Using Pseudo-Color Maps and Machine Learning Methods to Estimate Long-Term Salinity of Soils

Abstract

Soil salinity assessment methods based on remote sensing data are a common topic of scientific research. However, the developed methods, as a rule, estimate relatively small areas of the land surface at certain moments of the season, tied to the timing of ground surveys. Considerable variability of weather conditions and the state of the earth surface makes it difficult to assess the salinity level with the help of remote sensing data and to verify it within a year. At the same time, the assessment of salinity on the basis of multiyear data allows reducing the level of seasonal fluctuations to a considerable extent and revealing the statistically stable characteristics of cultivated areas of land surface. Such an approach allows, in our opinion, the processes of mapping the salinity of large areas of cultivated lands to be automated considerably. The authors propose an approach to assess the salinization of cultivated and non-cultivated soils of arid zones on the basis of long-term averaged values of vegetation indices and salinity indices. This approach allows revealing the consistent relationships between the characteristics of spectral indices and salinization parameters. Based on this approach, this paper presents a mapping method including the use of multiyear data and machine learning algorithms to classify soil salinity levels in one of the regions of South Kazakhstan. Verification of the method was carried out by comparing the obtained salinity assessment with the expert data and the results of laboratory tests of soil samples. The percentage of “gross” errors of the method, in other words, errors when the predicted salinity class differs by more than one position compared to the actual one, is 22–28% (accuracy is 0.78–0.72). The obtained results allow recommending the developed method for the assessment of long-term trends of secondary salinization of irrigated arable land in arid areas.

1. Introduction

Secondary salinization of arable land is one of the important negative factors limiting agricultural production in many regions of irrigated agriculture [1]. Over the past 30 years, the number of lands exposed to salinization has increased 2.4-fold [2] and covers up to 20% of the world’s irrigated land reserve [3]. Anthropogenic impacts and climate change, causing the increased use of soil and water resources, has led to increased salinity and land degradation, especially in arid areas [4]. Moreover, it is supposed that by 2050 the salinity of arable lands in the world could exceed 50% [5]. In Kazakhstan, secondary salinization of irrigated arable land is widespread in the Aral Sea basin [6,7]. This factor has a large-scale impact on irrigated areas in the middle and lower parts of the Syr Darya river basin in the Turkestan and Kyzylorda regions. Significant areas of salinization are observed in all southern regions of the country. To assess them and to develop the measures to reduce the negative impact, various monitoring methods are applied, one of which is the use of remote sensing of the earth’s surface. Remote methods for assessing soil salinity have been widely discussed by the scientific community since the 1980s [8,9,10,11]. However, a reliable and universal method has not yet been proposed [12]. The main difficulties are related to the fact that remote sensing methods assess the surface layer of soil, while the influence of salt is revealed in the root layer [13,14,15], the state of which is rarely directly related to the spectral characteristics of the soil surface [16]. Such a relationship can be detected only in a number of cases considering the local conditions [17,18,19]. In recent years, the development of machine learning methods has facilitated a solution to such a problem [20,21,22,23].

Nevertheless, soil salinization is a recurring phenomenon, caused not only by the specific conditions of one season but also by the long-term stable factors leading to an increased or decreased salt content in some parts of the regions under consideration [24].

In this study, a model of soil salinity assessment taking into account long-term remote sensing data is proposed. There are very few works of this type in the scientific literature. Only old studies [17,24] and a relatively recent study, which, however, is based on a low-resolution satellite product [25], can be noted. We hope that this paper will cover this gap.

The main contributions of this work are as follows:

• A method of salinity assessment for cultivated and non-cultivated soils of arid zones based on long-term averaged values of vegetation indices and salinity indices is proposed;
• The method is based on remote sensing data of the optical spectrum range and machine learning algorithms;
• The method allows revealing the stable relations between characteristics of corresponding spectral ranges and salinity parameters. In this sense, it is relatively stable.

An analysis of the literature shows that such an approach, which takes into account the long-term averaged data of the optical range of high spatial resolution, is being proposed for the first time.

The paper consists of the following sections:

• The second section describes the research area, which serves as an example for testing the proposed model.
• The third section provides a literature review describing the methods of soil salinity assessment based on RS data and machine learning.
• In the fourth section, we describe the research methodology.
• In the fifth section, we presented the results obtained.
• In the sixth section, we discuss the results obtained.

We conclude with a description of the advantages and limitations of the method, as well as the anticipated objectives for future research.

2. Research Field

A large irrigation massif—“Hungry Steppe”—with a total area of about 10,000 km², is located in the basin of the transboundary Syrdarya River. The river is part of the Aral Sea basin and ranks second in Central Asia (after the Amu Darya) in terms of annual runoff—23 km³. By territory, the main irrigated arable areas of this irrigation massif belong to Uzbekistan (upper part) and Kazakhstan (lower part). After the collapse of the USSR, the upstream hydrosystems with reservoirs, in particular Kayrakkum (4.2 km³; Tajikistan) and Toktogul (19.5 km³; Kyrgyzstan), changed their operation modes from irrigation to energy. Therefore, part of the river flow was used for power generation in the cold periods. This action significantly reduced the water potential in the irrigation season. In addition, after the collapse of the USSR, the processes of drainage system degradation were intensified, manifested in the siltation of drains and the abandonment of drainage wells. As a result, the developed process of secondary salinization of the “Hungry Steppe” arable land has taken place in recent decades.

There are about 140,000 hectares of irrigated arable land in the Kazakhstan sector of the “Hungry Steppe” irrigation array. The main crops are as follows: cotton, wheat, maize, melons, alfalfa, rice, etc. Relatively insignificant areas are allocated to rice crops. These arable lands administratively belong to two districts of the Turkestan district of Kazakhstan: Zhetysai and Maktaaral (see Figure 1). The latter is considered an object of application of the methodology for the estimation of the average multiyear salinity of lands, proposed in this article.

It should be noted that the task of assessing the secondary salinity of irrigated arable lands is complicated by the fact that we have to deal with different crops, crop rotations and agricultural practices. A significant role is played by seasonal weather conditions, air temperature, and water content of the river, which is the source of irrigation. Moreover, as a consequence, there is a great variety of soil surface conditions.

3. Related Works

Morphological diversity of soils (different depths and salinity of soil waters), different modes of economic use, and differences in chemistry and origin of certain types of salinity—all these factors complicate the task of predicting the degree of salinity. Soil salinity is measured both in the laboratory and in the field by assessing the salt content in the soil or the electrical conductivity of the water extract. The electrical conductivity of the water extract is more closely related to the negative effect of salinity on vegetation; therefore, it is used more often. Depending on the salt content in the soil, salinity is usually divided into five or six classes [26,27,28] (see

Table 1. Threshold values of 6 classes of soil salinity and its impact on plant growth.

Salinity Class	EC1:5 Range for Loams (dS/m)	Impact on Crop Growth	Types of Crops Growing at a Given Level of Salinity
Non-saline	0–0.18	Minor	Cereals other than corn: vetch, alfalfa
Slightly saline	0.19–0.36	The yield of crops that are sensitive to salinity is reduced.	Cotton, timothy, cocksfoot, sweet clover, wheat
Moderately saline	0.37–0.72	The yield of most crops is decreasing.	Fodder swede, fodder cabbage, wheatgrass, sorghum
Highly saline	0.73–1.45	Only salt-tolerant crops give a satisfactory harvest.	Sugar beets, sunflowers, western wheatgrass, French ryegrass, awnless brome
Severely saline	1.46–2.90	Most salt-tolerant crops can produce satisfactory yields.	Forage grasses: tall wheatgrass, brome grass, Canadian hair grass, tall fescue
Extremely saline	>2.90	Unsatisfactory yield	Fitting varieties

N.B. EC1:5 is the electrical conductivity of the soil solution (one weight fraction of soil dissolves in five parts of water).

). When divided into five classes, the “severely saline” and “extremely saline” classes are combined.

Field studies are labor-intensive, so their wide application is limited. During a ground survey, it is necessary to take soil samples from different depths (up to 1–2 m). Soil salinization is based on the physical process of fluid infiltration through porous media. The spatial feature of the processes of the three-dimensional percolation processes is a contrasting mosaic, manifested on the ground surface in the form of different-scale patchiness of local salinization zones [29]. Therefore, it is extremely difficult to conduct a detailed mapping of large areas of arable land on the basis of ground surveys alone.

To assess the level of soil salinity on a regional and global scale, remote sensing (RS) data, including aerial photography, satellite sensing, or UAV data are used [8,11,30].

One of the traditional approaches is the development of empirical models based on the correlation of the optical spectral indices and the soil electrical conductivity. The application of spectral indices reflecting the level of soil salinity in one way or another is widely used in regional studies. The literature devoted to the results of the application of spectral indices is extensive, and some idea of the variety of used indices and the reliability of modeling on their basis can be obtained from review papers [31,32,33,34,35]. Such models have different indicators of the correlation coefficient, which can sometimes be quite high, but the approach itself, as we have already noted, has limitations associated with its application on a local scale. In addition, the use of single satellite images, timed to the dates of ground-based measurements, does not provide an opportunity to assess the seasonal and multiyear dynamics of salinity. The authors of [36] note that “Surface salinity is a highly dynamic process, causing identification constraints derived from the proper behavior of the salt features, spectrally, spatially, and temporally” (see page 2). In addition, salt detection can also be blurred by the presence of vegetation and other surface features, which might contribute to creating spectral confusions with the salt reflectance properties (ibid.). These limitations make it necessary to develop methods for identifying and mapping saline soils using more complex, integrated approaches. One of the promising directions is the application of machine learning (ML) methods [37]. Machine learning technologies help to solve the nonlinear problem of the relationship between soil properties and environmental factors [38]. Among the many machine learning algorithms for solving the problem of estimating the soil salinity, the classical algorithms, boosting models, and random forest are most commonly used. For example, ref. [39] applied a complex calculation system that combined ground data, spectral indices, a regression model, and machine learning elements to map the salinity of sandy soils. The authors of [40] conducted a comparative study of several machine learning algorithms for the same set of ground measurements and showed that of the three used algorithms (gradient boost machine, X gradient boost and random forest), XGBoost finds the best result, despite the fact that the difference coefficient’s R2 between the three models is small and amounts to a few percent. Similar studies [41], which applied four machine learning algorithms (support vector machine, back propagation neural network, extreme learning machine, random forest) to calculate salinity from Sentinel-2 data, confirmed the similarity of the results of all algorithms used, with some differences in the calculation of salinity at different soil depths. A comparison of the least absolute shrinkage and selection operator (LASSO), multiple adaptive regression splines (MARS), classification and regression trees (CART), stochastic gradient treeboost (SGT), and random forest [42] models showed that SGT and random forest were the best models. Other models also have some drawbacks, such as negativity and outliers, binarization, an unreasonable range of values, and instability during multi-period predictions (large fluctuations in the range of soil salinity values).

The research [43] considers the basin of the Keriya River, northwest China. For this area, using the radar data (PALSAR and Radarsat-2) and the SVM classifier with optimal parameters, an accuracy of 93.01% was achieved. There are other examples of the application of machine learning models using radar data [23,44,45], optical data [46], or their combinations [5,47].

Table 2. Application of machine learning models for estimation of soil salinity.

Regressor	Abbreviation	Refs.
XGBoost [48]	XGB	[40,47,49,50]
Random forest [51]	RF	[20,40,41,42,52,53,54,55]
Support vector machines [56]	SVM	[41,45,54,57,58]
Linear regression [59]	LR	[39]
Lasso regression [60]	Lasso	[42]
Feed forward neural networks [61]	FFNN	[41,54,62,63,64,65,66,67]
Extreme learning machine [68]	ELM	[41]

lists the machine learning algorithms most commonly used for solving the problem of salinity estimation using satellite data.

The number of such researches is steadily growing, confirming the perspective and acceptable accuracy of the recognition and mapping of saline lands with the use of mixed methods of satellite data processing and, in particular, the perspective of machine learning for solving such a non-trivial task as the recognition of different types of salinity on different types of soils. At the same time, the listed methods of soil salinity assessment operate with operational data and are applied, as a rule, to relatively small areas of the earth’s surface at a certain time period.

This is especially true for methods using low-altitude multispectral images obtained using UAVs [69,70,71]. An exception is the method described in [55], which is aimed at constructing a global map. The authors used the SVM, CART, and random forest algorithms to develop a global salinity map on the Google Earth Engine platform. The authors claimed that the accuracy of mapping the various classes of saline soils reached 67–70%.

However, this method, as shown in Figure 2, has limited accuracy for marginal regions, such as Kazakhstan. Along with the location map (1), the figure shows changes in the forecast of the model developed by the authors of this work, based on XGBoost using Landsat 8 data of optical range (2), and the assessment of salinity using an RF model based on temperature data (hereafter Mtemp) [55] (3). To assess the salinity of the area marked in the figure on 23 May 2022, 80 samples of soil were collected, and several regression models were built; the best of these models was XGBoost (R2 = 0.54). ② in Figure 2 shows the model estimates using the satellite data from 1, 9, 10, 17, and 26 April 2022. It should be noted that the model predictions from 9 and 10 April are distorted by cloudy weather.

It can be seen that the predictions of the XGBoost model vary significantly even within one month and that the XGBoost model and the Mtemp model estimate salinity differently, with the Mtemp model predicting only two levels of soil salinity, which is not consistent with field survey data. In other words, for extensive arid zones with a sharp change in climatic conditions, which are the foothills and adjacent areas of southern Kazakhstan, the methods proposed in the literature can be applied in relatively localized areas with similar conditions, or they will have low accuracy. This is due to seasonal fluctuations in humidity, long-term changes in the water content of the seasons [72,73], and changes in irrigation. As a result, soil salinity is often unstable with active processes of salinization–desalination [17].

To identify statistically stable patterns of soil salinization, the authors of this paper propose to use the data of multiyear observations. It should be noted that the number of available ground measurements in Kazakhstan is still not great. For example, in [55], the data from the ground-based measurements of electrical conductivity in the territory of the Republic of Kazakhstan are absent. Therefore, it is reasonable to use the statistical data of the Ministry of Agriculture (MA) of the country, which contains information on the percentage of saline lands and develops a method that can reveal the distribution of salinity in the southern regions of Kazakhstan taking into account these data.

4. Method

The proposed method is based on multiyear values of vegetation and salinity indices and statistical data on salinity provided by the Ministry of Agriculture (MA) of Kazakhstan. To verify the method, the available data from expert assessments of salinity of large areas of cultivated lands in southern Kazakhstan are used. The methodological scheme of the study includes the steps below (see Figure 3).

4.1. Building a Pseudo-Color Map

The effect of arable land salinity on the spectral characteristics of the underlying surface is mainly associated with the depression of agricultural vegetation and possible salt outcrops (salt crusts) during the period of maximum salinization of the surface soil layer at the end of the growing season. The quantitative characteristics of these phenomena vary greatly from year to year, which is caused by variations in weather and hydrological conditions in individual seasons. More objective characteristics can be obtained as a result of the processing of the multiyear data and the transition to average long-term characteristics. In this case, the systematic effect of salinity will be reflected in quantitative characteristics.

In the case of low-dimensional (3- or 4-dimensional) phase spaces, which can be encoded in chrominance components, such as red-green-blue (RGB formula) or cyan-magenta-yellow-black (CMYK formula), it is possible to combine machine processing with expert analysis since the feature space is additionally encoded in a color image. In this case, the expert has the opportunity to choose the options for numerical processing and, as a first approximation, to control their effectiveness. In this study, a 3-dimensional feature space was used, in which 2 independent parameters were assigned to vegetation characteristics and one to the degree of salinity of the soil surface.

Monitoring of the satellite vegetation index values, for example, NDVI [74,75], during the vegetation season gives a fairly complete picture of the state of vegetation:

$$NDVI={\displaystyle \frac{nir-red}{red+nir}}$$

(1)

Parameterization of the vegetative growth curve with two independent variables automatically leads to the use of its two main points. The first is the maximum value of $NDV{I}_{max}$ on the graph: [calendar dates]—[NDVI]. The second is the area that the vegetation curve cuts off on its graph. If the same time period is always considered, for example, May–August, the area cut off by the vegetation curve is uniquely approximated by the mean value, $NDV{I}_{mean}$.

There are a large number of different salinity indices using satellite channels in the optical part of the spectrum (see, for example, [76]). They are most effective in the case of open soil, i.e., when the interfering (shielding) influence of green vegetation and mortuary mass is minimal. In the case of agricultural vegetation, open soil covers are typical for the sowing period in spring and in autumn after harvesting. For irrigated arable land, the maximum salinity is timed to the end of the growing season. The separate satellite scenes in the autumn period reflect fields in different phases. There are fields with ripening crops, either after harvesting or even after raising the seedbed. This condition is typical and related to the limited agricultural equipment, inability to sow the entire area in a limited time, as well as due to the use of different varieties and crops that have their own maturity.

The variety of soil surface conditions in the fields in the autumn period simultaneously presented on a separate satellite scene leads to the impossibility of uniform processing and interpretation of the results of estimating the salinity degree by means of salinity indices. In the case of multiyear data processing, the maximum values of the salinity index seem to be the most informative ones. The maximum values appear at the minimization of disturbing factors, and they characterize the underlying surface in the most uniform way in the analysis of multiyear data. The type of chosen salinity index can be different. Selecting the best index requires a separate study. In this study, the well-known VSSI salinity index [77] was chosen:

$$VSSI=2\ast green-5\left(red+nir\right)$$

(2)

Thus, three parameters were chosen to form a three-dimensional phase space with its own color image in the form of a composite pseudo-color image reflecting the peculiarities of vegetation dynamics and soil salinity: $NDV{I}_{max}$; $NDV{I}_{mean}$; $VSS{I}_{max}$. The following parameters were used to create the corresponding color formula, namely red-green-blue, to build the visual image so that the maximum of the VSSI index is fixed in the red channel at each pixel position:

$$VSS{I}_{max}=max\left(VSS{I}_1,VSS{I}_2,\dots ,VSS{I}_i,\dots ,VSS{I}_n\right)$$

(3)

The NDVI maximum is in the green channel:

$${NDVI}_{max}=max\left({\displaystyle \frac{ni{r}_1-re{d}_1}{re{d}_1+ni{r}_1}},{\displaystyle \frac{ni{r}_2-re{d}_2}{re{d}_2+ni{r}_2}},\dots ,{\displaystyle \frac{ni{r}_n-re{d}_n}{re{d}_n+ni{r}_n}}\right)$$

(4)

The average NDVI value is in the blue channel:

$${NDVI}_{mean}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\displaystyle \frac{ni{r}_i-re{d}_i}{re{d}_i+ni{r}_i}}$$

(5)

where $1\le i\le n$, and n is the number of Sentinel-2 satellite images annually from May to September, starting in 2016 and ending in 2022.

Since the purpose of the analysis was to rank the territory, in other words, a comparative analysis, the absolute values of the selected indices were not important. Therefore, the variability of each index was encoded by a byte of information (256 discrete values), and histogram processing of each channel was applied to form the most information-rich, from the point of view of an expert, visual image (see Figure 4).

4.2. Sampling of Local Zones with a Simple Spatial Organization of the Target Feature

Gradient fragments are used to construct a 5-class salinity classification according to [50]. Simple linear gradients determine the rank order of composite classes (colors) from non-salinity to very strong salinity (expert selection of gradient). The set of gradient fragments, through the centers of gravity of each class (color), forms the universal rank order (the ranks may vary slightly in each fragment) (see Figure 5).

4.3. Formation of a 5-Class Representation of Salinity

The created rank order of classes (colors) of the pseudo-color composite is used to construct a 5-class coarsening. Grouping of ranked colors into 5 classes attributed to 5 salinity classes is carried out. As a result, the formation of a 5-class salinity representation on the basis of a pseudo-color composite (see Figure 6) takes place.

4.4. Development and Training of the Classifier

When training the classifiers, a pseudo-color composite is used as the input data, and the salinity map corresponding to it is used as the target value (see Figure 7). During data preparation, the surface area is divided into fragments of 4 × 4 pixels, which forms the input vector of values 4 × 4 × 3, where 3 is the number of color channels (RGB). The target image is also fragmented into similar 4 × 4 areas, each of which is set to one of 6 colors. The color of the section is selected according to the maximum criterion. That is, for example, if the selected target fragment has 4 pixels of yellow, 2 pixels of green, and 10 of blue, then this fragment is colored completely in blue—there is no salinity.

4.5. Applying the Classifier to Specified Areas of the Earth’s Surface

The classifier model selected according to the maximum harmonic measure criterion (F1 score) is used to estimate the salinity of the entire significant area on the basis of its pseudo-color composite (see Figure 8).

The purpose of mapping is a numerical assessment of the relative state of the salinity of rural districts (or other fragments) in the area under consideration. The resulting map of the salinity of the district is analyzed for the balance of areas of salinity classes. It is estimated how close it is to the official data of the Ministry of Agriculture. Therefore, indirect validation of the results is performed. The above steps can be carried out repeatedly, so that the initial 5-class division is varied, and new training and validation is undertaken. The repetition of action continues until a given error level is obtained in the test section. Thus, the issue of the information content of the initial pseudo-color composite in the task of constructing a map of the average long-term salinity of arable land is solved.

To perform computational experiments, a software system was developed in Python using the numpy, sklearn, matplotlib, cv2, alive_progress, pickle, and tensorflow libraries, which provides the following:

• Cutting out part of the image;
• Segmentation of images by a given number of elements;
• Connecting image segments to form the final image;
• Converting the image fragments to one color;
• Color adjustment, which is necessary to exclude halftone pixels;
• Saving the prepared data sets and performing some additional functions related to obtaining the results and the application of machine learning models;
• Application of machine learning models.

Computational experiments were performed on a computer equipped with 32 GB of RAM, an Intel (R) Core (TM) i7-10750H processor, and a discrete video card, Nvidia GeForce GTX 1650 Ti.

The results of the machine learning models in this case were evaluated using the confusion matrix and the indicators of accuracy (Ac), precision (Precision), completeness (Recall) and harmonic measure (F1-Score):

$$Ac={\displaystyle \frac{N_t}{N}}$$

(6)

where $N_t$ is the number of correct answers, and N is the total number of possible answers of the model.

Precision:

$$P={\displaystyle \frac{TP}{\left(TP+FP\right)}}$$

(7)

Recall:

$$R={\displaystyle \frac{TP}{\left(TP+FN\right)}}$$

(8)

F1 score:

$$F1score={\displaystyle \frac{2\ast P\ast R}{\left(P+R\right)}}$$

(9)

where true positive (TP) and true negative (TN) are the cases of correct operation of the classifier, in other words, when the predicted class coincides with the expected one. Correspondingly, false negative (FN) and false positive (FP) are cases of misclassification.

It should be noted that when classifying the unbalanced datasets, when there are significantly more or fewer objects of some classes than others, it is possible to use two options for estimating the basic indicators. The first one, called the macro-average, assumes that each class, regardless of the number of its members, is equivalent in weight to the overall score of the indicator. This means that the quality metric is calculated within the objects of each class and then averaged over all classes. The second version of averaging, the micro-average, assumes that objects in all classes contribute equally to the quality metrics. Since it is important for us to consider the contribution of small classes, the macro-average will give a more objective estimate.

5. Results

The resulting pseudo-color composite is shown in Figure 8. A portion of this composite (40,000 pixels) was used to train and test machine learning models. In the process of conducting the computational experiments, nine machine learning models were used.

Table 3. Results of the operation of machine learning models on a test set of cultivated soil plots.

Machine Learning Model	F1_Score Macro Average	Accuracy
XGBoost classifier	0.83	0.89
Convolution neural network	0.84	0.89
Random forest classifier	0.76	0.84
KneighborsClassifier	0.76	0.84
SVC (kernel = “linear”, C = 0.025)	0.71	0.74
DecisionTreeClassifier (max_depth = 36)	0.66	0.77
GaussianNB	0.62	0.72
AdaBoostClassifier	0.60	0.60
MLP	0.59	0.72

shows the results of the classifiers.

As can be seen from the table, the best value of the class-averaged value of the harmonic measure is presented by the models based on convolution neural network and XGBoost classifier.

The XGBoost classifier model has the regularization parameters reg_lambda = 4, reg_alpha = 0.1 and a maximum tree depth of max_depth = 12.

The convolutional network contains three convolutional layers, two maxpooling layers, and a fully connected neural network of two output layers. The total number of layers including flatten and dropout layers is 12. The total number of trainable params is 724,230. The program code of the model and the graph illustrating the learning process of the network are given in Appendix A.

The detailed results of the XGBoost classifier (XGB) and convolution neural network (CNN) obtained using the classification_report utility are shown in Figure 9.

Comparison of the models’ performance (Figure 9) shows that the convolutional network is better at classifying non-saline (class 0) and very saline (class 4) surface areas. At the same time, both algorithms make more frequent errors in the region of extreme values, which may be due to the fact that there are fewer such data than average values.

6. Discussion

The obtained results show that the machine learning models perform the classification of salinity based on the pseudo-color composite quite accurately. However, a full evaluation of the method, including the construction of the composite, is not a trivial matter. To perform this, it is necessary to compare the obtained result with ground-based salinity estimates made by an independent method. Such a comparison of the current salinity with the long-term result will be approximate due to the statistical nature of the pseudo-color composite construction. To evaluate the results of this comparison, in addition to using the accuracy metric in the standard interpretation, an additional metric of “rough” errors (RE) was used, which shows the number of serious errors to the total number of responses excluding background pixels. To calculate RE, the class values are ordered from 0 (no salinity) to 4 (extreme salinity). In addition, an additional class number (5) is allocated to the background values (white is the unclassifiable zone). A classification error is considered serious if the predicted value of the class differs from the expected value by more than 1 up or down. For example, if the predicted value is 1—low salinity, but the real value is 0 or 2, in this case it is considered an error. If the predicted value is 1 and the real value is 3, this situation is considered a serious (“rough”) error. Accordingly,

$${Ac}_{RE}=1-RE$$

(10)

To evaluate the method in this paper, we compared the obtained results with expert estimates and laboratory studies of soil salinity.

6.1. Comparison with Expert Evaluation

The trained model was applied to assess the salinity of one of the sites in the Maktaaral district of the Kyzylorda region. This area is characterized by a pseudo-color composite (see Figure 10a). To verify the obtained results, a salinity map of this area, created within the framework of a survey on the assessment of the salinity of rural areas of the Maktaaral district, carried out by the South Kazakhstan Hydrogeological and Meliorative Expedition of the Committee on Water Resources of the Ministry of Agriculture of the Republic of Kazakhstan in 2019, was used. Each field was evaluated by the expedition’s experts as one of five salinity levels in whole or in large fragments (see Figure 10b).

The map proposed in Figure 10a is a multiyear average salinity estimate for the period of 2018–2022. Accurate assessment of the reliability of the developed map requires many years of laboratory measurements. It is impossible to obtain such data due to the absence of relevant archives. Therefore, it is reasonable to evaluate the accuracy of the obtained results using a special metric. The metric should evaluate serious errors of the model by comparing expert estimates (ground measurements) of salinity and model predictions.

Figure 11 shows the result of applying the pre-trained model to classify the specified area of the map using the pseudo-color compository.

In Figure 11b, the black dots indicate “rough” classification errors, i.e., when the value predicted by the classifier differs by more than one class up or down from the expert estimate shown in Figure 10b. Of the total number of the 5239 significant pixels (not background), the gross errors are 1175. That is, RE = 22.4%, or ${Ac}_{RE}=0.78$. The total number of errors is 3506, or 66.9% (see Figure 11c). It should be noted that hydrogeology experts of the meliorative expedition perform an estimation at once, and it is approximate, as a whole, for the big areas of a field based on its visible condition at the moment of estimation. However, the fields often have a patchy mosaic structure of salinity. It should also be taken into account that the simulation results refer to a multiyear period. Under these conditions, the method shows very high correspondence with one-time expert assessments.

6.2. Comparison with Laboratory Conductivity Measurements

In June 2022, the ground surveys of salinity were carried out during a field expedition. Surface soil samples were collected in a localized area during the expedition.

In total, 75 samples were collected from a depth of up to 30 cm. Each sample was processed in the standard way in the following order: drying, sieving to remove insoluble fractions, mixing with water in the proportion of 1:5, settling the solution for at least 24 h, and measuring the electrical conductivity using Hanna GroLine HI9814. The results of laboratory conductivity measurements are given in Appendix B. The results of the comparison of the long-term salinization map and the performed electrical conductivity measurements are shown in Figure 12.

Part of the samples (9) were received from sites where modeling data are missing (white background). The number of gross errors is 19 (28%, ${Ac}_{RE}=0.72$); in other words, the laboratory measurements show approximately the same correspondence to the forecast as the expert estimates shown in Figure 10. At the same time, the main gross errors appear where the method predicts low salinity.

6.3. Long-Term Salinity Map of the Maktaaral District and Method Restrictions

This indicates both the stability of the salinization processes in the area under consideration and the stability of the method itself, which allows revealing the long-term regularities of salinization processes in the cultivated arable land. The resulting long-term salinity map of the Maktaaral District is shown in Figure 13.

The obtained results allow drawing the following conclusions:

• The method based on taking into account the long-term salinization patterns demonstrates good correspondence to the actual salinization of fields in the Maktaaral region.
• Salinization in this area has a stable nature. The salinization pattern is repeated year after year with slight variations in the level of electrical conductivity.
• Although we have obtained a much more detailed salinity map, it is necessary to note the limitations of the proposed method.
• The method allows us to construct a map of long-term salinity trends, but its compliance with the current state of the field will not be complete due to both weather anomalies and long-term climate changes.
• Accurate verification of the obtained result is generally difficult. It requires many years of work on salinity assessment.
• The spatial resolution of the map is limited by the resolution of satellite images, which may not be sufficient in cases of small-scale salinity.

7. Conclusions

As a result of the performed research, a method of estimating the salinity of the cultivated soil areas, based on the account of long-term statistically stable states of the earth surface, is offered.

The method comprises two stages. In the first stage, a pseudo-colour model of a local area of the earth’s surface is built. Then, it is used to tune the machine learning models. The quality of the machine learning models is assessed quantitatively using the standard metrics. The obtained estimates are presented in Table 2. The tuned model is then used for salinity mapping.

The method is verified by comparing the resulting map with expert estimates obtained in 2019 and laboratory measurements of soil samples obtained in 2022.

Comparison of the obtained result with the data available to the authors from ground assessments of fields’ salinity in the Maktaaral district of Kyzylorda region (2019) shows rather good agreement between the assessments of specialist hydrogeologists of the meliorative expedition and the model conclusions (RE = 22%, or ${Ac}_{RE}=0.78$).

Comparison of the results with the employment of laboratory studies of electrical conductivity (2022) showed a slightly worse result (RE = 28% or ${Ac}_{RE}=0.72$).

The performed double verification demonstrates the stability of the method. This allows extending the model from a local area to the entire Maktaaral region, where the conditions of irrigated agriculture are approximately the same. In addition, verification estimates allow expecting positive results in other arid zones where irrigated agriculture is used.

The proposed method for assessing the salinity of cultivated lands has the following advantages:

• Relatively high resistance to annual fluctuations in weather conditions;
• The possibility of extrapolating the results obtained on the relatively small areas of the earth’s surface to large areas and regions;
• The gradient of the relief, the amount of precipitation (water availability of the year) and the state of the irrigation system are indirectly taken into account in the process of constructing a pseudo-color composite and subsequent training of the machine learning.

The main limitations of the method are as follows:

• Significant weather changes and the conditions of a single year can make serious adjustments to the salinity of certain areas of the earth’s surface. Such rapid changes cannot be captured using the proposed method;
• The pseudo-color composite, which is formed at the Nth step of the method, essentially depends on the conditions in a particular area of the earth’s surface and requires a significant share of manual labor.

Future Research:

• Improving the accuracy of the method by using a wider spectral range of remote sensing data, including radar;
• Increasing the coverage of the method by collecting more field data on soil electrical conductivity;
• Increasing the spatial accuracy of the method.