Estimating soil profile salinity under vegetation cover based on UAV multi-source remote sensing

doi:10.21203/rs.3.rs-4971758/v1

Download PDF

Article

Estimating soil profile salinity under vegetation cover based on UAV multi-source remote sensing

https://doi.org/10.21203/rs.3.rs-4971758/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Soil salinization is the most common land degradation problem in arid, semi-arid and coastal areas of China, which seriously affects local crop yield, economic development, and environmental sustainability. There are few studies on estimating soil salinity at different depths under vegetation cover. In this study, field soil control experiments were employed to collect multi-source remote sensing data under barley growth, and soil salt content (SSC) with various depths. Three types of feature variables were built based on images and were filtered by the boosting decision tree (BDT) method. Besides, four machine learning algorithms coupling with seven variable combination groups were used to comprehensively establish soil salinity estimation model. Finally, the performances of estimation model for different crop over ratios were evaluated. The results showed that the gaussian process regression (GPR) model based on the full variable group at the depths of 0 ~ 10 cm and 30 ~ 40 cm is more accurate than other models. The validation R² is 0.774 and 0.705, and the RMSE is 0.185% and 0.31%;The random forest (RF) models based on spectral index and texture data at 10 ~ 20 cm and 20 ~ 30 cm depths are more accurate, with R² of 0.666 and 0.714. SSC may be quantitatively inverted at various depths using the machine learning model based on multi-source remote sensing, which also serves as a guide for monitoring soil salinization.

Earth and environmental sciences/Biogeochemistry

Earth and environmental sciences/Environmental sciences

Physical sciences/Engineering

soil salt content

UAV

vegetation cover

soil depth

multi-source remote sensing data

In China, the soil salinization area extends 9.21 million hectares, accounting for 6.62% of the total cultivated land. The process of saline soil formation is intricate, posing significant challenging for the detection and dynamic monitoring of soil salinization. Conventionally, soil salinization was assessed through field sampling and chemical analysis, these methods that were both time consuming, and labor-intensive¹. In contrast, remote sensing technology, a rapid and expansive means to gather data on ground objects across different temporal and spatial, is an ideal tool for soil salinity monitoring². The spectral response of soils varied with salt content, with high-salinity soils exhibiting stronger responses in the visible and near-infrared bands than those with lower salinity³. Leveraging remote sensing for the dynamic monitoring of soil salinization is crucial for the efficient utilization of soil and water resources, providing essential insights into the timing, patterns, and locations of potential changes in soil salinity, thus facilitating improved resource management and planning.

In recent years, Unmanned Aerial Vehicle (UAV) and other aerial remote sensing platforms have experienced rapid advancements, increasingly integrating into civilian applications and becoming prominent in agricultural research. UAVs boast advantages like portability, high flexibility, and customizable flight durations. Zhao et al.⁴ used multispectral remote sensing data from three research locations to establish soil salinity inversion models based on support vector machine (SVM), RF, backpropagation neural network (BPNN), and extreme learning machine (ELM). The results showed that the four spectral index-based models could get a high level of inversion accuracy. Similarly, Wei et al.⁵ used a UAV equipped with Micro-MCA multispectral sensors to capture pictures for evaluating the soil salinity in a small area of the Hetao Irrigation District. Chen et al.⁶ developed estimation models of SSC in sunflower fields at different soil depths during the budding and blooming stages using UAV multispectral data. IVUSHKIN et al.⁷ found that UAVs equipped with multiple sensors of hyperspectral, multispectral, thermal infrared, and LiDAR cameras have great potential for monitoring soil salinization. Feature indices like the soil salinity index and vegetation index, derived from spectral band reflectance transformations, serve as important variables for estimating soil salinity content⁸. Qi et al.⁹ collected spectrum reflectance and spectral indices based on a UAV ground cooperative system, and then used machine learning algorithms like BPNN to build a salinity inversion model. The findings indicated that the constructed model accurately captured the level of salinization in the research area.

The relationship between soil salinity and remote sensing feature variables was frequently nonlinear due to the combination of complex factors such as soil, vegetation, and atmospheric signatures¹⁰. A common method for estimating SSC was to use the mathematical statistical model, especially the linear regression model including partial least squares regression (PLSR)^11,12. But the relationship between spectral covariates and soil properties in nature was rarely linear¹³. Machine learning (ML) algorithms, renowned for their ability to address nonlinear problems and process high dimensional data, often outperform statistical regression models in predicting soil salinity. Algorithms like RF、SVM and BPNN could capture these intricate nonlinear relationships through multi-layered results, making them particularly effective for soil salinity estimation¹⁴. Hu¹⁵ compared PLSR and RF methods using hyperspectral first order differentiation, broad band spectral indices, and narrow band spectral indices as independent variables. It was reported that the RF model could better predict soil salinity and the model for the bare soil area had the highest prediction accuracy compared to the other sample areas. Zhang et al.¹⁶ found that the SVM method based on texture features had an improved effect on the monitoring accuracy of soil salinity information. Wei et al.⁵ established various machine learning models to evaluate and optimize the best salt estimate model. However, the prediction accuracy of a single machine learning algorithm could fluctuate under varying conditions. Therefore, evaluating the performance of multiple machine learning regression algorithms is essential for constructing reliable soil salinity prediction models that could adapt to diverse environmental factors.

There is a strong correlation between vegetation growth and soil salinity, as evidenced spectrally in two primary ways: differences in leaf spectral reflectance and significant variations in the texture features of spectral images influenced by changes in soil salinity or leaf characteristics^17,18 Studies have shown that texture features are extensively utilized to reveal vegetation characteristics variations, combining them with spectral information could effectively improve the accuracy of predictive models^{19,20,21,22,23}. Huang et al.²⁴ used Sentinel-2 image combined with texture features to significantly improve the classification effect of moderate saline soil in the Yellow River Delta region. Nevertheless, the prevalent use of remote sensing data for soil salinity detection primarily relied on spectral indices⁵, with limited studies investigating the use of texture features in soil salinity estimation²⁴. In the case of bare soil, the spectrum could directly determine the salt content of the soil surface. Conversely under vegetation coverage, soil salinity could be indirectly assessed by vegetation canopy spectral signatures²⁵.

Few investigations on estimating soil salinity under vegetation cover have been conducted in the past. Therefore, the purpose of this study is to construct soil salinity estimation models at various depths beneath barley growth. The main research contents were to:(1) evaluate the potential and feasibility of UAV remote sensing for SSC estimation under barley coverage; (2) optimize the feature variables of spectral band, spectral index and texture data based on BDT method; (3) Validate the accuracy of soil salt estimation model for different vegetation coverage conditions. This study presented a novel method for the dynamic monitoring of soil salinity in agricultural production helped to achieving precise irrigation and fertilization practices, and improved agricultural productivity efficiency and sustainable utilization of soil resources.

2.1 Study area and experimental design

The study area is located at the ecological experimental station of Yangzhou University, situated in the Jianghuai Plain of Jiangsu Province in eastern China (119°24′E, 32°21′N), at an altitude of 5 meters. This region is characterized by a subtropical monsoon climate, with an average annual frostfree period, precipitation, evaporation, and air temperature of 223 days 937 mm, 1063 mm, and 14.8°C, respectively. Soil samples for the experiment were collected from Tiaozini reclamation area in Dongtai City, Jiangsu Province. This area was subject to marine intrusion and groundwater topdressing, with high salinity in the soil tillage layer.

The crop studied was barley (Hordeum vulgare L.). Four different soil salinity treatments were set: control (no salt), low salinity (3‰), medium salinity (5‰), and high salinity (10‰). The salinity experiment was carried out using a box planting setup with dimensions of 100 cm in length, 40 cm in width, and 40 cm in height, each containing 120 kg of base soil. Each treatment was replicated thrice, resulting in a total of 12 experimental units. The cultivation adhered to local management practices, encompassing weeding, pest, and disease prevention and control. Barley was planted annually at the early of November, with each box accommodating two rows and was harvested at the late of May of the following year. To maintain the soil salinity status of each treatment, the experimental box devices were designed as sealed containers to prevent salt leaching by precipitation and irrigation.

2.2 Data collection and acquisition

2.2.1 Soil salt measurement

During the barley growth stages of reviving-jointing, jointing-filling, and grain filling maturity, soil electrical conductivity data were measured every 7–10 days, in conjunction with the acquisition of multispectral imagery. A total of 84 dataset were collected during barley growth. The soil electrical conductivity was measured using the EC450 conductivity meter (Spectrum Technologies Co., Ltd., Chicago, IL, USA). First, the electrode was calibrated using a calibration solution (conductivity: 1413 µS/cm). After calibration, the electrode was inserted into the soil profile to measure conductivity at soil depths of 0ཞ10 cm, 10ཞ20 cm, 20ཞ30 cm, and 30ཞ40 cm. The soil conductivity values were directly recorded by the handheld meter. Subsequently, the SSC was derived using the empirical formula (SSC = 0.2882EC + 0.0183, %).

2.2.2 Multispectral data acquisition and processing

This study leveraged remote sensing data from multiple UAV based sensors to enhance monitoring of soil properties through data fusion. The DJI Inspire 2 UAV platform (DJI Inc., Shenzhen, China) used in this study, which was equipped with a Altum multi-spectral and infrared camera (MicaSense, Inc., Seattle, WA, USA). This camera was capable of capturing images across six spectral bands (blue, green, red, red edge, near-infrared, and thermal infrared) simultaneously. Variables The UAV flight operations were conducted under optimal conditions of clear skies and calm winds, between 11 a.m. and 2 p.m. local time. To ensure high-quality data, the sensor was prewarmed for 5 minutes and the reference plate was used for radiometric calibration before each flight. The flight altitude and the cross-track overlap were maintained to 25 meters and 75%, respectively. The camera was oriented vertically downward to achieve a ground resolution of 1.1 cm per pixel. After preprocessing the collected multispectral images by radiation correction, geometric correction, and image mosaic, the reflectance data of each pixel were generated during crop growth period. To extract canopy reflectance, a region of interest (ROI) was preset near the center of each plot.

2.2.3 Vegetation coverage calculation

After collecting multispectral imagery, the canopy light interception of barley was measured using the AccuPAR LP-80 ceptometer (Decagon Devices Inc, Pullman, WA, USA). The LP-80 is equipped with 80 independent sensors, each spaced uniformly at 1 cm intervals, capable of measuring solar radiation within the 400–700 nm band in different modes. During the measurement, the sensor was strategically positioned in the center of the plot, aligned with the row direction. With the acquired photosynthetically active radiation at the top and bottom of the canopy, the instrument automatically calculated the crop LAI through a built-in algorithm with the other variables. In accordance with the reports and employing a classification method that considers similar geographical features and vegetation types²⁶, vegetation cover was classified according to the LAI value as follows: low coverage (0, 0.45), medium coverage (0.45, 0.75) and high coverage (0.75, 1).

2.3 Construction of feature variables

Spectral indices were classical variables that integrated the spectral characteristics of each band of ground objects and enhanced specific information through mathematical transformations and combinations of reflectance values from different bands. Salinity indices, which were frequently employed for the rapid assessment of soil salinization, exhibited a strong correlation with bare soil salinity²⁷. Similarly, vegetation indices were frequently used for the quantitative assessment of vegetation growth²⁸. In this study, a selection of widely recognized spectral indicators were employed for the monitoring of soil salinization, including 15 salinity indices and 15 vegetation indices.

The selected salinity indices include: Normalized Difference Salinity Index (NDSI), R-edge Normalized Difference Salinity Index (NDSI-reg), Brightness Index (BI), Salinity Index 1 (SI1), R-edge Salinity Index 1 (SI1-reg), Salinity Index 2 (SI2), R-edge Salinity Index 2 (SI2-reg), Salinity Index 3 (SI3), R-edge Salinity Index 3 (SI3-reg), Salinity Index (SI-T), Salinity Index S1, Salinity Index S2, Salinity Index S3, Salinity Index S5, Salinity Index SI; The selected vegetation indices include: Normalized Difference Vegetation Index (NDVI), R-edge Normalized Difference Vegetation Index (NDVI-reg), Difference Vegetation Index (DVI), R-edge Difference Vegetation Index (DVI-reg), Enhanced Vegetation Index (EVI), R-edge Difference Vegetation Index (EVI-reg), Triangular Vegetation Index (TVI), Normalized Greenness Index (NDGI), Simple Ratio Index (SR), Modified Soil Adjusted Vegetation Index (MSAVI), Optimized Soil Adjusted Vegetation Index (OSAVI), Soil Adjusted Vegetation Index (SAVI), Visible Light Band Difference Vegetation Index (VDVI), Visible Atmospherically Resistant Index (VARI), Green Normalized Difference Vegetation Index (GNDVI), and a normalized relative canopy temperature (NRCT). The calculation formulas for these indices were shown in Table 1.

Table 1

Spectral index and calculation formula
Feature	Formulation	Reference
NDSI	NDSI= \(\:\frac{R-NIR}{R+NIR}\)	29
NDSI-reg	NDSI-reg=\(\:\frac{RedEdge-NIR}{RedEdge+NIR}\)	30
BI	BI=\(\:\sqrt{{R}^{2}+{NIR}^{2}}\)	31
SI1	SI1=\(\:\sqrt{G\times\:R}\)	28,32
SI1-reg	SI1-reg=\(\:\sqrt{G\times\:RedEdge}\)	30
SI2	SI2=\(\:\sqrt{{G}^{2}+{R}^{2}+{NIR}^{2}}\)	28,32
SI2-reg	SI2-reg=\(\:\sqrt{{G}^{2}+{RedEdge}^{2}+{NIR}^{2}}\)	30
SI3	SI3=\(\:\sqrt{{G}^{2}+{R}^{2}}\)	28,32
SI3-reg	SI3-reg=\(\:\sqrt{{G}^{2}+{RedEdge}^{2}}\)	30
SI-T	SI-T=\(\:100\times\:R/NIR\)	33
S1	S1=\(\:B/R\)	34
S2	S2=\(\:\frac{B-R}{B+R}\)	34
S3	S3=\(\:G\times\:R/B\)	34
S5	S5=\(\:B\times\:R/G\)	34
SI	SI=\(\:\sqrt{B\times\:R}\)	35
NDVI	NDVI=\(\:\frac{NIR-R}{NIR+R}\)	36
NDVI-reg	NDVI-reg=\(\:\frac{NIR-RedEdge}{NIR+RedEdge}\)	30
DVI	DVI= \(\:NIR-R\)	28
DVI-reg	DVI-reg= \(\:NIR-RedEdge\)	30
EVI	EVI=\(\:2.5\times\:\frac{NIR-R}{NIR+6R-7.5B+1}\)	37
EVI-reg	EVI-reg=\(\:2.5\times\:\frac{NIR-RedEdge}{NIR+6RedEdge-7.5B+1}\)	30
TVI	TVI=\(\:0.5\times\:\frac{120\times\:\left(NIR-G\right)}{200\times\:\:\left(R-G\right)}\)	9
NDGI	NDGI=\(\:\frac{G-R}{G+R}\)	38
SR	SR=\(\:NIR/R\)	39
MSAVI	MSAVI=\(\:\frac{\left(2NIR-1\right)-\sqrt{{\left(2NIR+1\right)}^{2}-8\left(NIR-R\right)}}{2}\)	30
OSAVI	OSAVI=\(\:1.16\times\:\frac{NIR-R}{NIR+R+0.16}\)	30
SAVI	SAVI=\(\:1.5\times\:\frac{NIR-R}{NIR+R+0.5}\)	40
VDVI	VDVI=\(\:\frac{2G-\left(R+B\right)}{2G+\left(R+B\right)}\)	41
VARI	VARI=\(\:\frac{G-R}{G+R+B}\)	42
GNDVI	GNDVI=\(\:\frac{NIR-G}{NIR+G}\)	9
NRCT	NRCT=\(\:\frac{Ti-Tmin}{Tmax-Tmin}\)	43
Note: B, R, G, NIR, and Rededge represent the reflectance of blue, red, green, near-infrared, and red edge bands, respectively.

The image texture was indicative of variations in soil surface’s color and gray level, which was closely related to the soil's salt and water content. In this study, the statistical Gray Level Cooccurrence Matrix (GLCM) method was used to extract the textural features of the images. GLCM is a prevalent method widely used for image feature extraction, texture analysis, and quality evaluation⁴⁴, describing the correlation between pixel gray levels within images. Eight characteristic variables including mean (MEA), variance (VAR), uniformity (HOM), contrast (CON), difference (DIS), entropy (ENT), second-order moment (SEC) and correlation (COR) were calculated for each band of UAV multispectral images utilizing the second-order statistical filtering tool. Variables In total, this study incorporated 5 band reflectance values, canopy temperature, 31 spectral indices, and 40 texture feature data, amounting to 77 feature variables serving as independent variables.

2.4 Feature variable selection and set construction

2.4.1 Optimization of characteristic variables

To optimize the complexity of the model input variables, this study employed the BDT method to refine the selection of the 77 characteristic variables. BDT is one of the embedded type methods which used to compute estimates of Predictor Importance for the tree by summing changes in the mean squared error (MSE) due to splits on every predictor and dividing the sum by the number of branch nodes⁴⁵. The selection steps involved in this optimization process include: (1) the data were divided into twenty groups; (2) importance score of feature variables in each set of data was calculated based on BDT; (3) normalizing the important scores for various data categories. (4) A statistical analysis and sorting were performed on the normalized data of 20 groups. The method used in this study was as follows: The obtained 84 datasets were randomly divided, with 70% allocated to the training subset and the remaining 30% reserved for the test subset. This process was repeated 20 times. The results from each iteration were normalized across different groups (spectral band reflectance group, spectral index group, and texture data group). Based on the contribution degree of characteristic variables within each group, the threshold was set to 0.1 to select the pivotal variables for model training and establishing.

2.4.2 Construction of model training

The construction of a predictive model necessitates the careful extraction of input features and the judicious selection of appropriate machine learning algorithms. To explore the influence of different features on model performance, the feature combinations were categorized as follows:

Single variable groups: (1) Band reflectance. Composed of the original reflectance values of each spectral band. These values were directly reflected the physical and chemical properties of the soil, serving as the primary features for predicting SSC. (2) Spectral Index. Consists of various spectral indices (such as NDVI, RVI, etc.) calculated from different spectral bands. These indices were designed to enhance specific soil or vegetation characteristics. (3) Texture data. Extracting from spectral data, such as the gray level cooccurrence matrix. These features captured subtle changes and spatial structure of the soil surface.

Pairwise variable groups: (1) Spectral band reflectance and indices combination: This combination retained the original spectral information while incorporating enhanced information, potentially providing more comprehensive responses of soil salinity changes. (2) Spectral band reflectance and texture feature combination: By integrating the complementary advantages of both types of data, this combination revealed spectral characteristics and captured spatial heterogeneity of the soil surface. (3) Spectral indices and texture feature combination: This combination enriched the enhanced information with spatial structure data, potentially improving the model's generalization.

Full Variable Group: This group contained all types data of spectral bands reflectance, spectral indices, and texture features. The comprehensive inclusion of all potentially relevant information was intended to enhance the model's generalization capabilities and predictive performance.

2.5 Machine learning models

Based on the SSC and the selected feature variables, four machine learning methods of RF, SVM, GPR, and BPNN methods was used to construct soil salt estimation models.

RF was an ensemble learning method that used to establish prediction models for classification and regression problems by using randomly unrelated decision trees⁴⁶. In recent years, RF has been widely used in vegetation growth parameters estimate and the estimation of soil physical and chemical parameters. For instance, Huang et al.⁴⁷ established several soil salinity estimation models based on Landsat-8 OLI images in a study of oasis soil salinity in arid areas, and found that the estimation accuracy of the RF modeling method was higher than that of classical statistical models. Similarly, Sui et al.⁴⁸ developed a soil salinity estimation model based on original observations and satellite data in a study of coastal soil salinity, utilizing hydrological connectivity measurements and the RF algorithm.

SVM was a method that implemented the concept of structural risk minimization, effectively addressing small sample sizes, nonlinearity, and high-dimensional data. SVM offered strong expression capability, generalization ability, and learning efficiency. It easily integrated with multi-source information, thereby achieving higher estimation accuracy⁴⁹. For example, Cai et al. (2010) combined multispectral and texture features and used the SVM classifier to identify soil affected by salt, confirming that the SVM classifier effectively extracted soil salinization distribution information in the Yinchuan Plain. Similarly, Guan et al.⁵⁰ introduced SVM theory into the dynamic prediction of soil EC value, constructing a dynamic prediction model for soil salinity aimed at managing irrigation water in salinized areas.

GPR was a nonparametric Bayesian regression method that predicted by assuming the data distributed as a multivariate Gaussian distribution, providing an estimate of prediction uncertainty. It was flexible and efficient for small datasets but can be computationally expensive and difficult to scale for larger ones. Additionally, selecting and adjusting the appropriate kernel function required considerable experience and testing.

BPNN was a feedforward network composed of multiple neurons capable of learning and identifying nonlinear relationships in complex systems. It exhibited strong self-learning ability, adaptability, and resistance to interference, making it highly promising for the estimation of soil physical and chemical parameters. For instance, Wang et al.⁵¹ successfully established a prediction model for soil moisture and salinity using BPNN with Landsat-8 satellite data.

2.6 Technical workflow

This study aims to identify the most effective variables set for accurately predicting SSC, thereby improving the overall performance and applicability of the model in diverse soil depths. SSC was designated as the dependent variable, while different variable groups serving as the independent variables. The sample data were randomly divided into two groups, with 70% allocated for model training and 30% for validation. Four different machine learning methods (RF, SVM, GPR, and BPNN) were employed to estimate SSC. Each method constructed corresponding prediction models, optimizing the performance by adjusting the model hyperparameters. To further evaluate model accuracy, 10-fold cross-validation was used to construct and test the SSC estimation models (Fig. 1).

3.1 Statistical of soil salinity distribution

The salt content of sampling points at different soil depths was categorized as follows: non-saline soil (< 0.2%), mild salinization (salt content 0.2 ~ 0.5%), severe salinization (0.5 ~ 1.0%), and saline soil (> 1.0%). The non-salt treatment had an average salt content of 0.214%, the low-salt treatment had an average salt content of 0.257%, the medium-salt treatment had an average salt content of 0.418%, and the high-salt treatment had an average value of 1.353%. The statistical of obtained salt content data were as following: in the 0 to 10 cm soil depth, the measured SSC were relatively lower than those of other depths, which varied from 0.031% to1.04%, with an average of 0.428%. At a depth of 10 to 20 cm, the SSC ranged from 0.087–1.406%, with an average of 0.561%. The SSC in the 20 to 30 cm depth were the highest, with the average SSC of 0.633%, varying between 0.055 ~ 1.806%. The SSC distribution in 30 ~ 40 cm was similar to that in the 10 ~ 20 cm soil layer, with an average of 0.619%. According to these criteria, the measured salt grade distribution in the study area was shown in Fig. 2.

3.2 Contribution and selection of feature variables

In this study, the important factor of 77 feature variables of three types data for different soil depths were calculated (Fig. 3). After normalizing the feature variables important factor of different groups and setting a threshold of 0.1, the feature variables were screened. Common variables across different depths were selected for the ML model training. Finally, the selected variables for different groups were as follow: Band reflectance: B, R-edge, and NIR; spectral Index: SI2, SI2-reg, BI, SI-T, DVI, EVI, SAVI, OSAVI, DVI-reg, EVI-reg, and MSAVI; Texture Data B-MEA, B-VAR, B-CON, B-ENT, B-SEC, G-VAR, RE-MEA, NIR-MEA, NIR-HOM, NIR-ENT, and NIR-SEC. Altogether, 25 feature variables were filtered for modeling.

3.3 Comparision of model performance for various ML methods and data groups

The optimal variables of the three variables types of band reflectance, spectral index, and texture data and their combinations were used as the independent variables and the SSC as the target variable imputing into the machine learning (RF, SVM, GPR, and BPNN) model to establish the soil salt prediction model. The prediction accuracies of these models based on different variable combinations were shown in Figs. 4 and 5.

Figures 4 and 5 demonstrated that for each algorithm, the band reflectance and spectral index variables performed well in estimating SSC at a depth of 0 to 10 cm within the single variable groups. Among these, the BPNN model, based on the spectral index, achieved the best performance with an R² of 0.74 and RMSE of 0.26%. The whole variable combination-based GPR model was the best in the multivariate combination group. It exhibited a stableR² value around 0.77, an RMSE of 0.24%, and great overall stability. The RF model identified the band reflectance as the most single variable for SSC at a depth of 10 to 20 cm, whereas the SVM model produced the worst estimation. Across the four approaches tested, texture data yielded suboptimal results, with an average RMSE exceeding 0.34%. However, The RF model that incorporated both texture and spectral index data outperformed others in the multivariate combination, achieving a mean RMSE of 0.31% and an R² of approximately 0.61, making it the most effective model overall. Additionally, the GPR model's results outperformed those of other approaches. In contrast, the SVM and BPNN models consistently delivered less accurate predictions across all groups.

At a depth of 20 to 30 cm, the RF model built using band reflectance and spectral index demonstrated superior performance, with average R² values of 0.62 and 0.67 and mean RMSE of 0.31% and 0.29%, respectively. This model outperformed than the models built by the other three algorithm. Within the multivariate combination, the spectral index and texture data performed better outcomes in the RF model compared to other combinations, yielding an R² of 0.67 and an RMSE of 0.29%. Similarly, the GPR algorithm, when applied to the texture data showed higher accuracy than the other three algorithms. The SVM model achieved its highest accuracy when all variables were combined, while the accuracy of models based on band reflectance and spectral index data in the GPR and BPNN algorithms was marginally higher than that of other combinations, with R² values of 0.64 and 0.58, and RMSE values of 0.30% and 0.35%, respectively.

The RF model built using textural data as a single variable had the lowest accuracy at a depth of 30 to 40 cm, with an average RMSE of 0.39% and R² of 0.37. In contrast, the RF model constructed using the spectral index showed higher accuracy than the other three models. Within the multi-variable combination group, the SVM model retained consistent accuracy, with the R² of approximately 0.55. The GPR model, when applied to the full variable set, demonstrated superior accuracy compared to other combinations. Notably, the BPNN model exhibited better accuracy when based on band reflectance and spectral index data than other models.

In conclusion, soil salinity estimation at different depths were improved by applying a variety of machine learning techniques, with the effectiveness of these methods differing across modeling groups. The most effective method for estimating soil salinity was the whole variable group based GPR model, particularly for both surface (0ཞ10 cm) and deep (30ཞ40 cm) measurements. However, for middle depth soils (10–20 cm and 20–30 cm), the RF model based on spectral index and texture data displayed the best outcomes. The RF model effectively reduced the variance of a single tree by integrating multiple decision trees, thereby improving the model's stability and estimation accuracy. While the SVM and BPNN models performed marginally worse overall than GPR and RF, they were still fairly good at predicting within particular categories.

3.4 Outer-sample validation and analysis

The performances of optimal estimation models have been evaluated 20 times using the outer-samples (not training) data (Fig. 6). At a soil depth of 0ཞ10 cm, the R² values of the validation set fluctuated between 0.6 and 0.9, with an average R²of 0.77. The RMSE values typically ranged from 0.1 to 0.3%, with an average RMSE of 0.185%. This indicated that the GPR model had high prediction accuracy and stability in predicting shallow soil salinity. For the 10ཞ20 cm depth, the RF model showed favorable results, with an average R² of 0.67. The RMSE values ranged between 0.2 and 0.4%, with an average RMSE of 0.325%. At the 20 ~ 30 cm depth, the RF model achieved a maximum R² of 0.87 and minimum RMSE of 0.212%. At a deepest soil layer, the model's R² ranged from a minimum of t 0.52 to a maximum of 0.92, indicating high overall accuracy and stability. The RMSE values fluctuated between 0.169% and 0.485%, with an average RMSE of 0.31%.

The modeling and validation accuracies of four soil salinity estimation models at different depths were compared. Models were selected at each depth to achieve better estimation results in both modeling and validation. The average R² values in the modeling set at each depth exceeded 0.6, and the average RMSE were below 0.4%, indicating that the model had high accuracy and good stability.

3.5 Performance of salinity estimations under different vegetation coverage

In this study, the LAI was used as a proxy for vegetation cover, which is divided into three different levels: low, medium and high. Correspondingly, soil salinity data are stratified based on these cover classes. The results of SSC estimation under various covers were presented in Fig. 7.

The accuracy of the salt estimation models at various depths was assessed based on the salt data collected under different vegetation coverage levels, and corresponding scatter plots were created. As shown in Fig. 5, all models achieved R² values greater than 0.4, with the highest value reaching 0.83, indicating that the model's accuracy was highest for soil surface with intermediate vegetation covering. The models demonstrated good accuracy for soil depths of 0 to 10 cm under each vegetation coverage level, with R² values exceeding 0.7 and low RMSE values. At a depth of 10 to 20 cm, the model accuracy was highest under low vegetation covering, while the accuracy under medium vegetation coverage was lowest. However, at a depth of 20 to 30 cm, the maximum validation accuracy was achieved with medium coverage, with R² values exceeding 0.7 across all three coverages levels. With considerable vegetation coverage, the maximum accuracy at 30 to 40 cm was achieved, with an RMSE of 0.241% and an R² of 0.78. Generally, when using the data selected under each vegetation coverage level as the validation set, the favored models at different depths showed good estimation performance.

Soil salinity estimation by UAV multispectral remote sensing data plays a positive role in salinity monitoring and management. In this study, the spectral data and salinity data acquired during the vegetation cover period were used to construct and validate the estimation model for soil salinity at different depths. In addition, the model performances were validated under different vegetation cover ratios and the impact of cover on salt estimation in different soil layers was also explored.

Feature selection is an important part of constructing the soil salinity estimation model. Through the BDT method, 25 key features were finally identified, including bands reflectance (e.g., blue, red-edge, and near-infrared), spectral indices (e.g., SI2-reg, EVI, and DVI, etc.), and texture data (e.g., the mean, variance, and second-order moments of the GLCM features, etc.). These features efficiently captured the spectral properties of the soil and the spatial structural information, providing the model with sufficient input variables. Numerous studies have demonstrated that the red edge band is more sensitive to soil salinity and is strongly correlated the spectral properties of the vegetation canopy, thereby improving the accuracy of soil salinity estimation^30,52. To facilitate modeling and prediction, Hu et al.⁵³ employed UAV to gather remote sensing images of areas with both vegetation cover and barren ground. The regions with vegetation cover showed the most accurate prediction results. As a result, the spectral index produced by incorporating the red edge band significantly increases the accuracy of the soil salinity estimation in saline agricultural lands with vegetation cover. The red edge and near-infrared bands showed high importance at all depths in this study, as they sensitively revealed the changes in soil moisture and salt content, serving as important indicators for soil salinity prediction. This result aligned with the reported of Taghadosi et al.⁵⁴. Spectral indices such as EVI, EVI-reg and BI indirectly indicated the salinity status of the soil by enhancing the vegetation characteristics. Lobell et al.⁵⁵ showed that EVI is more reliable than NDVI for salinity monitoring. Spectral index groups, such as SI2-reg, EVI-reg, and DVI-reg, involved the calculation of the red-edge band, further proving the importance of the red-edge band in soil salinity monitoring, consistent with the results of Ma et al.⁵⁶ Textural features, such as the mean, variance, and contrast in the GLCM, were instrumental in capturing subtle soil surface changes and spatial structure information. Tai⁵⁷ showed that incorporating texture features could effectively improve the accuracy of soil salinity estimation models under vegetation cover conditions.

Different modelling groups also exerted an influence on estimation model performance. In the single-variable modelling group, spectral indices had a greater contribution to soil salinity estimation, with the RF model demonstrating high accuracy across all four depths (Fig. 4). Bian et al.⁵⁸ found that there is a certain correlation between spectral index and soil salinity, which has a great contribution to the application of estimation. Although the accuracy of the model built by band reflectance was slightly lower than that of spectral indices, the overall difference was not significant, which was in agreement with Wang et al.⁵⁹, who evaluated the prediction accuracy of different variable groups for soil salinity in oasis. In contrast, texture data performed poorly among the single variables. However, when combined with other variables, the texture information from multispectral images could improve the accuracy of soil salinity estimation at different depths of vegetation cover. Zheng et al.¹⁸ showed that the combination of texture data and spectral information significantly improved the accuracy of rice biomass estimation. The texture features of the UAV multispectral images provided rich information⁶⁰, making them applicable for estimating soil salinity at different depths. Incorporating texture data significantly enhances model performance, particularly in the RF model. When texture data is combined with other data, the accuracy of the model is notably improved. This study also demonstrates that using sensitive bands and spectral indices as the input variables leads to better modeling and validation outcomes in soil salinity estimation models. Nevertheless, in multi-variate groups, variables may affect and constrain each other, and combining variables of high importance not always achieve optimal results. For instance, the accuracy of all groups in the BPNN model was not optimal compared to other groups. Introducing an excessive number of independent variables could lead to information redundancy, overfitting, and a decrease in model accuracy⁶¹.

Previous studies have highlighted the superiority of machine learning methods in estimating soil salinity content due to the complexity and indirect relationships between variables⁶². In this study, RF, SVM, GPR and BPNN were used to model soil salinity at different depths and to select the most effective model. The evaluation of model criteria revealed that RF and GPR have advantages in estimating soil salinity. Although SVM has strong generalization capability for handling nonlinear problems, it is sensitive to noisy data⁶³. The models exhibited varying levels of accuracy in soil salinity estimation at different depths. The GPR model demonstrates better prediction accuracy than other methods for surface and deep soils, while the RF model performed well in intermediate soils. Due to the complexity of soil salinization mechanism and the intricate nonlinear relationship between soil spectral characteristics and soil salinity, machine learning is well-suited for revealing these relationships. Their strong nonlinear fitting and generalization capabilities, make them ideal for simulating the complex interactions among variables. Ma et al.⁵⁶ found that the RF model had better estimation results when predicting the soil salinity in Ebinur Lake wetland using multispectral and Digital Elevation Model (DEM) data. Therefore, machine learning algorithms have shown superior performance in both modeling -and validation of soil salinity estimation model. For instance, Wei et al.⁵ developed SSC estimation models using RF, SVM, and BPNN algorithms, with the RF model achieving the highest prediction accuracy (R² = 0.84). Additionally, Zhu et al.⁶⁴ and Yu et al.⁶⁵ concluded that the RF based SSC estimation model had a high accuracy.

The essence of estimating soil salinity during the vegetation cover period is to indirectly obtain information on soil salinity through the crop spectral response to soil salinity. The vegetation cover ratio of the given crop in the same period may correlate with the degree of soil salinity. The optimal estimation model for various soil depth varied with vegetation cover conditions, with the optimal model on depths ranging from 0ཞ10 cm under both low and medium vegetation cover conditions, and from 20ཞ30 cm under high vegetation cover. During the vegetation cover period, the absorption of soil water by crops was mainly done by the lateral root. Consequently, the soil salinity in the soil layer where the lateral roots were located will significantly affect crop water absorption⁶⁶, causing crop growth stress. Higher soil salinity levels result in more severe stress and poorer vegetation growth, which would be indirectly expressed in the vegetation canopy.

In this study, SSC at different depths was estimated and analyzed based on UAV multi-spectral data by composing seven type groups from band reflectance, spectral index and texture feature. Moreover, feature variables selection based on BDT method and four machine learning algorithms (RF, SVM, GPR and BPNN) were incorporated in estimation model establishment. (1) The estimation accuracy of the RF and GPR models surpasses that of the SVM and BPNN models across all four soil depths in the SSC estimation models. (2) The GPR model, based on the entire variable group, is the most accurate for estimating soil salt content at depths of 0 ~ 10 cm and 30 ~ 40 cm. The R² values for the validation set were 0.774 and 0.705, and the modeling set, 0.77 and 0.62, with corresponding RMSE of 0.185% and 0.31%, respectively. The RF model, based on the spectral index and texture data set, is the most accurate estimation model for depths of 10 to 20 cm and 20 to 30 cm. (3) The model was further validated using salt data under varying vegetation covers. The highest validation accuracy was obtained with medium vegetation coverage at depths of 0 ~ 10 cm and 20 ~ 30 cm, as well as high vegetation coverage at depths of 10 ~ 20 cm and 30 ~ 40 cm. (4) The model imputes were screened using the BDT approach to select the independent variables at each depth. The selected results demonstrated that the model constructed using only texturing data group had low accuracy, but the accuracy increased dramatically when additional variables were included. There are still some problems and limitations in this study. This experiment was conducted under box planting conditions, and the experimental results have some limitations, which may not fully reflect the soil salinity conditions under actual field conditions. In addition, the study has not yet considered the effects of crop type and soil water content on the estimation model, which may have a significant impact on model accuracy in practical applications. Therefore, future studies need to consider more influencing factors in order to further optimize the estimation model and improve the accuracy and stability of soil salinity prediction.

Competing Interests Statement

The authors declare no competing interests.

Funding

This research was funded by the National Natural Science Foundation of China (52379049, 52209071), the Postgraduate Research & Practice Innovation Program of Jiangsu Province (Yangzhou University) (No. SJCX23_1945), the China Postdoctoral Science Foundation (2023T160552, 2020M671623), “Chunhui Plan” Cooperative Scientific Research Project of Ministry of Education of China (HZKY20220115), the “Blue Project” of Yangzhou University and the Priority Academic Program Development of Jiangsu Higher Education Institutions (PAPD).

Author Contribution

Conceptualization, C.Z.; methodology, Z.L.; software, Z.L.; validation, M.D. and M.T.; formal analysis, Z.L.; resources, M.D.; data curation, M.D. and R.L.; Writing—original preparation, Z.L.; Writing—review and editing, C.Z. and M.T.; visualization, Z.Z.; supervision, C.Z. and S.F.; project administration, C.Z.; funding acquisition, C.Z. and S.F. All authors have read and agreed to the published version of the manuscript.

Acknowledgement

The authors would like to acknowledge the Agricultural Water and Hydrological Ecology Experiment Station, Yangzhou University, for providing the experimental facilities.

Data Availability

The datasets used and analyzed during the current study are available from the corresponding author upon reasonable request.

Seifi, M., Ahmadi, A., Neyshabouri, M. R., Taghizadeh-Mehrjardi, R. & Bahrami, H. A. Remote and Vis-NIR spectra sensing potential for soil salinization estimation in the eastern coast of Urmia hyper saline lake, Iran20100398 (Society and Environment, 2020).
Wang, L. W. & Wei, Y. X. Estimating the total nitrogen and total phosphorus content of wetland soils using hyperspectral models. Acta Ecol. Sin. 36 (16), 5116–5125 (2016).
Rao, B. R. M. et al. Spectral behaviour of salt-affected soils. Int. J. Remote Sens. 16 (12), 2125–2136 (1995).
Zhao, W., Zhou, C., Zhou, C., Ma, H. & Wang, Z. Soil salinity inversion model of oasis in arid area based on UAV multispectral remote sensing. Remote Sens. 14 (8), 1804 (2022).
Wei, G. et al. Estimation of soil salt content by combining UAV-borne multispectral sensor and machine learning algorithms. PeerJ, 8, e9087. (2020).
Chen, J. et al. UAV remote sensing inversion of soil salinity in field of sunflower. Trans. Chin. Soc. Agricultural Mach. 51 (7), 178–191 (2020).
Ivushkin, K. et al. UAV based soil salinity assessment of cropland. Geoderma. 338, 502–512 (2019).
Nicolas, H. & Walter, C. Detecting salinity hazards within a semiarid context by means of combining soil and remote-sensing data. Geoderma. 134 (1–2), 217–230 (2006).
Qi, G., Zhao, G. & Xi, X. Soil salinity inversion of winter wheat areas based on satellite-unmanned aerial vehicle-ground collaborative system in coastal of the Yellow River Delta. Sensors. 20 (22), 6521 (2020).
Wang, N. et al. Integrating remote sensing and landscape characteristics to estimate soil salinity using machine learning methods: A case study from Southern Xinjiang, China. Remote Sens. 12 (24), 4118 (2020).
Farifteh, J., Van der Meer, F., Atzberger, C. & Carranza, E. J. M. Quantitative analysis of salt-affected soil reflectance spectra: A comparison of two adaptive methods (PLSR and ANN). Remote Sens. Environ. 110 (1), 59–78 (2007).
Xu, L. & Wang, Q. Retrieval of soil water content in saline soils from emitted thermal infrared spectra using partial linear squares regression. Remote Sens. 7 (11), 14646–14662 (2015).
Ge, X. et al. Combining UAV-based hyperspectral imagery and machine learning algorithms for soil moisture content monitoring. PeerJ, 7, e6926. (2019).
Poblete, T., Ortega-Farías, S., Moreno, M. A. & Bardeen, M. Artificial neural network to predict vine water status spatial variability using multispectral information obtained from an unmanned aerial vehicle (UAV). Sensors. 17 (11), 2488 (2017).
Hu, J. Estimation of Soil Salinity in Arid Area Based on Multi-Source Remote Sensing (Doctoral dissertation, Zhejiang University). (2019).
Zhang Fei, Z. F. et al. I., Extracting of soil salinization by SVM and accuracy evaluation based on texture characteristic. (2009).
Zhang, Z. et al. Inversion of soil salt content by UAV remote sensing under different vegetation coverage. Trans. Chin. Soc. Agric. Mach. 53, 220–230 (2022).
Zheng, H. et al. Enhancing the nitrogen signals of rice canopies across critical growth stages through the integration of textural and spectral information from unmanned aerial vehicle (UAV) multispectral imagery. Remote Sens. 12 (6), 957 (2020).
Yang, H., Wang, Z., Cao, J., Wu, Q. & Zhang, B. Estimating soil salinity using Gaofen-2 imagery: A novel application of combined spectral and textural features. Environ. Res. 217, 114870 (2023).
Ren, J., Li, X., Zhao, K., Fu, B. & Jiang, T. Study of an on-line measurement method for the salt parameters of soda-saline soils based on the texture features of cracks. Geoderma. 263, 60–69 (2016).
Andrade, R. et al. Proximal sensor data fusion and auxiliary information for tropical soil property prediction: Soil texture. Geoderma. 422, 115936 (2022).
Drielsma, M. J., Love, J., Taylor, S., Thapa, R. & Williams, K. J. General Landscape Connectivity Model (GLCM): a new way to map whole of landscape biodiversity functional connectivity for operational planning and reporting. Ecol. Model. 465, 109858 (2022).
Cai, S., Zhang, R., Liu, L. & Zhou, D. A method of salt-affected soil information extraction based on a support vector machine with texture features. Math. Comput. Model. 51 (11–12), 1319–1325 (2010).
Huang, C. et al. Land cover mapping in cloud-prone tropical areas using Sentinel-2 data: Integrating spectral features with Ndvi temporal dynamics. Remote Sens. 12 (7), 1163 (2020).
Jia, K. L. & Zhang, J. H. Impacts of different alkaline soil on canopy spectral characteristics of overlying vegetation. Spectrosc. Spectr. Anal. 34 (3), 782–786 (2014).
Zhang, J. et al. Estimating soil salinity with different fractional vegetation cover using remote sensing. Land. Degrad. Dev. 32 (2), 597–612 (2021).
Wang, F. et al. Sensitivity analysis of soil salinity and vegetation indices to detect soil salinity variation by using Landsat series images: applications in different oases in Xinjiang, China. Acta Ecol. Sin. 37 (15), 5007–5022 (2017).
Chi, Y., Sun, J., Liu, W., Wang, J. & Zhao, M. Mapping coastal wetland soil salinity in different seasons using an improved comprehensive land surface factor system. Ecol. Ind. 107, 105517 (2019).
Khan, N. M., Rastoskuev, V. V., Shalina, E. V. & Sato, Y. Mapping salt-affected soils using remote sensing indicators-a simple approach with the use of GIS IDRISI. In 22nd Asian conference on remote sensing (Vol. 5, No. 9). (2001), November.
Zhang, S. et al. Integrated satellite, unmanned aerial vehicle (UAV) and ground inversion of the SPAD of winter wheat in the reviving stage. Sensors. 19 (7), 1485 (2019).
Song, C. Y., Ren, H. X. & Huang, C. Estimating soil salinity in the Yellow River Delta, Eastern China—an integrated approach using spectral and terrain indices with the generalized additive model. Pedosphere. 26 (5), 626–635 (2016).
Abbas, A., Khan, S., Hussain, N., Hanjra, M. A. & Akbar, S. Characterizing soil salinity in irrigated agriculture using a remote sensing approach5543–52 (Physics and chemistry of the Earth, 2013).
Tripathi, N. K., Rai, B. K. & Dwivedi, P. Spatial modeling of soil alkalinity in GIS environment using IRS data. In Proceedings of the 18th Asian Conference on Remote Sensing, Kuala Lumpur, Malaysia (Vol. 20, No. 25, pp. 81–86). (1997), October.
Allbed, A., Kumar, L. & Aldakheel, Y. Y. Assessing soil salinity using soil salinity and vegetation indices derived from IKONOS high-spatial resolution imageries: Applications in a date palm dominated region. Geoderma. 230, 1–8 (2014).
Zhou, L. et al. Research on quick dynamic monitoring of soil salinization in Qaidam Basin[J]. Science of Surveying and Mapping, 46(7): 99 ~ 106, 114. (2021).
Khan, N. M., Rastoskuev, V. V., Sato, Y. & Shiozawa, S. Assessment of hydrosaline land degradation by using a simple approach of remote sensing indicators. Agric. Water Manage. 77 (1–3), 96–109 (2005).
Liu, H. Q. & Huete, A. A feedback based modification of the NDVI to minimize canopy background and atmospheric noise. IEEE Trans. Geosci. Remote Sens. 33 (2), 457–465 (1995).
Lyon, J. G., Yuan, D., Lunetta, R. S. & Elvidge, C. D. A change detection experiment using vegetation indices. Photogram. Eng. Remote Sens. 64 (2), 143–150 (1998).
Birth, G. S. & McVey, G. R. Measuring the color of growing turf with a reflectance spectrophotometer 1. Agron. J. 60 (6), 640–643 (1968).
Arachchi, M. H., Field, D. J. & McBratney, A. B. Quantification of soil carbon from bulk soil samples to predict the aggregate-carbon fractions within using near-and mid-infrared spectroscopic techniques. Geoderma. 267, 207–214 (2016).
Ma, S., He, B., Ge, X. & Luo, X. Spatial prediction of soil salinity based on the Google Earth Engine platform with multitemporal synthetic remote sensing images. Ecol. Inf. 75, 102111 (2023).
Gitelson, A. A., Kaufman, Y. J., Stark, R. & Rundquist, D. Novel algorithms for remote estimation of vegetation fraction. Remote Sens. Environ. 80 (1), 76–87 (2002).
Elsherbiny, O., Zhou, L., Feng, L. & Qiu, Z. Integration of visible and thermal imagery with an artificial neural network approach for robust forecasting of canopy water content in rice. Remote Sens. 13 (9), 1785 (2021).
Xiao, Y., Dong, Y., Huang, W., Liu, L. & Ma, H. Wheat fusarium head blight detection using UAV-based spectral and texture features in optimal window size. Remote Sens. 13 (13), 2437 (2021).
Drucker, H. & Cortes, C. Boosting decision trees. Advances in neural information processing systems, 8. (1995).
Mutanga, O., Adam, E. & Cho, M. A. High density biomass estimation for wetland vegetation using WorldView-2 imagery and random forest regression algorithm. Int. J. Appl. Earth Obs. Geoinf. 18, 399–406 (2012).
Huang, X. Y., Wang, X. M. & KAWUQIATI, B. Inversion of soil salinity of an oasis in an arid area based on Landsat8 OLI images. Remote Sens. Nat. Resour., 35(1). (2023).
Sui, H. et al. Soil salinity estimation over coastal wetlands based on random Forest algorithm and hydrological connectivity metric. Front. Mar. Sci. 9, 895172 (2022).
Tan, K., Ye, Y. Y., Du, P. J. & Zhang, Q. Q. Estimation of heavy metal concentrations in reclaimed mining soils using reflectance spectroscopy. Spectrosc. Spectr. Anal. 34 (12), 3317–3322 (2014).
Guan, X., Wang, S., Gao, Z. & Lv, Y. Dynamic prediction of soil salinization in an irrigation district based on the support vector machine. Math. Comput. Model. 58 (3–4), 719–724 (2013).
Wang, J., Wang, W., Hu, Y., Tian, S. & Liu, D. Soil moisture and salinity inversion based on new remote sensing index and neural network at a salina-alkaline wetland. Water. 13 (19), 2762 (2021).
Kaplan, G. & Avdan, U. Evaluating the utilization of the red edge and radar bands from sentinel sensors for wetland classification. Catena. 178, 109–119 (2019).
Hu, J. et al. Quantitative estimation of soil salinity using UAV-borne hyperspectral and satellite multispectral images. Remote Sens. 11 (7), 736 (2019).
Taghadosi, M. M., Hasanlou, M. & Eftekhari, K. Retrieval of soil salinity from Sentinel-2 multispectral imagery. Eur. J. Remote Sens. 52 (1), 138–154 (2019).
Lobell, D. B. et al. Regional-scale assessment of soil salinity in the Red River Valley using multi‐year MODIS EVI and NDVI. J. Environ. Qual. 39 (1), 35–41 (2010).
Ma, G., Ding, J., Han, L. & Zhang, Z. Digital mapping of soil salinization in arid area wetland based on variable optimized selection and machine learning. Trans. Chin. Soc. Agric. Eng. 36, 124–131 (2020).
Tai, X. Soil Salinity Monitoring Model on UAV Multi Spectral Remote Sensing Vegetation Cover[D]. Northwest Agricultural and Forestry University, DOI: (2022).
10.27409/d
Bian, L., Wang, J., Guo, B., Cheng, K. & Wei, H. Remote sensing extraction of soil salinity in Yellow River Delta Kenli County based on feature space. Remote Sens. Technol. Application. 35 (1), 211–218 (2020).
Wang, F. et al. Environmental sensitive variable optimization and machine learning algorithm using in soil salt prediction at oasis. Trans. Chin. Soc. Agric. Eng. 34, 102–110 (2018).
Guo, Y. et al. Integrating spectral and textural information for identifying the tasseling date of summer maize using UAV based RGB images. Int. J. Appl. Earth Obs. Geoinf. 102, 102435 (2021).
Morellos, A. et al. Machine learning based prediction of soil total nitrogen, organic carbon and moisture content by using VIS-NIR spectroscopy. Biosyst. Eng. 152, 104–116 (2016).
Cui, X. et al. Estimating soil salinity under sunflower cover in the Hetao Irrigation District based on unmanned aerial vehicle remote sensing. Land. Degrad. Dev. 34 (1), 84–97 (2023).
Shieh, H. L. & Kuo, C. C. A reduced data set method for support vector regression. Expert Syst. Appl. 37 (12), 7781–7787 (2010).
Zhu, C., Ding, J., Zhang, Z. & Wang, Z. Exploring the potential of UAV hyperspectral image for estimating soil salinity: Effects of optimal band combination algorithm and random forest. Spectrochim. Acta Part A Mol. Biomol. Spectrosc. 279, 121416 (2022).
Yu, X., Chang, C., Song, J., Zhuge, Y. & Wang, A. Precise monitoring of soil salinity in China’s Yellow River Delta using UAV-borne multispectral imagery and a soil salinity retrieval index. Sensors. 22 (2), 546 (2022).
Zhang, Z. T. et al. UAV Multispectral Remote Sensing Soil Salinity Inversion Based on Different Fractional Vegetation Coverages[J]. Trans. Chin. Soc. Agricultural Mach., 53(08): 220–230. (2022).

No competing interests reported.

Download PDF

Editorial decision: Revision requested
30 Oct, 2024
Reviewers agreed at journal
19 Oct, 2024
Reviews received at journal
19 Oct, 2024
Reviews received at journal
09 Oct, 2024
Reviewers agreed at journal
02 Oct, 2024
Reviewers agreed at journal
28 Sep, 2024
Reviewers invited by journal
26 Sep, 2024
Editor assigned by journal
26 Sep, 2024
Editor invited by journal
05 Sep, 2024
Submission checks completed at journal
03 Sep, 2024
First submitted to journal
25 Aug, 2024

You are reading this latest preprint version

Estimating soil profile salinity under vegetation cover based on UAV multi-source remote sensing

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials and methods

2.1 Study area and experimental design

2.2 Data collection and acquisition

2.2.1 Soil salt measurement

2.2.2 Multispectral data acquisition and processing

2.2.3 Vegetation coverage calculation

2.3 Construction of feature variables

2.4 Feature variable selection and set construction

2.4.1 Optimization of characteristic variables

2.4.2 Construction of model training

2.5 Machine learning models

2.6 Technical workflow

3. Results

3.1 Statistical of soil salinity distribution

3.2 Contribution and selection of feature variables

3.3 Comparision of model performance for various ML methods and data groups

3.4 Outer-sample validation and analysis

3.5 Performance of salinity estimations under different vegetation coverage

4. Discussion

5. Conclusion

Declarations

Funding

Author Contribution

Acknowledgement

Data Availability

References

Additional Declarations

Status:

Version 1