Soil Salinity Inversion of Winter Wheat Areas Based on Satellite-Unmanned Aerial Vehicle-Ground Collaborative System in Coastal of the Yellow River Delta

Soil salinization is an important factor affecting winter wheat growth in coastal areas. The rapid, accurate and efficient estimation of soil salt content is of great significance for agricultural production. The Kenli area in the Yellow River Delta was taken as the research area. Three machine learning inversion models, namely, BP neural network (BPNN), support vector machine (SVM) and random forest (RF) were constructed using ground-measured data and UAV images, and the optimal model is applied to UAV images to obtain the salinity inversion result, which is used as the true salt value of the Sentinel-2A image to establish BPNN, SVM and RF collaborative inversion models, and apply the optimal model to the study area. The results showed that the RF collaborative inversion model is optimal, R2 = 0.885. The inversion results are verified by using the measured soil salt data in the study area, which is significantly better than the directly satellite remote sensing inversion method. This study integrates the advantages of multi-scale data and proposes an effective “Satellite-UAV-Ground” collaborative inversion method for soil salinity, so as to obtain more accurate soil information, and provide more effective technical support for agricultural production.


Introduction
Soil salinization is a form of soil degradation. Soil salinization will not only cause a series of problems such as ecological deterioration, but also have a negative impact on the growth of crops [1]. Therefore, it is of great practical significance for agricultural production to carry out research on soil salinization in coastal winter wheat planting areas and grasp the spatial distribution of salinization.
Traditional soil salt information is obtained mainly through field survey sampling and chemical analysis method. This method is relatively accurate, but it is time-consuming and labor-intensive, and has obvious limitations in terms of spatial globalization and effectiveness. In addition, field sampling can also cause damage to winter wheat and other crops. Satellite remote sensing data can make up for these shortcomings. Because of its high timeliness, economy and large-area simultaneous observation ability, it has become an important method for quantitative extraction of saline soil information in a large area. A large number of scholars have carried out related research and achieved excellent results. Most of them have realized a large-scale quantitative inversion of soil salt content based on multi-source satellite images [2][3][4][5][6][7][8][9][10][11], and they inverted soil salinity by studying the quantitative sunlight but little precipitation and evaporation large, uneven droughts and floods, and obvious seasonal alternating wet and dry. The annual mean precipitation, evaporation and air temperature are 511.6 mm, 1928.2 mm and 12.4 °C, respectively [28]. The terrain of the study area is low and flat, decreasing from southwest to northeast. The source of surface water is natural precipitation and water from the Yellow River. The groundwater level is relatively shallow and the salinity is high. The main soil type is gleyic solonchaks with a high sand proportion and high salinization. Soil texture is light, organic matter is generally lack, nitrogen and phosphorus are less, soil overall nutrient is poor, the pH value of the soil is greater than 7 [29]. The main crops are winter wheat, corn, rice and cotton, but the overall management is extensive and the yield is low. Wheat planting areas are mainly distributed in the higher terrain area in the southwest and the Yellow River coast area in the northeast. The variation of soil salinization is obvious, which is an ideal area for this study. Based on the investigation of Kenli area, landform, soil and crop distribution, two test areas A and B (Figure 1) were selected for the concentrated distribution of winter wheat in the southwest and northeast of the study area respectively to carry out field soil salinity measurement and UAV flight test. Test area A was a square area of 200 m × 200 m, and a 100 m × 50 m area was selected to arrange sampling points. Test area B was a rectangular area of 200 m × 100 m, and a 50 m × 50 m area was selected to arrange sampling points. An investigation suggests that planting time, farming methods, and fertilization conditions are all basically the same in A and B test areas, while the growth of winter wheat there is obviously different, with all soil salt content levels distributed, making the experimental area more typical and representative.

Acquisition of Ground Soil Salinity Data
There is less rainfall in spring in Kenli District, and the surface salinization is obvious and stable. Winter wheat is at the reviving stage, while other major crops are not sown, which is convenient for extracting spectral characteristics [30]. A field survey in the study area was conducted from April 10 to 16, 2019. To ensure uniform distribution, three samples were pre-arranged in the study area every 5 km × 5 km grid, and finally 77 winter wheat sample data were collected. At the same time, for the two test areas, a ground card was placed every 10m at the outer boundary of the sampling point area of the A and B test areas, and connect them with a measuring rope to form a 10 m × 10 m grid of sample points, taking the intersection of the grid as sampling points. This 10 m × 10 m grid is the same as the pixel size of Sentinel-2A. A total of 102 sample points was collected, including 66 in the A test area and 36 in the B test area. Four outliers in the sampling points were eliminated, and the remaining 98 samples were used to construct and verify the soil salinity inversion model of winter wheat. An EC110 portable salinity meter (Spectrum Technologies, Inc., Aurora, USA) equipped with a 2225FST series probe (conductivity temperature correction was performed) was used to make multiple measurements of the electrical conductivity (EC) of the soil surface layer 10 cm below the plant at each sample point and make a record after stabilization. The average of the measured values is taken as the EC value of each sample point, in ds/m. According to the results of earlier research which in our laboratory, the measured EC data were converted into soil salt content (SSC) in g/kg by using the formula SSC = 2.18 EC + 0.727 which was obtained by chemical analysis of soil in the same study area in spring [31]. At the same time, the orientation, topography, soil, wheat growth and other relevant information were recorded.

Acquisition and Processing of UAV Imagery
A multispectral camera (Parrot Sequoia, Parrot Inc., Paris, France) was mounted on a Dajiang Matrice 600 Pro UAV (loaded mass: 5.5 kg; flying time: 18 min) (SZ DJI Technology Co., Ltd. Shenzhen, Guangdong Province, China). The camera can receive a total of 4 bands of information, which are green light (G), red light (R), red edge (RE), near infrared (NIR). The wavelengths are 550 nm, 660 nm, 735 nm, and 790 nm, and the band widths are 40 nm, 40 nm, 10 nm and 40 nm. The Sequoia multi-spectral camera is mounted on the head of the UAV, and the radiation sensor is mounted on the top of the UAV to write the radiometric correction data into the image during flight.
The data collection time was from 11:00 to 15:00 on 14 April 2019. The weather was clear and cloudless with low wind force when the UAV was flying. Before takeoff, the Sequoia multispectral camera and radiation sensor were calibrated, and the ground standard whiteboard image was collected. The flight height was 50 m, the flight speed was 5 m/s, and the image acquisition interval was 1.5 s. After the data are collected, they will be imported into Pix4D Mapper software (Pix4D, S.A., Prilly, Switzerland) for splicing, radiation correction and other processing to obtain the high-resolution orthophoto image of the test area, with a spatial resolution of 4-5 cm. Finally, in ENVI5.3, the decision tree method is used to remove the soil background. In order to eliminate the random error caused by the reflectance of a single point, a 5 × 5 pixels image is taken with the sampling point as the center, and the average reflectance value is taken as the reflectance data of the sampling point.

Acquisition and Processing of Sentinel-2A Satellite Data
Sentinel-2 satellite is a multispectral imaging satellite with high resolution, revisit rate and update rate. It includes two small satellites A and B. The revisit period is 5 days. The main payload is MSI multispectral imager, covering 0.4-2.4 µm spectral range, including 10 m (four bands), 20 m (six bands), 60 m (three bands) ground resolution, which can monitor the growth, coverage and health of land vegetation, and obtain information on crop planting, land use changes, etc. In this paper, the Sentinel-2A products were downloaded from the ESA Copernicus data sharing website  the acquisition time of ground and UAV data and the  quality of the image, the Sentinel-2A Level-1C multispectral image on 17 April 2019 was selected for  modeling and inversion of the soil salt content, and the images of 3 November 2018 and 26 June 2019 were used for extraction of winter wheat planting areas.
The downloaded L1C data are orthophoto with geometric precision correction, without radiometric calibration and atmospheric correction. First, radiometric calibration and atmospheric correction are carried out by using Sen2cor, a plug-in published by ESA. Then, the data are re-sampled by the Sentinel Application Platform (SNAP) software to generate 10m spatial resolution images, and the data are exported in ENVI format. Finally, the Sentinel-2A true color images of the research area were obtained by splicing the images in ENVI 5.3 (Exelis Visual Information Solutions, Inc., Colorado, USA) and clipping the images using kenli District administrative boundary vector documents ( Figure 1). In order to be consistent with the UAV image band, the green (G), red (R), red edge (RE), and near infrared (NIR) of Sentinel-2A image are selected in this study. The wavelengths are 560 nm, 665 nm, 740 nm, 865 nm, and the band widths are 45 nm, 38 nm, 18 nm and 33 nm.

Calculation and Optimization of Vegetation Indices
Studies have shown that different levels of soil salinization have an impact on vegetation growth and morphology, plasma membrane permeability, photosynthetic pigments of leaves, gas exchange parameters, chlorophyll fluorescence characteristics, etc. [32]. Therefore, there are differences in spectral information of vegetation at different levels of soil salinization, which can indirectly reflect the level of soil salinization [33,34]. The vegetation index can highlight the characteristics of vegetation and effectively reflect the health and growth of vegetation. When the soil salt content increases, the reflectivity of visible red light of the salt-sensitive vegetation will increase, and the near-infrared reflectance will decrease [35].In order to better reflect the vegetation conditions, 8 vegetation indexes related to red light and near-infrared are selected in this study, including normalized difference vegetation index (NDVI), normalized difference red edge index (NDRE), optimized soil adjusted vegetation index (OSAVI), green normalized difference vegetation index (GNDVI), triangle vegetation index (TVI), difference vegetation index (DVI), Improved chlorophyll absorption vegetation index based on PROSPECT and SAILH radiation transfer model (MCARI2) and renormalized difference vegetation index (RDVI).
The multi-band spectrum collected by the UAV is used to calculate the 8 vegetation indexes, and the formulas are shown in Table 1. Then the correlation coefficient between each vegetation index and soil salinity was calculated, and the variance inflation factor (VIF) between vegetation indexes was calculated by using the formula VIF = 1/(1-r*r) (r is the correlation coefficient between vegetation indexes) [12], excluding the low correlation or VIF > 10, which is the parameter that cannot be diagnosed by collinearity. The sensitive vegetation indexes are selected for soil salt modeling. The 98 samples were sorted from small to large, and the modeling set and the validation set were sampled at equal intervals in a ratio of 2:1 to ensure the same range and uniform distribution of the model samples and the validation samples. 68 samples were selected for modeling and 30 samples for validation.
Taking the sensitive vegetation index as the input variable of the model, three methods were used to construct the winter wheat soil salinity inversion model, namely, BPNN, SVM and RF. BPNN is a multi-layer feedforward neural network trained according to the error back propagation algorithm, and it has also been applied to the salt inversion problem [39,40]. SVM is a new machine learning method from linear separable to linear nonseparable based on the principle of minimizing structural risk according to the statistical theory. It has been widely applied in image recognition and classification and has also been applied in regression problems in recent years [41,42]. The RF algorithm is an integrated learning algorithm obtained by combining the bagging algorithm with the decision tree algorithm. In recent years, many scholars have applied it to remote sensing technology [43,44]. All the three methods were implemented in MATLAB R2016b (MathWorks, Inc., Natick, MA, USA). The BPNN sets the number of training iterations to 1000, the accuracy to 0.003, the learning rate to 0.01; The SVM method was set as V-SVR, Gaussian kernel function was selected, the best penalty factor C and kernel parameter gamma were selected through network search and cross validation, and the SVM model was trained. The RF method called MATLAB random forest toolbox, and the parameter leaf node Leaf = 5 and the number of trees Ntrees = 200 were finally determined by Bayesian optimization.
The accuracy of model modeling and verification was evaluated by the coefficient of determination (R 2 ) and root mean square error (RMSE) [45][46][47]. R 2 is used to measure the fitting degree of the model, and RMSE reflects the deviation between measured value and predicted value. The closer R 2 is to 1, the smaller the RMSE, which means the higher the accuracy of the model, the better the effect. The model with the best accuracy and effect was selected for soil salinity inversion of winter wheat. The degree of soil salinization is divided into five grades according to relevant criteria [48], non-salinization (<1 g/kg), mild salinization (1-2 g/kg), and moderate salinization (2-4 g/kg), severe salinization (4-6 g/kg) and saline soil (>6 g/kg), and we get the distribution map of soil salinity grade.

Information Extraction of Winter Wheat Planting Area
The planting area of winter wheat in the study area was extracted by using the time series features composed of the NDVI of the Sentinel-2A images on 3 November 2018, 17 April 2019 and 26 June 2019. According to the investigation and analysis of various vegetation types in the study area, only winter wheat was in the seedling stage in early November, its growth reached the peak stage in late April of the second year, and entered the maturity stage at the end of June. Its NDVI time series curve showed a rapid rise from early November to late April, while other vegetation types showed little change in NDVI, the NDVI of winter wheat declined rapidly from late April to the end of June, while the NDVI time series curves of other vegetations were in varying degrees of rising stages, as shown in Figure 2. The valley-peak-valley in the NDVI time series at the beginning of November, late April and the end of June is a significant feature that distinguishes winter wheat from other vegetation. Therefore, by calculating the changes in NDVI from early November to late April, and late April to the end of June, a decision tree is established to extract this feature. Combined with the training sample data, the following decision rules are established: Thus, the distribution of winter wheat planting areas on 17 April 2019 was obtained. Sentinel-2A images have mixed pixels, and it is difficult to accurately obtain the measured data of soil salinity within the corresponding range of pixels. However, the resolution of UAV images is up to centimeter level, so the measured data of soil salinity within the corresponding range of pixels are easy to obtain and accurate. Therefore, the measured ground salt data and the UAV image vegetation index are used to construct a high-precision inversion model to obtain the soil salt content of the test areas A and B, then the salt values corresponding to sentinel-2 image pixels in the test area were calculated and taken as the "salt true value" of the Sentinel-2A image construction inversion model. It is combined with the Sentinel-2A image vegetation index to construct three inversion models including BPNN, SVM and RF.
First, the 10 m × 10 m surface vector data corresponding to the size positions of Sentinel-2A image pixels were successively established in the test area A and the test area B. In order to ensure the objectivity of the data, in the formed surface vector data grid, every three rows and three columns of surface vector data are used as a unit to extract the surface vector data at the center of the unit. If the center position is not in the unit, the surface vectors at other positions in the unit are extracted. Figure 3 is the distribution map of the area vector data extracted from the test area A and a total of 75 surface vector data are extracted. The vegetation index corresponding to the Sentinel-2A image pixel of the surface vector was counted and entered into the attribute information, and the combination of the vegetation indexes were used as the input variable of the model. Secondly, A vector surface corresponds to 40,000 image UAV pixels, and the average salt value of 40,000 UAV pixels is calculated as the salt "truth value" of the surface vector data. The soil salt content of 75 surface vector data was recorded into the attribute information as the output variable of the model. Finally, BPNN, SVM and RF were used to construct soil salt content collaborative inversion model based on Sentinel-2A image. All three methods were implemented in MATLAB R2016b.The BPNN sets the number of training iterations to 1000, the accuracy to 0.003, and the learning rate to 0.02; The SVM method was set as V-SVR, and the training set cross-validation and network search method were used to optimize the parameters. According to the principle of minimum variance, the penalty coefficient was determined as C = 10000, γ = 0.01; the RF method called MATLAB random forest toolbox, and the parameter leaf node, Leaf = 5, and the number of trees, Ntrees = 300, was finally determined by Bayesian optimization. Thus, the distribution of winter wheat planting areas on 17 April 2019 was obtained.

Construction of the Satellite-UAV-Ground Collaborative Inversion Model of Soil Salinity in Wheat Area Based on Sentinel-2 Images
Sentinel-2A images have mixed pixels, and it is difficult to accurately obtain the measured data of soil salinity within the corresponding range of pixels. However, the resolution of UAV images is up to centimeter level, so the measured data of soil salinity within the corresponding range of pixels are easy to obtain and accurate. Therefore, the measured ground salt data and the UAV image vegetation index are used to construct a high-precision inversion model to obtain the soil salt content of the test areas A and B, then the salt values corresponding to sentinel-2 image pixels in the test area were calculated and taken as the "salt true value" of the Sentinel-2A image construction inversion model. It is combined with the Sentinel-2A image vegetation index to construct three inversion models including BPNN, SVM and RF.
First, the 10 m × 10 m surface vector data corresponding to the size positions of Sentinel-2A image pixels were successively established in the test area A and the test area B. In order to ensure the objectivity of the data, in the formed surface vector data grid, every three rows and three columns of surface vector data are used as a unit to extract the surface vector data at the center of the unit. If the center position is not in the unit, the surface vectors at other positions in the unit are extracted. Figure 3 is the distribution map of the area vector data extracted from the test area A and a total of 75 surface vector data are extracted. The vegetation index corresponding to the Sentinel-2A image pixel of the surface vector was counted and entered into the attribute information, and the combination of the vegetation indexes were used as the input variable of the model. Secondly, A vector surface corresponds to 40,000 image UAV pixels, and the average salt value of 40,000 UAV pixels is calculated as the salt "truth value" of the surface vector data. The soil salt content of 75 surface vector data was recorded into the attribute information as the output variable of the model. Finally, BPNN, SVM and RF were used to construct soil salt content collaborative inversion model based on Sentinel-2A image. All three methods were implemented in MATLAB R2016b.The BPNN sets the number of training iterations to 1000, the accuracy to 0.003, and the learning rate to 0.02; The SVM method was set as V-SVR, and the training set cross-validation and network search method were used to optimize the parameters. According to the principle of minimum variance, the penalty coefficient was determined as C = 10000, γ = 0.01; the RF method called MATLAB random forest toolbox, and the parameter leaf node, Leaf = 5, and the number of trees, Ntrees = 300, was finally determined by Bayesian optimization. Sensors 2020, 20, x 8 of 18

Soil Salinity Inversion Results and Accuracy Analysis in Wheat Area
Among the three machine learning models constructed, the optimal model was selected according to R 2 and RMSE. The optimal model was used to invert the soil salinity of the winter wheat planting area in Kenli District, and the soil salinity distribution map of winter wheat in Kenli District was obtained.
At the same time, this method was compared with the direct inversion method of satellite remote sensing, namely, based on Sentinel-2A images and ground-measured salt data, the optimal model was directly constructed and selected to invert the winter wheat soil salt distribution in Kenli District. In order to verify the accuracy of the two methods, the R 2 and RMSE of the inversion results and the data of 77 winter wheat soil sampling points in Kenli District were calculated for quantitative evaluation.

Screening of Soil Salt-Sensitive Vegetation Index
Correlation analysis was conducted between UAV image vegetation index and measured soil salinity content respectively, as is shown in Table 2. In the correlation matrix, the DVI, MCARI2, and TVI of the eight vegetation indices had low correlations with soil salinity. The VIF value of OSAVI and NDVI was greater than 10, and there was strong multicollinearity. Therefore, the combination of vegetation index NDVI, NDRE, GNDVI and RDVI was selected as independent variables for modeling.

Soil Salinity Inversion Results and Accuracy Analysis in Wheat Area
Among the three machine learning models constructed, the optimal model was selected according to R 2 and RMSE. The optimal model was used to invert the soil salinity of the winter wheat planting area in Kenli District, and the soil salinity distribution map of winter wheat in Kenli District was obtained.
At the same time, this method was compared with the direct inversion method of satellite remote sensing, namely, based on Sentinel-2A images and ground-measured salt data, the optimal model was directly constructed and selected to invert the winter wheat soil salt distribution in Kenli District. In order to verify the accuracy of the two methods, the R 2 and RMSE of the inversion results and the data of 77 winter wheat soil sampling points in Kenli District were calculated for quantitative evaluation.

Screening of Soil Salt-Sensitive Vegetation Index
Correlation analysis was conducted between UAV image vegetation index and measured soil salinity content respectively, as is shown in Table 2. In the correlation matrix, the DVI, MCARI2, and TVI of the eight vegetation indices had low correlations with soil salinity. The VIF value of OSAVI and NDVI was greater than 10, and there was strong multicollinearity. Therefore, the combination of vegetation index NDVI, NDRE, GNDVI and RDVI was selected as independent variables for modeling.

UAV-Ground Collaborative Inversion Model of Soil Salinity Based on UAV Images
Soil salinity inversion model was established by taking 4 vegetation indexes of 68 modeling samples as independent variables and soil salinity content as dependent variables, and 30 validation samples were used to verify the model. The results are shown in Table 3. respectively. The inversion model of the RF algorithm has the highest accuracy. Therefore, The RF inversion model was selected to conduct soil salinity inversion for UAV images in the test areas A and B, and the salinity levels were divided according to five levels to obtain the distribution diagram (Figure 4a,b). Figure 4a,b are the interpolation maps of the measured soil salinity in the sampling area. Generally speaking, the inversion results of the wheat field soil salinity are basically the same as the interpolation results of measured data, and the change trend of the area proportions of each grade is roughly the same, but in comparison, the reflection of the inversion results on spatial distribution of soil salinity is more refined. Therefore, the soil salinity inversion model of winter wheat based on RF is better and has the best stable effect.

UAV-Ground Collaborative Inversion Model of Soil Salinity Based on UAV Images
Soil salinity inversion model was established by taking 4 vegetation indexes of 68 modeling samples as independent variables and soil salinity content as dependent variables, and 30 validation samples were used to verify the model. The results are shown in Table 3. It can be seen that all three inversion models show good stability, with their Rm 2 and Rv 2 both exceeding 0.6, and there is no "overfitting phenomenon". The highest modeling Rm 2 of the RF algorithm is 0.878, which is 0.089 and 0.27 higher than the Rm 2 of the BPNN and SVM algorithms, and the RMSEm is the lowest of 0.511, which is 0.159 and 0.38 lower than the RMSEm of the BPNN and SVM algorithms, respectively. From the validation results, the highest Rv 2 of the RF algorithm is 0.827, which is 0.16 and 0.226 higher than the Rv 2 of the BPNN and SVM algorithms, and the RMSEv is the lowest 0.473, which is 0.216 and 0.340 lower than the RMSEv of the BPNN and SVM algorithms, respectively. The inversion model of the RF algorithm has the highest accuracy. Therefore, The RF inversion model was selected to conduct soil salinity inversion for UAV images in the test areas A and B, and the salinity levels were divided according to five levels to obtain the distribution diagram (Figure 4a,b). Figure 4a,b are the interpolation maps of the measured soil salinity in the sampling area. Generally speaking, the inversion results of the wheat field soil salinity are basically the same as the interpolation results of measured data, and the change trend of the area proportions of each grade is roughly the same, but in comparison, the reflection of the inversion results on spatial distribution of soil salinity is more refined. Therefore, the soil salinity inversion model of winter wheat based on RF is better and has the best stable effect.

Collaborative Inversion of Soil Salinity Based on "Satellite-UAV-Ground"
Based on the 75 surface vector data of test areas, the NDVI, NDRE, GNDVI and RDVI of the Sentinel-2A images of 52 modeling samples were taken as independent variables, and the soil salt content estimated by UAV was taken as dependent variable to establish soil salt inversion model, and 23 validation samples were used to verify the model. Figure 5 shows the accuracy of the modeling set and validation set of the three models. By comparison, RF model has the highest accuracy, modeling set R 2 = 0.886, RMSE = 0.456, validation set R 2 = 0.850, RMSE = 0.505, followed by SVM model, the modeling set and validation set R 2 are 0.796 and 0.649 respectively, and BPNN model has the lowest accuracy, the modeling set and validation set R 2 are 0.734 and 0.630 respectively. Therefore, RF model is the optimal model for collaborative inversion of soil salinity based on "Satellite-UAV-Ground". (a)

Collaborative Inversion of Soil Salinity Based on "Satellite-UAV-Ground"
Based on the 75 surface vector data of test areas, the NDVI, NDRE, GNDVI and RDVI of the Sentinel-2A images of 52 modeling samples were taken as independent variables, and the soil salt content estimated by UAV was taken as dependent variable to establish soil salt inversion model, and 23 validation samples were used to verify the model. Figure 5 shows the accuracy of the modeling set and validation set of the three models. By comparison, RF model has the highest accuracy, modeling set R 2 = 0.886, RMSE = 0.456, validation set R 2 = 0.850, RMSE = 0.505, followed by SVM model, the modeling set and validation set R 2 are 0.796 and 0.649 respectively, and BPNN model has the lowest accuracy, the modeling set and validation set R 2 are 0.734 and 0.630 respectively. Therefore, RF model is the optimal model for collaborative inversion of soil salinity based on "Satellite-UAV-Ground".

Collaborative Inversion of Soil Salinity Based on "Satellite-UAV-Ground"
Based on the 75 surface vector data of test areas, the NDVI, NDRE, GNDVI and RDVI of the Sentinel-2A images of 52 modeling samples were taken as independent variables, and the soil salt content estimated by UAV was taken as dependent variable to establish soil salt inversion model, and 23 validation samples were used to verify the model. Figure 5 shows the accuracy of the modeling set and validation set of the three models. By comparison, RF model has the highest accuracy, modeling set R 2 = 0.886, RMSE = 0.456, validation set R 2 = 0.850, RMSE = 0.505, followed by SVM model, the modeling set and validation set R 2 are 0.796 and 0.649 respectively, and BPNN model has the lowest accuracy, the modeling set and validation set R 2 are 0.734 and 0.630 respectively. Therefore, RF model is the optimal model for collaborative inversion of soil salinity based on "Satellite-UAV-Ground".

Analysis of Inversion Results of Soil Salinity in Wheat Area
A large-scale inversion of the soil salinity in the winter wheat planting area in Kenli district was carried out through the optimal RF model of the "Satellite-UAV-Ground" collaborative inversion, and the soil salinity level distribution map was obtained, as shown in Figure 6. Table 4 shows the area ratio of each salinity level. It can be seen that the distribution of non-saline soil in the wheat planting area of the study area is very small, accounting for only 0.05%. The proportion of winter wheat soil area shows a trend of decreasing with the increase of salinization, which is consistent with the actual situation of the survey. Among them, mildly salinized soil in wheat area is widely distributed, accounting for 61.32% of the total area. It is concentrated in the southwest where the terrain is relatively high and the northeast area affected by the fresh water of the Yellow River. The moderately salinized soil area was the second, accounting for 19.53%, which was distributed in all wheat regions. Severe salinized soil and salinized soil accounted for 19.1% of the total area and were scattered in wheat area. 3.3.2. Accuracy Comparison between "Satellite-UAV-Ground" Collaborative and "Satellite-Ground" Direct Inversion Figure 7 shows the soil salinity inversion results of the RF model constructed directly based on Sentinel-2A images and measured salinity data. The R 2 of the model is 0.56, which is far lower than the RF model of collaborative inversion. From the comparison of Figure 6 and Figure 7, it can be seen that the trend of soil salinity distributions obtained by the two inversion methods are basically consistent. However, Figure 7 is generally lower than the salt value in the survey. This is mainly due to the influence of the mixed pixels during the directly satellite remote sensing inversion method. Figure 6 has fewer non-saline areas and more saline soil areas, which is consistent with the field survey. As a result, using the UAV inversion result as the medium data can effectively reduce the influence of mixed pixels on the inversion result.

Analysis of Inversion Results of Soil Salinity in Wheat Area
A large-scale inversion of the soil salinity in the winter wheat planting area in Kenli district was carried out through the optimal RF model of the "Satellite-UAV-Ground" collaborative inversion, and the soil salinity level distribution map was obtained, as shown in Figure 6. Table 4 shows the area ratio of each salinity level. It can be seen that the distribution of non-saline soil in the wheat planting area of the study area is very small, accounting for only 0.05%. The proportion of winter wheat soil area shows a trend of decreasing with the increase of salinization, which is consistent with the actual situation of the survey. Among them, mildly salinized soil in wheat area is widely distributed, accounting for 61.32% of the total area. It is concentrated in the southwest where the terrain is relatively high and the northeast area affected by the fresh water of the Yellow River. The moderately salinized soil area was the second, accounting for 19.53%, which was distributed in all wheat regions. Severe salinized soil and salinized soil accounted for 19.1% of the total area and were scattered in wheat area. 3.3.2. Accuracy Comparison between "Satellite-UAV-Ground" Collaborative and "Satellite-Ground" Direct Inversion Figure 7 shows the soil salinity inversion results of the RF model constructed directly based on Sentinel-2A images and measured salinity data. The R 2 of the model is 0.56, which is far lower than the RF model of collaborative inversion. From the comparison of Figures 6 and 7, it can be seen that the trend of soil salinity distributions obtained by the two inversion methods are basically consistent. However, Figure 7 is generally lower than the salt value in the survey. This is mainly due to the influence of the mixed pixels during the directly satellite remote sensing inversion method. Figure 6 has fewer non-saline areas and more saline soil areas, which is consistent with the field survey. As a result, using the UAV inversion result as the medium data can effectively reduce the influence of mixed pixels on the inversion result.  In order to further compare the inversion results of the "Satellite-UAV-Ground" collaborative method and the direct satellite remote sensing method, the salt values of 77 sampling points in Kenli District at the two inversion results were extracted and compared with the measured data. As is  In order to further compare the inversion results of the "Satellite-UAV-Ground" collaborative method and the direct satellite remote sensing method, the salt values of 77 sampling points in Kenli District at the two inversion results were extracted and compared with the measured data. As is In order to further compare the inversion results of the "Satellite-UAV-Ground" collaborative method and the direct satellite remote sensing method, the salt values of 77 sampling points in Kenli District at the two inversion results were extracted and compared with the measured data. As is shown in Figure 8. The R 2 = 0.741 and RMSE = 0.776 of the "Satellite-UAV-Ground" collaborative inversion results and the measured values, while the R 2 = 0.591 and RMSE = 0.831 of the satellite remote sensing method and the measured values, indicating that the results of the "Satellite-UAV-Ground" collaborative inversion are highly consistent with the measured salinity, and the direct satellite remote sensing inversion results have a large deviation.
shown in Figure 8. The R 2 = 0.741 and RMSE = 0.776 of the "Satellite-UAV-Ground" collaborative inversion results and the measured values, while the R 2 = 0.591 and RMSE = 0.831 of the satellite remote sensing method and the measured values, indicating that the results of the "Satellite-UAV-Ground" collaborative inversion are highly consistent with the measured salinity, and the direct satellite remote sensing inversion results have a large deviation. Therefore, "Satellite-UAV-Ground" collaborative inversion effectively improves the accuracy of soil salt inversion in winter wheat planting areas and obtains a more reliable soil salt distribution.

Discussion
The measured data on the ground is the basis for quantitative analysis of soil, UAV near-earth remote sensing is the link between the satellite and the ground, and satellite remote sensing is the platform for large regional inversion. Combining the three will be an important way to obtain soil salt information at present and in the future. Therefore, in this study, the three platforms of "Satellite-UAV-Ground" data were used to perform soil salinity inversion. Compared with the direct satellite image inversion results, the inversion quality was significantly improved. The inversion results of two models are verified with ground-based data. The results show that the R 2 based on the "Satellite-UAV-Ground" collaborative inversion model has been significantly improved, and RMSE and distribution pattern is also improved, but not so obviously. It might be because the salinity inversion result of UAV image is used as an intermediate bridge, its scope is limited and it is hard to fully express the situation of the entire study area. We will continue to improve research on collaborative approach, models, vegetation spectra, data sampling, etc. to improve the performance in terms of scale, accuracy, and temporal and spatial resolution.
The selected test areas in this study were located in two concentrated winter wheat planting areas in the southwest and northeast of Kenli District. The growth of winter wheat was significantly different, and the soil salt content was distributed in all grades, which was typical and representative. In order to improve the spectral and salinity data accuracy of winter wheat, the grid intersection method was used to determine the location of ground data collection, which could ensure the uniformity and accuracy of sample points more than the traditional random sampling point determination method. In order to ensure the representativeness of modeling samples, the samples are sorted and sampled at equal intervals at the ratio of 2:1 between the calibration set and the validation set. Compared with the traditional method of random partition modeling set and validation set [5,49], the data is more uniform and effective, thus ensuring the strong universality and high stability of the model.
The soil salt contents in the study area vary greatly, so the smaller the pixel and the closer the sampling point, the more accurate the information reflected. In this study, an inversion model was built based on the measured ground data and UAV images. The salinity value obtained from the inversion was regarded as the "salinity truth value", one pixel of Sentinel-2A corresponds to 40,000 UAV pixels, the average value of the soil salinity corresponding to these 40,000 UAV pixels was Therefore, "Satellite-UAV-Ground" collaborative inversion effectively improves the accuracy of soil salt inversion in winter wheat planting areas and obtains a more reliable soil salt distribution.

Discussion
The measured data on the ground is the basis for quantitative analysis of soil, UAV near-earth remote sensing is the link between the satellite and the ground, and satellite remote sensing is the platform for large regional inversion. Combining the three will be an important way to obtain soil salt information at present and in the future. Therefore, in this study, the three platforms of "Satellite-UAV-Ground" data were used to perform soil salinity inversion. Compared with the direct satellite image inversion results, the inversion quality was significantly improved. The inversion results of two models are verified with ground-based data. The results show that the R 2 based on the "Satellite-UAV-Ground" collaborative inversion model has been significantly improved, and RMSE and distribution pattern is also improved, but not so obviously. It might be because the salinity inversion result of UAV image is used as an intermediate bridge, its scope is limited and it is hard to fully express the situation of the entire study area. We will continue to improve research on collaborative approach, models, vegetation spectra, data sampling, etc. to improve the performance in terms of scale, accuracy, and temporal and spatial resolution.
The selected test areas in this study were located in two concentrated winter wheat planting areas in the southwest and northeast of Kenli District. The growth of winter wheat was significantly different, and the soil salt content was distributed in all grades, which was typical and representative. In order to improve the spectral and salinity data accuracy of winter wheat, the grid intersection method was used to determine the location of ground data collection, which could ensure the uniformity and accuracy of sample points more than the traditional random sampling point determination method. In order to ensure the representativeness of modeling samples, the samples are sorted and sampled at equal intervals at the ratio of 2:1 between the calibration set and the validation set. Compared with the traditional method of random partition modeling set and validation set [5,49], the data is more uniform and effective, thus ensuring the strong universality and high stability of the model.
The soil salt contents in the study area vary greatly, so the smaller the pixel and the closer the sampling point, the more accurate the information reflected. In this study, an inversion model was built based on the measured ground data and UAV images. The salinity value obtained from the inversion was regarded as the "salinity truth value", one pixel of Sentinel-2A corresponds to 40,000 UAV pixels, the average value of the soil salinity corresponding to these 40,000 UAV pixels was calculated, which was used as the salt value corresponding to the pixel of Sentinel-2A. Compared with the salt value measured on the ground, the "true salt value" is a more comprehensive representation of the true salt status in the Sentinel-2A pixel. Most of the previous studies have realized the up-scaling inversion method of UAV model by establishing the relationship of spectral information between UAV image and satellite image [5,50]. The method in this paper is a collaborative inversion model based on satellite images which is constructed by taking the UAV "salt truth value" as the bridge of "Satellite-UAV-Ground" collaborative inversion. It can better maintain the information content and spatial structure characteristics of the original remote sensing data, and better guarantee the accuracy of salt inversion. It is an effective way to realize the integrated inversion of "Satellite-UAV-Ground".
Previous research results show that the machine learning inversion model is superior to the statistical model [51][52][53][54]. The machine learning method can show strong nonlinear fitting ability and excellent data mining ability and can better simulate the complex nonlinear relationship between soil salinity and remote sensing image characteristics. Therefore, three machine learning modeling methods were directly adopted in this study, and the comparison found that the RF model was particularly effective, especially the "Satellite-UAV-Ground" collaborative inversion model, with R 2 and RMSE up to 0.886 and 0.456, respectively. Therefore, if there is a non-linear relationship between the predictor variable and the response variable, a non-linear model such as RF will usually have a better fitting effect and produce excellent estimation accuracy.
The growth of winter wheat in coastal salinized areas is mainly affected by soil salinity, while other factors such as soil texture, pH value, climate, water, nutrients, and fertility have a more balanced impact on crop growth, which is a systematic error. Therefore, salt content is mainly considered for the impact on vegetation index. The vegetation index can be used to invert soil salinity indirectly, which has been confirmed by previous research [11,55,56]. However, the relationship between the vegetation index and soil salinity is different under different environmental conditions. Consequently, the constructed model is only applicable to wheat fields in coastal saline soil in spring. In order to make the obtained soil salinity better support crop production, the next step is to perform salinity inversion for different crops in the study area in different seasons, and use satellite data at different times to verify the model to further improve the reliability of the model.
When salinity inversion is carried out by using data from different remote sensing platforms, matching between data is very important. The predecessors usually take the average of multiple measurements within the ground range corresponding to the satellite pixel as the salt value corresponding to the pixel [57,58]. In this paper, the salt value of the corresponding range of Sentinel-2A pixels is obtained by calculating the average value of 40,000 pixels of the corresponding UAV. Compared with the previous methods, the accuracy of salt value corresponding to satellite pixel is improved, thus the spatio-temporal matching accuracy of the data is improved. Due to the uncertainty of remote sensing data, the band response functions of different sensors are different. When building a collaborative inversion model, it is necessary to fully consider the radiation characteristics of the data and select similar bands to reduce the impact caused by radiation and improve the accuracy of data spectrum matching.

Conclusions
In this study, satellite images, UAV images and measured soil salt data were combined to build an inversion model to realize the "Satellite-UAV-Ground" collaborative inversion of soil salt in the coastal area of winter wheat. The main conclusions are as follows: (1) The correlation between 8 vegetation indexes based on the multi-spectral bands of UAV images in the winter wheat test area and soil salinity was all greater than 0.5. According to the correlation coefficient and variance expansion factor, four sensitive vegetation indices, including NDVI, RDVI, GNDVI, and NDRE, were selected for modeling, and three machine learning inversion models, BPNN, SVM and RF were constructed. The RF model modeling set has R 2 = 0.878, RMSE = 0.511, and its accuracy is higher than BPNN and SVM. It is the best salt estimation model. The inversion results are in good agreement with the actual distribution of soil salt in the test area. The model has good predictive ability and applicability for the estimation of soil salinity of winter wheat in spring in coastal salinization areas. (2) The soil salinity in the test area obtained from the inversion of the best model of UAV imagery is used as the "true value of salinity" for satellite image modeling, and three collaborative inversion models are constructed. The RF inversion model has R 2 = 0.885, which is significantly better than the BPNN and SVM models. The model is applied to the study area to obtain a large-scale distribution map of soil salinity in the winter wheat area. The soil in the winter wheat planting area of the study area is dominated by light and moderate salinization, accounting for 80.85% of the area, mainly distributed in the southwest and northeast regions. The area of heavily salinized and saline soil is smaller, only accounting for 19.1%, and is scattered in the wheat area. (3) The result of "Satellite-UAV-Ground" collaborative inversion method and the satellite remote sensing direct inversion method were quantitatively compared and evaluated by using the measured salinity data of 77 sample points in the wheat field in the study area. The results show that the R 2 = 0.741 and RMSE = 0.776 of the "Satellite-UAV-Ground" collaborative inversion result and the measured value, while the R 2 = 0.591 and RMSE = 0.831 of the satellite remote sensing method and the measured value. Therefore, the "Satellite-UAV-Ground" collaborative inversion method can effectively improve the accuracy of the soil salt inversion results and obtain more accurate spatial distribution of winter wheat soil salt in spring in coastal salinization areas.
This study makes full use of the advantages of satellites, UAV images, and ground data to construct soil salinity inversion model, using UAV soil salinity inversion results as intermediate data, and machine learning modeling methods to obtain a grade distribution map of winter wheat soil salinity in the study area which is consistent with the actual distribution of soil salt. The study proposed an effective "Satellite-UAV-Ground" collaborative inversion method for soil salinity, which provides a scientific basis for accurately grasping the distribution of winter wheat soil salinity levels and guiding agricultural production in the study area.