Soil Salinity Inversion Based on a Stacking Integrated Learning Algorithm

1. Introduction

Soil salinization is considered to be significant issues with ecological impacts, severely proscribing the safety and improvement of regional ecological areas [1,2,3,4]. Cultivated soil salinization is recognized to motive land degradation, harm crop growth, and preclude agricultural improvement [5,6,7,8]. Being in a position to gather soil salinity statistics rapidly and precisely ought to assist with efficaciously evaluating the diploma of soil salinization to facilitate the enhancement and utilization of salinized land. An ordinary technique for achieving soil salinity statistics has been fixed-point sampling with the use of a conductivity meter to measure the statistics [4,9,10,11,12]. Although this method has been shown to be effective, it has the shortcomings of being time consuming, being labor intensive, having poor representation of the measurement points, including only a small coverage area, and so forth and has limitations for soil salinity monitoring across large spatial scales [13]. In latest years, as remote sensing technology has been integrated with agriculture, it has been a means of achieving rapid acquisition of records at a lower value [4,14,15,16]. A growing number of remote sensing techniques are being applied to soil salinity monitoring [17,18,19,20]. Feature indices that are more sensitive to salinity are approached from remote sensing images, and a model is structured by incorporating feature indices with soil salinity content to facilitate the monitoring current status of regional soil salinity [8]. This method can make up for the shortcomings of previous field surveys and allow researchers to study regional soil salinization from a larger spatial scale, with the advantages of obtaining information quickly, being less affected by the ground, and being able to continuously and dynamically monitor regional salinization status, making it one of the most widely used quantitative soil salinization monitoring methods today [21].

Inversion of empirical statistical regression models for biophysical parameters of vegetation using remote sensing is usually classified into simply linear regression models on the one hand and non-linear regression models on the other hand. In order to explore the feasibility of using multiple vegetation indexes to invert soil salinity content, Wu et al [22] estimated soil salinity primarily on the basis of linear regression, synergistic kriging, regression kriging, and geographically weighted regression, and the findings indicated that geographically weighted regression method was the most accurate (RMSE = 0.31), respectively. Although linear regression models could effectively reduce the uncertainty in the inversion process [23], these models are difficult to solve for data that are nonlinear or have high correlation between features. In order to accurately estimate soil water in semiarid areas in the western part of Khorasan-Razavi province in (northeast) Iran, Hamed et al. [24] explored the sensitivity of vegetation indexes calculated from Landsat 8 remote sensing imagery to soil water by using random forest (RF), elastic net regression, and linear regression models, and findings indicated that RF regression model was the most accurate (RMSE = 0.04). Nonlinear regression models can provide an explanation for the correlation between the bio-physical and model variables [25], are easy to parallelize, and have a relatively strong model generalization ability; however, it is often necessary to trade off the balance of such a model with its accuracy if the prediction error of a single regression model is relatively low [26]. Ghosh et al [27] estimated biomass of mangrove forests in India making use of series of machine learning algorithms just like RF model, gradient boosting model and extreme gradient boosting model as well as integration of multiple machine learning algorithms. They found that the accuracy of inverting the aboveground biomass was further improved, showing RMSE to be 72.87 t/ha, using a stacking algorithm in a multi-temporal image stacked dataset. This RMSE was 1.62 t/ha less than that of the single RF regression model, indicating that an integrated learning regression algorithm based on stacking could integrate multiple underlying regression models and provide enhanced generalization capabilities [28].

At present, the integrated learning regression model based on stacking has shown good performance in soil water inversion [29,30,31]; however, there is still a need for in-depth research in soil salinity content inversion. Therefore, we attempted to use the stacking integrated learning regression model in soil salinity inversion to obtain more accurate results. The study objectives are: (1) to establish an inverse soil salinity model which is mainly used for RF model, back propagation neural network (BPNN) model, SVR model, convolutional neural network (CNN) model, and integrated learning algorithms and (2) to evaluate and analyze accurately the established soil salinity model, to quantitatively characterize the characteristics of changes in the soil salinity content of the region, to provide accurate monitoring and an inversion model of soil salinity to provide technical support and a conceptual foundation for the future use of inversion models.

2. Materials and Methods

2.1. Study Area

The Manas River Basin Oasis Area lies in the mid-region of the northern foothills of the Tianshan Mountains in China, in the southern edge of the Junggar Basin with a longitude of 85°01′–86°32′E and a latitude of 43°27′–45°21′N, as shown in Figure 1. The area includes Lower Nodi Irrigation District, Anzhihai Irrigation District, Jinguhe Irrigation District, Shihezi Irrigation District, Xinhu General Field Irrigation District, and Mosuo Bay Irrigation District [32]. The basin is arid with an average annual temperature of 4.7°C–5.7°C, with the highest temperature in July and the lowest in January, an average annual precipitation of 100–200 mm, and, mainly concentrated in summer, an average annual evaporation of 1,500–2,100 mm [33,34]. Ice and snow meltwater in the region carry salts from rock weathering into farmland. This perennial salt aggregation has resulted in serious salinization of the oasis, and due to the irrationality of irrigation, the salinization of farmland in the Manas River Basin is frequent, seriously affecting improvement of resource utilization and economic development [35]. Therefore, exploring fashions with greater accuracy is indispensable to understand regional soil salinity monitoring.

2.2. Data Sources

Soil salinity historical data (from April 2014) were sourced from reference 36 [36]. Remote sensing image data were synchronized with surface data, and the data were obtained from Landsat 8 OLI remote sensing image data collected by the United States Geological Survey (http://glovis.usgs.gov/). The data had been subjected in preprocessing steps such as geometric correction, radiometric correction, FLAASH atmospheric correction, synthesis, cropping, and so on. Moreover, visual interpretation and supervised classification had been applied to categorize the land use status to avoid any confusion regarding information about water bodies such as lakes, rivers, ponds, and puddles in the study area during salinity inversion.

2.3. Salinity Index Construction

The spectral index is an easy as well as valid method for measuring characteristic distributions on the land surface. As such, it was already widely used in global and regional land cover monitoring, vegetation classification and monitoring of environmental change [37,38,39]. In this study, according to the spectral characteristics of the features, we combined the spectral reflectance in various bands into the spectral index, and used them as indexes for remote sensing evaluation. Overall, five vegetation indexes, seven salinity indexes, one water index, and one brightness index were selected [5], and their formulas were shown in Table 1.

2.4. Model Construction and Accuracy Evaluation

2.4.1. Model Construction and Model Parameters Determination

The input and output data of the soil salinity inversion model were soil spectral index values calculated by Landsat 8 and surveyed soil salinity data, respectively, and model were built by using RF, BPNN, SVR and CNN models.

The grid search method [47] was used to identify the best variables in the machine learning regression model. The number of decision trees of the RF model was set to 200, and the minimum samples required for the division of internal nodes was set to 5. The number of iterations of the BPNN model was set to 1,500, the error threshold was set to 1e-5, and the learning rate was set to 0.001. The number of iterations of the CNN model was set to 1,000, the error threshold was set to 1e-6, and the learning rate was set to 0.01. Finally, the kernel parameter type of the SVR model was set to 0.01, the kernel function of the SVR model was set to “linear”, and the box constraint was set to 1.

The main steps of the stacking integrated learning regression model were as follows: (1) use MATLAB logic statements to optimize the results of the model which has the best model precision in a single machine learning regression algorithm, (2) use the trained RF model and BPNN model to predict a training set and a test set, (3) link the training set and the test set to the predicted results of RF and BPNN model to form a new training set and test set, (4) construct an stacking integrated learning regression model, and (5) use the trained integrated model to predict a training set and test set to evaluate performance of the model.

There are structure and rationale of the stacking integrated learning regression model as shown in Figure 2. The technology roadmap is shown in Figure 3.

The accuracy of the model that was assessed by the coefficient of determination (R²), RMSE, and relative percentage difference (RPD) [4,8]. Generally，higher R² values and smaller RMSE values indicate a more favourable model. The RPD was classified in three levels, which were Class A (RPD > 2.0, the constructed model is considered to be highly reliable), Class B (1.40 ≤ RPD ≤ 2.0, the constructed model is considered to be moderately reliable), and Class C (RPD < 1.40, the constructed model is not reliable) [4,5,36]. The formula to calculate the evaluation indicators is as follows:

$$R^2={\displaystyle \frac{{{\displaystyle {\sum }_{i=1}^n\left({\widehat{y}}_i-{\overline{y}}_i\right)}}^2}{{\displaystyle {\sum }_{i=1}^n(y_i-{\overline{y}}_i}{)}^2}}$$

(1)

$$RMSE=\sqrt{{\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n{\left({\widehat{y}}_i-{\overline{y}}_i\right)}^2}}{n}}}$$

(2)

$$RPD={\displaystyle \frac{\sqrt{\frac{{\displaystyle {\sum }_{i=1}^n{(y_i-{\overline{y}}_i)}^2}}{n}}}{RMSE}}$$

(3)

where ${\widehat{y}}_i$, $y_i$, and ${\overline{y}}_i$ are the predicted, measured, and average measured values of soil salinity, respectively, and n is the number of samples.

3. Results

3.1. Correlation Analysis between Spectral Indexes and Soil Salinity

There were 14 soil spectral indicators and soil salinity derived from Origin software, which conformed to the normal distribution test, the results of Pearson correlation analysis are shown in Figure 4.

As seen in Figure 4, among the triangular vegetation index, salinity index, water index, and brightness index, only the brightness index has no significant correlation with soil salinity, indicating that the relationship between soil salinity and the brightness index is relatively weak. Normalized Difference Snow Index (NDSI) and Standardized Reservoir Supply Index (SRSI) are positively correlated with soil salinity, and these indexes have large values in high salinity areas, reflecting the accumulation of soil salinity. Normalized Difference Water Index (NDWI) also shows a positive correlation with soil salinity. The Normalized Difference Vegetation Index (NDVI), Difference Vegetation Index (DVI), Soil-Adjusted Vegetation Index (SAVI), and Green Normalized Difference Vegetation Index (GNDVI) are negatively correlated with soil salinity, which indicates that in areas of high salinity, the growth and cover of vegetation was poorer and high salinity inhibited the normal growth of plants, resulting in a negative correlation.

3.2. Evaluation of Machine Learning Regression Models

The spectral index data (NDSI, SRSI, NDWI, NDVI, DVI, SAVI, and GNDVI) with good correlation and the corresponding soil salinity data were used as the input data of models. Four machine learning methods (RF, BPNN, CNN, and SVR) were used to construct the soil salinity inversion models, prediction results are shown in Table 2. Forecast values of four inversion models of soil salinity had been in comparison against the measurements of soil salinity, as shown in Figure 4.

As seen in Table 2 and Figure 5, the single optimal machine learning regression model was the RF. On the training set, both the RF and BPNN achieved better inversion results with R² above 0.5, RMSEs of 0.30 and 0.53, respectively, and RPDs of 1.4 or more, which reflects high accuracy. meanwhile, the CNN and SVR models underperform poorly on the training set, with R² of 0.20 and 0.11, RMSEs of 0.51 and 0.60, and RPDs below 1.4, respectively, indicating that the constructed models were unreliable. On the test set, the best inversion model was RF which had R2 of 0.49, RMSE of 0.43, and RPD of 1.40, with accuracies 23%, 31%, and 47% higher than those of the BPNN, CNN, and SVR models, respectively; the R² of the four models, in descending order, was RF > BPNN > CNN > SVR, and the RMSEs were the opposite of this. According to the model performance evaluation index, it was reasonable to use the spectral index to establish the model. Therefore, in this paper, the RF and BPNN models with better inversion results were selected as the base models for the stacking integrated learning regression.

The stacking integrated learning regression model can integrate many basic regression models and so forth to create much stronger predictions and deliver enhanced capabilities of generalization during inversion. The prediction results of the training set and test set of stacking integrated learning regression model are shown in Figure 6.

A performance comparison of the results of stacking integrated learning, RF, BPNN, CNN, and SVR regression models is shown in Figure 7. As seen in Figure 6, on the training set, the R² of the stacking integrated learning regression model relatively improved by 16.22%, the RMSE relatively reduced by 23.33%, the RPD relatively improved by 34.85% compared with the single optimal RF model, which changed from being a moderately reliable model to a model with higher reliability that can be used for model analysis. On the test set, the R² of the stacking integrated learning regression model was relatively improved by 8.16%, the RMSE was relatively reduced by 13.95%, and the RPD was relatively improved by 6.47% compared with the RF. In summary, the integrated learning regression model had a stronger generalization ability and improved the estimation accuracy of soil salinity once again on the basis of the superior results obtained in the training of the single optimal machine learning regression model.

3.4. Spatiotemporal Distribution of Soil Salinity

To better reflect inversion effect of model, the spatial distribution of soil salinity in study area in April and October 2016–2020 was mapped based on the constructed stacking integrated learning regression model. As seen in Figure 8, there was a significant difference in soil salinity. From a spatial point of view, the soil salinity content in study area in April and October of 2016, 2017, and 2018 showed a decreasing trend from north to south, with more pronounced salinization in the east-central part (near the Manas power plant and the town of Anjihai). In 2016, 2017, and 2018, the average salt content in April was 2.71, 2.67, and 2.61 g/kg, respectively, and the average salt content in October was 2.53, 2.57, and 2.58 g/kg, respectively. The values from October were lower by 0.18, 0.10, and 0.03 g/kg, respectively, compared to April. There was no significant change in soil salinity in most areas in April and October 2019. The average value of soil salinity in October 2020 was 2.71 g/kg, which was significantly higher, by 7.11%, compared to that of April of the same year.

As can be seen from Figure 9, the percentage of pixels with high soil salinity content of 2.75–2.8 g/kg in the study area in April 2016, October 2016, April 2020, and October 2020 was 19.78%, 1.16%, 0.14%, and 0.51%, respectively. These values decreased by 19.64% in April 2020 compared to April 2016, indicating a reduction in salinization. From 2016 to 2018, 55.30% of the area’s soil salinity shifted from 2.65–2.80 g/kg to 2.50–2.65 g/kg, especially from October to April. Soil salinity was maintained at 2.50–2.55 g/kg in 70.07% of the area in 2019. However, soil salinity intensified in October compared to April in 2020, with 96.34% of the area shifting from 2.55–2.65 g/kg to 2.65–2.80 g/kg, probably due to the disturbance of vegetation cover when acquiring remote sensing images. Overall, in April, the degree of soil salinization in the oasis area of the Manas River Basin decreased gradually, and the saline-alkaline land improvement measures were still effective.

4. Discussion

Accurately monitoring soil salinity content is essential to both food production with precision agriculture construction [48]. In this paper, a Pearson correlation analysis was introduced to filter the characteristics in spectral indices [49], and the findings of this study indicated that dominant factors of soil salinity were mainly vegetation index, salinity index, and water index, and this is in agreement to Alhammadi et al. [41,42,43,44]. The difference was that this study obtained a significant correlation between water index and soil salinity. Soil salinity migrates when soil moisture moves, and there will be an increase in soil electrical conductivity and higher soil moisture content in areas with higher soil salinity [50,51]. Tang et al [52] also find that the albedo of salinized soil increases due to the specular effect of water bodies as soil moisture content increases towards a critical point. Therefore, the water index, vegetation index, and salinity index were input as variables for the building of soil salinity inversion models.

The development of machine learning algorithms has accelerated the remote sensing modeling process. Many machine learning algorithms often exceed the accuracy of traditional regression modeling [40,41,42]. In this research, four machine learning regression models, RF, BPNN, CNN and SVR, had been made to simulate soil salinity, of which RF obtained best inversion results with R² of 0.74, RMSE of 0.30 and RPD of 1.98. Wei et al. [53] quantitatively estimated soil salinity by using multispectral imagery from unmanned aerial vehicles and established the BPNN, SVR, and RF models, and their findings indicated RF model was the most effective. Zhang et al. [54] explore the problem of rapid monitoring of soil salinity. They used support vector machine, BPNN, RF, and multivariate linear regression models to establish soil salinity, and their findings indicated RF model is the optimal with an R2 of 0.72, this is in agreement well with this research.

Because of both plant individual differences and changes in canopy over the seasons, there is often the shortcoming of a single machine learning model with poor generalizations [8]. Yang et al [8,55] presented a stacking approach to accomplish the task of efficiently estimating highly accurate forecasts by combining a number of weak forecasters into one powerful forecaster, which can effectively improve the estimation accuracy. Presently, while integrated learning Models on the basis of stacking algorithms are broadly applied for the domains of machine vision and natural language processing, the application of integrated learning models for soil salinity have not been explored in terms of stacking strategy [8]. In this study, we constructed a stacking-based integrated regression model that produces higher accuracy, with an R² increase more than 16.22% compared to a single machine learning regression model based on a training set. Thus, there is a significant advantage of the stacking algorithm in predicting soil salinity. In other words, the more complex integration algorithm, that is, the stacking integration model, is significantly more capable of handling complex problems than the single machine learning regression model, a finding also supported by Pham et al. [56]. The first layer of the stacking method consists with various underlying models with inputs from the original training set, while the outputs are the predicted values of the various base models. The second layer consists of only one meta-model, which is trained on forecast and true values of the various models in the first layer to form the completely integrated model. Similarly, the process of predicting the test set goes through the predictions of all the base models to compose the features of the second layer, next, a second layer of meta- model is implemented to predict final outputs in order to keep approaching the true value [57], which could be prone to overfitting if the training set of the base models is used directly to generate the training set of the meta-model. This study focused on RF and BPNN based models rather than those based on the CNN and SVR. This was because the RF and BPNN based models obtained better results than the CNN and SVR based models when we tested the model performance. However, the approach to combining the advantages of a single model to build an integrated model remains the key to future research.

The stacking integrated learning regression model was constructed to carry out soil salinity inversion with the oasis zone of the Manas River Basin as the research object. Findings of research show soil salinity content in this area had a decreasing trend from north to south in 2016, 2017, and 2018. Region of study is a typical mountain-basin structural system and sees differences in soils, water table, and climate based on latitude and location [58]. Zhang et al. found that a low-latitude area had a long development time, high soil maturity, and low soil salinity. Meanwhile, a mid-latitude area had a high water table, and with the promotion of membrane drip irrigation technology, the crop area could expand rapidly. Moreover, the membrane drip irrigation technology did not take away the salinity, so the mid-latitude area had a higher soil salinity. Finally, a high-latitude area that situated at the lower river basin and had a texture that was dominated by sandy loam and sandy soil would have a low water table. The soil salinity would also be higher than that in the mid-latitude area [59,60,61], this is in agreement with this study's results. In 2016–2018, soil salinity was reduced in October compared with April, probably because the change in groundwater burial depth during the year showed a mining type. Then in October, the groundwater declined rapidly, and the deeper groundwater burial depth and frequent agricultural irrigation made soil salinity decrease in the month. It is in accordance with the findings of Chen et al. [62] on soil salinity and nutrients in different landscape types. However, in April and October 2019, there was no significant change in soil salinity in the vast majority of the area. Soil salinity became stronger in 2020 compared to 2019, especially in October 2020, when soil salinity increased significantly. As temperatures increased, soil water rose vigorously, and salts then increased and accumulated in the soil surface. This has been similar to Zhao et al.'s [63] study on 121 Corps, 8th Agricultural Division, Manas River Basin, where crop soil salinity showed a salt accumulation trend in early May and mid-September. The integrated learning regression model constructed in this study is a fast and accurate inversion of soil salinity content for quantitative assessment and monitoring of saline soils regionally. Nevertheless, the distribution patterns of soil salinity for various crops and seasons may have significant differences that would need to be further investigated in the future.

5. Conclusions

We constructed soil salinity inversion models by combining Landsat 8 OLI remote sensing data and four methods of machine learning. Findings indicated that the single optimal machine learning regression model was the RF. In addition, the innovative stacking integrated learning regression model that was constructed in this study yielded better soil salinity inversion results, with a relative increase of 16.22% in the R², a relative decrease of 23.33% in the RMSE, and a relative increase of 34.85% in the RPD on the training set compared with the single optimal machine learning regression model. On the test set, the R² was relatively improved by 8.16%, the RMSE was relatively reduced by 13.95%, and the RPD was relatively improved by 6.47%. Soil salinity in the oasis area of the Manas River Basin was accurately predicted and characterized. The soil salinity content in this area in April and October of 2016, 2017, and 2018 shows a trend of decreasing from north to south. And it gradually decreased in April, with the percentage of pixels with high soil salinity content of 2.75–2.8 g/kg in the study area decreasing by 19.64% in April 2020 compared to April 2016. Overall, the regression model on the basis of stacking had high accuracy, which could be of important relevance towards the quick abstraction of salinity status and soil salinity maintenance and management in the Manas River Basin. However, further exploration is needed to combine advantages of multiple single models to construct an integrated model, and this consideration is the key to future research.