting factor. Though researchers have used different automated techniques, regional studies in relation to tropical and semi-arid regions are meagre. The global assessment of SAL provides an overall picture (Ivushkin et al. 2019 ), however, does not give the exact details required for management at a regional scale, due to limited sample points. This study aims to assess the SAL changing pattern using different resolution images in the semi-arid tropical region of the densely populated coastal District.

Materials and methods

Study area

The study area of the present study is the Thoothukdu district of Tamil Nadu, India. Extended between 8º 19’ N and 9º 20’ N latitude, 77º 40’ E, and 78º 10’ E longitude, coastal length and spatial extent of 163.5 km and 4621sq.km respectively (Fig. 1). The Gulf of Mannar is an important ecological important site in the district. It is home to 3,600 marine species of flora and fauna. The district is the largest salt producer in the state, producing 70% of the total salt of the state and 30% of the country’s salt requirements. The district experiencing the semi-arid tropical type of climate with many climate change impacts such as sea-level rise (Sheikh 2011), seawater intrusion, shoreline change (Satheeskumar et al. 2021), increase in salt-affected land (Selvam et al. 2013) and meteorological drought (Sheik and Chandrasekar 2011). The annual mean rainfall of the district is 661.6 m with a maximum and minimum temperature ranging from 29.5 to 40.5 °C and 18.4 to 26.7 °C, respectively.

Fig. 1
figure 1

Study area map showing sample point location in Thoothukudi District, Tamil Nadu, India

Full size image

A total of 593 soil samples were used in the present study. Of 593 soil samples, 258 Soil samples from different locations were collected from 30 July to 5 August 2020 for EC analysis (Fig. 1). The remaining samples were taken from the Indian Council of Agricultural soil database. Each soil sample’s geographic coordinate was measured and stored by GPS TDC 600 with less than 2 m positional accuracies. At each location of soil samples, four topsoils from the four corners of the quadrant were collected and mixed well. The samples were dried completely in the air and passed through a 2 mm sieve to remove non-soil material. Soil leachate was prepared at a 1:2.5 soil: water ratio, after that soil EC was determined with a digital multiparameter meter (Systronics EC - TDS meter 308) at room temperature 25 °C. EC values ​​were classified as non-saline (< 2ds/m), slightly saline (2–4 ds/m), moderately saline (> 4, < 8 ds/m), highly saline (8–16 ds/m)) and extremely saline. (> 16) (Brown et al. 1954; Aksoy et al. 2022). After the EC analysis of all soil samples in the laboratory, a shapefile was created using ArcGIS 10.5, the EC value, stored location detail, and corresponding salinity classes were assigned to each point in the attribute table of the shapefile.

Data used

This study used Sentinel-2 and Landsat-8 (Table 1) of 2020 satellite imagery from May to August month with less than 40% cloud cover available at the Google Earth Engine (GEE) platform (Table 1). GEE have an extensive geospatial dataset, including Sentinel, Moderate Resolution Imaging Spectroradiometer (MODIS), ASTER, Landsat, and SRTM imagery, as well as analytical abilities to handle a large number of dataset. Topographic maps (58 K/3,4,7,8; 58 H/9,13–15; 58 L/1-3.5; 58G/12,15,16) from the Survey of India (SOI) were utilized to demarcate the features like administrative boundaries, rivers, etc.

Table 1 Spectral bands of Sentinel-2 used in the present study
Full size table

Data Processing

Selected all satellite images were imported from the GEE data catalog to the GEE code editor section. Selected image collections were then filtered between May to August month, 2020 using .filterDate() script in the GEE code editor. A total of 7 images of Landsat-8 and 47 images of Sentinel-2 were used. The study region shapefile and soil sample shapefile were then uploaded through ‘assets’ section and imported to the code editor in GEE and using the .filterBounds() script, clipped the satellite image collection to the study area extent. Finally, using .median() script, calculated the median of all image collections falling under the May to August month 2020 to reduce the data volume and for faster analysis (Carrasco et al. 2019).

Selection of predictors

The spectral indices are important to assess soil salinity in arid and semi-arid areas (Aksoy et al. 2022). By considering the various studies and nature of the study, commonly used soil salinity indicators such as vegetation indices, and salinity indices were selected for the present study to generate a powerful combination in assessing soil salinity. Various spectral bands of Sentinel-2 such as B2, B3, B4, and B8 of 10 m spatial resolution and spectral bands of Landsat-8 such as B2, B3, B4, B5, B6, and B7 of 30 m spatial resolution were selected for the study. Other spectral bands of Sentinel-2 and Landsat-8 were not considered in the analysis to make the model prediction output of 10 and 30 m. Various spectral bands, vegetation indices, and soil salinity indices (Tables 1 and 2) were combinedly named as predictors. .select() expression were used to select particular spectral band and .expression() script is used to calculate all predictors. A total of 19 predictors (fifteen indices and four spectral bands) using Sentinel-2 (Figs. 3) and 21 predictors (fifteen indices and 6 spectral bands) using Landsat-8 and collected soil sample’s salinity class were used to assess soil salinity. .addBands() expression were used to add particular predictors into the predictor list to predict soil salinity.

Table 2 Spectral indices of vegetation, soil salinity, and topographical attributes used in the model
Full size table
Fig. 2
figure 2

Methodology flow chart of automated delineation of salt-affected land using Sentinel-2 and Landsat-8 satellite image

Full size image
Fig. 3
figure 3

Different soil salinity predictors used to derive salt-affected land for Thoothukudi District, India

Full size image

Random forest modelling

The RF model was used to predict soil salinity in the present study. 70% of the collected soil sample’s salinity category was used to train the model and the remaining 30% were used for validation purposes using the .filter() expression. 593 soil samples were used in the model. 178 soil samples were used for training and the remaining samples were used for the model validation. All the predictors and collected soil sample’s salinity class have been integrated into the model and using the ‘ee.Classifier.smileRandomForest()’ expression the model was performed. The hyperparameter was used to find the optimum number of trees and bag fraction with the highest training accuracy. Under different settings ranging from 1 to 500 number of trees at an interval of 10 and bag-fraction ranging from 0.1 to 0.9 with an interval of 0.1 were calculated to get the optimum number of trees and bag-fraction with the highest training accuracy. Through the bag-fraction method, unused samples can participate in making the decision tree creation process to evaluate each tree’s accuracy to improve model performance by considering the mean accuracy value of all trees. The model was trained using .sampleRegions() expression. All predictors and their importance in the model prediction were generated using “ee.Feature(null, ee.Dictionary().get(‘importance’))” expression.

Using .confusionMatrix() expression confusion matrix was calculated and .accuracy() expression the overall training accuracy was calculated. Similarly, validation accuracy for classified raster were calculated using the .errorMatrix() and .accuracy() expression. 5-fold cross-validation method, training and validation accuracy were calculated to validate the model performance. Very limited amount of points in moderately, highly, and extremely saline regions were selected because of less soil samples in these regions. Area of different classes of soil salinity of the classified raster was calculated using “ee.Image.pixelArea()” expression. The detailed flowchart methodology was given in Fig. 2.

593 samples were collected and used for the user’s accuracy and producer’s accuracy. F1 score of each class and micro-averaged, macro-averaged F1 score were calculated using the formula given below:

$$MiP=\frac{{\sum }_{i=1}^{r}{TP}_{i}}{{\sum }_{i=1}^{r}({TP}_{i}+{FP}_{i})}$$
(1)
$$MiR=\frac{{\sum }_{i=1}^{r}{TP}_{i}}{{\sum }_{i=1}^{r}({TP}_{i}+{FN}_{i})}$$
(2)

Both Mip and MiR, the denominator is the sum of all the elements (diagonal and off-diagonal) of the confusion matrix. TPi is the i-th diagonal element, FPi is the total of off-diagonal elements of the i-th row and FNi is the total of off-diagonal elements of the i-th column. Finally, the micro-averaged F1 (MiF1) score is computed as the harmonic mean of these quantities:

$$M{iF}_{1}=2\times \frac{{MiP}_{i}\times {MiR}_{i}}{Mi{P}_{i}+{MiR}_{i}}$$
(3)

To calculate the macro-averaged (MaF1), first need to evaluate precision (Pi) and recall (Ri) using the formula given below within each class, I = 1, …., r:

$${P}_{i}=\frac{{TP}_{i}}{({TP}_{i}+{FP}_{i})}$$
(4)
$${R}_{i}=\frac{{TP}_{i}}{({TP}_{i}+{FN}_{i})}$$
(5)

F1 score within each class (F1i ) is calculated by using the formula given below

$${F}_{1i}=2\times \frac{{P}_{i}\times {R}_{i}}{{P}_{i}+{R}_{i}}$$
(6)

The macro-averaged F1 (MaF1) score is defined as the simple arithmetic mean of F1i :

$$MaF1=\frac{1}{r}\sum _{i=1}^{r}{F}_{1i}$$
(7)

A popular technique for assessing the effectiveness of the RF models is the receiver operating characteristic (ROC) curve. The true positive rate and the false positive rate for all potential cut-off value are plotted on the ROC curve.

Results

Soil sample assessment for electrical conductivity

The EC values ​​of the collected soil varied from 0.02 to 72 ds/m, with an average value of 2.87 ds/m at a depth ranging from 0 to 20 cm. The study area’s major part belongs to agricultural land, also the dominance of saltpan close to the coastline. The maximum portion of the saline samples was located near the coastline of the study area.

Machine learning with Random Forest (RF) model

Two parameters, numbers of trees and bag-fraction were used in the hyperparameter tunning process. Result suggests that 20 number of trees with 0.6 bag-fraction has the highest training accuracy of 98.58% using Sentinel-2, whereas 30 number of trees with 0.5 bag-fraction has the highest training accuracy of 97.16% using Landsat-8. Variables of importance (VIMP) ranking were evaluated for the 19 predictors. Among the 19 predictors in the RF model using Sentinel-2, the top 10 important predictors were NDVI, SI5, SAVI, SI1, CRSI, SI2, SI4, GDVI, EVI and SI-I with VIMP scores of 8.87, 8.45, 8.36, 7.98, 7.66, 6.95, 6.55, 6.35, 6.26 and 5.7 respectively (Fig. 4). Among the 21 predictors in the RF model using Landsat-8 satellite image, the top 10 important predictors were CRSI, SI-III, SI-I, SI-II, NDVI, Blue, NDSI, GDVI, SI2 and Green with VIMP score of 8.7, 8.64, 8.11, 7.84, 6.94, 6.92, 6.87, 6.76, 6.75, and 6.36 (Fig. 4) respectively.

Fig. 4
figure 4

Variables of importance ranking of model predictors using Sentinel-2 and Landsat-8

Full size image

Spatial extent of salt-affected lands

Using the RF model of machine learning in the GEE platform SAL of Thoothukudi District was assessed (Fig. 5a and 5b). The estimated total SAL using Sentinel-2 in the Thoothukudi district in 2020 was 13,225 ha, including 7055 ha (Fig. 6) as moderately saline soil, 4070 ha as highly saline soil, and 2100 ha as extremely saline soil region. The overall training and validation accuracy was 98.97% and 98.31% respectively. The user’s accuracy of slightly saline, moderately saline and extremely saline was 100% followed by non-saline (97.81%). The producer’s accuracy of non-saline and extremely saline was 100%, followed by slightly saline (93.75%), moderately saline (92.85%) and highly saline (88.88%) (Table 3). Whereas using Landsat-8, the total assessed SAL was 12,446 ha, including 5249 ha as moderately saline, 3721 ha as highly saline and 3476 ha as extremely saline (Table 4) with a training and validation accuracy of 97.60% and 97.19% respectively. The user’s accuracy of slightly saline, moderately saline, highly saline and extremely saline was 100%, followed by non-saline soil (97.10) and the producer’s accuracy of non-saline and extremely saline was 100%, followed by highly saline (88.88%), slightly saline (87.5%), moderately saline (85.71%) (Table 5).

Table 3 Accuracy assessment of different class-wise soil salinity using Sentinel 2 of 2020
Full size table
Table 4 Spatial distribution of different soil salinity classes
Full size table
Table 5 Accuracy assessment of different class wise soil salinity using Landsat- 8 of 2020
Full size table
Fig. 5a
figure 5

Salt-affected land of 2020 using Landsat-8 satellite images

Full size image
Fig. 5b
figure 6

Salt-affected land in 2020 using Sentinel-2 satellite data

Full size image
Fig. 6
figure 7

Class wise extension of salt-affected land using Sentinel-2 and Landsat-8 satellite images

Full size image

F1 scores of both Landsat-8 and Sentinel-2 were calculated for non-saline, slightly saline, moderately saline, highly saline and extremely saline. Using the Sentinel-2 image, the highest F1 score was for extremely saline soil and slightly saline (1), followed by non-saline (0.99), moderately saline (0.96) and highly saline (0.94). The overall micro-averaged and macro-averaged F1 scores were 0.98 and 0.97 respectively. Whereas using Landsat-8 the highest F1 score was for Extremely saline (1), followed by non-saline (0.98), moderately saline (0.96), highly saline (0.94), and slightly saline (0.87). The overall micro-averaged and macro-averaged F1 scores were 0.98 and 0.95 respectively.

ROC and AUC for model validation

The RF model performance using two different datasets of Sentinel-2 and Landsat-8 is illustrated in Fig. 7. The figure shows that Sentinel-2 data is more capable of detecting salt-affected regions accurately than Landsat-8. The area under curve (AUC) value using the Sentinel-2 dataset was 0.813, whereas the AUC value using Landsat-8 data was 0.792 (Fig. 7).

Fig. 7
figure 8

Model validation using ROC curve

Full size image

Discussion

Multispectral optical remote sensing satellite images and machine learning have been used to assess the SALs in the coastal Thoothukudi district of Tamil Nadu, India. Combination of spectral bands and indices were used as predictors in the model. Two different spatial resolution satellites of Sentinel-2 and Landsat-8 were used in the RF model and the predicted result shows the significance of the model with an overall validation accuracy of 98.31% and 97.19% respectively.

Satellite remote sensing data plays a vital role in assessing EC as saline soil reflects specific surface reflectance and helps to predict saline soil (Aksoy et al. 2022). White salt crust areas refer to high to extreme saline areas (Pessoa et al. 2016). However, high spectral surface reflectance in every band does not necessarily indicate high salinity in multispectral satellite data. This makes it difficult to use multispectral bands and their spectral indices to assess SAL directly (Davis et al. 2019). However, the RF model evaluates the VIMP and represents the importance of each factor from highest to lowest importance in the model prediction. Several studies have suggested that the RF model is the best model with higher accuracy for monitoring soil salinity (Aksoy et al. 2022; Li et al. 2021).

Vegetation indices, salinity indices, and spectral bands of Sentinel-2 and Landsat-8 were considered as predictors in this study to assess SAL in 2020. Vegetation indices and salinity indices are the most widely used spectral indices to assess soil salinity (Aksoy et al. 2022; Li et al. 2021), but usually, spectral indices of vegetation and soil salinity response to EC are influenced by many factors which include resistance to salinity, percentage of vegetation cover, soil moisture and soil type (Peng et al. 2019). The results can vary significantly under different environmental settings. Since several other physical properties such as color, texture, and moisture affect the surface reflectance of saline soils, the unique soil salinity index may not provide an accurate estimate for all cases (Daliakopoulos et al. 2016).

Many strategies have been followed to get better accuracy results, such as (a) existing spectral indices selection based on the environmental condition of the study area, (b) new spectral indices creation based on local environmental conditions, (c) sensitive spectral indices selection based upon vegetation coverage, for example, salinity indices can be given more priority in the less vegetative cover region, whereas vegetation indices can be used in a high percentage of vegetation cover areas.

In the present study, the RF model was applied by considering all the predictors such as indices of vegetation, soil salinity, and spectral bands of Sentinel-2, Landsat-8 and collected soil samples EC. Vegetation indices, soil salinity indices and spectral bands have been used by many researchers worldwide for assessing SAL (Morgan et al. 2018; Aksoy et al. 2022; Allbed and Kumar 2013). Among the spectral band of Sentinel-2, green, blue, and red, among the Landsat-8, shortwave infra-red2, and red were found to be more important spectral bands in SAL prediction. Worldwide various studies on soil salinity assessment suggested that near-infrared, shortwave infrared, and blue bands were useful in identifying SAL (Nguyen et al. 2020; Khan and Sato 2001; Li et al. 2021).

Very less mix was observed between saline and non-saline regions while carrying out accuracy assessment by considering only saline and non-saline soil samples using both Sentinel-2 and Landsat images, also the overall accuracy was improved by 98.87 for both the Sentinel-2 and Landsat-8. Therefore, both Sentinel-2 and Landsat-8 were well capable in distinguishing saline regions from non-saline regions.

Prediction results obtained from sentinel-2, non-saline, and extremely saline classes had the highest producer’s accuracy (100%) followed by slightly saline (93.75%), moderately saline (92.85%) and highly saline (88.88%), whereas prediction result from Landsat-8 shows non-saline soil has the highest producer accuracy, followed by moderately saline soil (92.85%), highly saline (88.88%), slightly saline (87.5%) and extremely saline (71.43%) class. This shows the Sentinel-2 image is more capable of predicting different soil salinity classes with more accuracy as compared to the Landsat-8 image. The F1-score result also shows that Sentinel-2 is more capable of distinguishing different levels of soil salinity as compared to Landsat-8.

In addition, Sentinel-2 data spatial resolution is 10 m, it is much more capable of detecting soil salinity with better accuracy as compared to Landsat-8 as increasing spatial resolution gives better accuracy (Li et al. 2021). Saline areas are largely found in nearby mining areas, near the coast, saltpan, and river mouths of both Thamirabarani and Vaippar river. This region experiences meteorological drought, shoreline change, tsunamis, and seawater intrusion. SAL results show high to extreme soil saline areas mostly concentrated around the saltpan regions and the Thamirabarani river mouth nearby. As seawater enters through both the river (Thamirabarani and Vaippar river) to the inland region, the surroundings become moderately to highly saline. In addition, due to saline groundwater and the high evaporation rate, the salt is deposited on the soil surface. Various studies reported seawater dominance in the Thamirabarani river region (Selvam et al. 2013; Satheeskumar et al. 2021) and also various landuse such as sandy beaches, dunes and mudflats conversion to saltpan may be the reason of SAL in the region. Soil salinity varies by location due to the influence of geological, hydrological, biological, and climatic factors that affects the influence of the soil and water balance (Wang et al. 2020). The spectral range of salt-effected land does not have a single spectral signature to detect, also surface cover creates mixed spectral responses which cause difficulties in identifying saline soil. Also, land surface cover, soil EC and sodicity prevent in the assessment of soil salinity using moderate resolution satellite images and spectral indices of vegetation and soil salinity (Kılıc et al. 2022). Yearly long-term rainfall trends can be incorporated and it can help in understanding the SAL over the study region in the future. The SAL detection and demarcation will help local stakeholders to manage and develop alternate livelihood options, also limiting the saltpan region to safeguard other coastal ecosystems. The current study model can be applied to other coastal regions in delineating the SAL at the state or national level which may help in making action plans to control and manage SALs.

Conclusion

Combination of satellite images and machine learning techniques coupled with field-level soil EC measurements has been used to detect the SALs of different saline classes from non-saline to extremely saline class. Spectral indices of vegetation, soil salinity, and spectral bands were incorporated into the model. The Hyperparameter tuning was used to evaluate the optimum number of decision trees and bag-fraction of highest training accuracy to improve the prediction accuracy. To delineate SAL, Sentinel-2, and Landsat-8 both can be utilized, both satellites were more or less equally sensitive to different classes of SAL. But Sentinel-2 satellite images is little better in detecting the different level of saline soil regions as compared to Landsat-8. The model can assist in making regional or national scale policies to control and manage SAL regions for different livelihood options. To demarcate precisely SALs, high-resolution satellite image is necessary. In addition, assessing soil salinity using spectral indices, land use patterns, long-term rainfall, and seawater rise in the region can provide a better framework to make policy planning to support alternative livelihood options for the local people.