Wujiaqu City is located in the urban agglomeration of the northern slope of Tianshan Mountain, the northern foothills of Tianshan Mountain and the southern edge of the Junggar Basin (43°59′25″ to 44°39′00″N and 87°17′42″ to 87°43′15″E). As warm and humid air masses are difficult to enter the Xinjiang Basin due to the obstruction of mountain ranges, resulting in scarce precipitation and large temperature variations throughout the year (-38~42°C); scarce precipitation (150mm) and strong evaporation (2000mm). In order to better explore the ability and ubiquity of UAV hyperspectral data to indirectly monitor soil salinization through vegetation spectral information in oasis farmland in arid zones, a field study was conducted in June 2022 in a typical cotton field in Wujiaqu City ( Figure 1 ).

The field sampling time was selected on June 12, 2022, which was in the cotton bud stage, with no interference from precipitation, irrigation, or anthropogenic factors, and the unmanned data hyperspectral data were sampled at the same time as the soil samples to ensure the reliability of the measured data. Thirty-three surface (0-10 cm) soil samples were collected using the five-point sampling method, and the coordinates of each sampling point were recorded using a portable GPS device. Soil samples were air-dried and ground in the laboratory, and sieved through a 2-mm sieve. Soil solutions were prepared according to the method of 1:5 soil leachate, and soil parameters such as solution EC_1:5 and pH were measured using a multiparameter meter (WTWinoLab^® Multi3420 set B, WTW GmbH, Germany) at a constant room temperature of 25°C. Soil EC_1:5 can be used to indicate soil salinity.

In order to acquire the UAV hyperspectral data in the study area, a Nano-Hyperspec (Headwall Photonics Inc., Bolton, MA, USA) ultra-miniature airborne hyperspectral imager (UHI) was mounted on a M600 Pro six-rotor UAV flight platform for spectral data acquisition. for spectral data acquisition. The parameters of the hyperspectral imager are shown in Table 1 . Hyperspectral images with a spatial resolution of 4.4 centimeters and a spectral range of 400~1000 nm can be acquired at a flight altitude of 100 meters. The hyperspectral image data were acquired on June 12, 2022 at 13:00 BST in a clear, cloudless environment with a wide field of view, and a calibrated white cloth with an area of 2.5m×2.5m and a reflectance of 60% was used for reflectance correction and dark current correction. After data acquisition, the hyperspectral images were processed using Hyperspec III (version 1.3) and SpectralView (version 1.0) software for radiance-to-brightness conversion, atmospheric correction, and geometric correction.

Savitzky-Golay (SG) smoothing is the most efficient way to process spectral reflectance data of saline soils (Ge et al., 2021). The hyperspectral images acquired in this study had a resolution of 4.4 cm and covered 272 bands, and in order to ensure that this dense spectral information can objectively and accurately represent the spectral features of cotton that are subjected to abnormal changes due to salinity, the hyperspectral images were first subjected to Savitzky-Golay (S-G) smoothing with eight points and a polynomial order of nine. Standard normalized variate (SNV) is often used to remove spectral signal variance and is widely used in spectral analysis (Wang H.-P. et al., 2022). In order to correct the spectral errors between samples due to scattering, the hyperspectral images were corrected individually for each spectrum using the SNV. These preprocessing procedures were implemented using the MATLAB R2021b software was implemented.

Hyperspectral data contain information from different wavelengths, which may contain noise (Ge et al., 2022). Fractional order derivative (FOD) can be used to smooth and remove this noise, thus improving the quality of the data (Wang J. et al., 2022), and can also be used to enhance features in hyperspectral data (Jiang et al., 2022). Since the G-L definition is relatively concise and has a better performance in terms of for transform and spectral processing (Hong et al., 2019), In this study, the G-L definition form is chosen to process the hyperspectral data, and the G-L definition expression is shown in Equation (1).

Where α is the order, Γ is the Gamma function, λ is the wavelength, and n is the D-value of the constraint limit of the differential equation.

Descriptive statistics of 33 soil samples, as shown in Figure 3 , showed that the EC_1:5 of the samples ranged from 2.26 dS m^-1 to 20.09 dS m^-1, with a mean of 5.19 dS m^-1, a standard deviation of 4.05 dS m^-1, and a coefficient of variation (CV) of 78.06% (10%<CV<90%, which is a moderate variation), which suggests that the data have a sufficiently discrete degree for model construction and prediction.

Comparing the original spectrum ( Figure 4A ), the SG-filtered spectrum ( Figure 4B ) and the SNV-treated spectrum ( Figure 4C ), the influential noise is further eliminated after SG smoothing, and the spectral reflectance range is expanded to -2 to 2.5, which well highlights the spectral characteristics of cotton. Soil salinity mainly images cotton’s absorption of water and nitrogen, which in turn affects chlorophyll synthesis, and the degree of salinity leads to differences in the degree of reflection and absorption of electromagnetic waves in the bands at 555 nm and 680-690 nm affected by chlorophyll, and also affects the reflectance intensity of the red-edge bands, as well as the degree to which the electromagnetic waves at 763 nm, 821 nm, and 935 nm are absorbed by water.

The soil reflectance spectra were processed by FOD with the order of 0-2 and the step size of 0.1, and the processed results are shown in Figure 5 . With the increase of FOD order, the values in the whole spectral reflectance range gradually decrease and converge to 0. The small differences between spectra are highlighted to avoid the loss of important information and enhance the effect of data preprocessing. In the 0.1-1 order spectra, positive peaks appear at 506 nm, 606 nm and 770 nm, and negative peaks appear at 761 nm and 932 nm; in the 1-2.0 order spectra, positive peaks appear at 501 nm, 766 nm and 939 nm, and negative peaks appear at 759 nm and 930 nm.

The soil samples were divided into training set and test set input models according to 7:3, and the SOA-RF performance results of the three modeling strategies are shown in Table 4 and Figure 12 , which shows that the R² and RPD of the three modeling strategies gradually increase, the RMSE and MSE gradually decrease, and there is a significant improvement in the modeling accuracy of modeling strategy III. At the same time, the MSE of SOA-RF and traditional RF models are compared, and the MSE of the three modeling strategies of SOA-RF are much smaller than those of RF models, which shows that the introduction of SOA algorithms improves the model performance greatly.

The spatial distribution of top soil salinity in cotton fields was mapped using the SOA-RF model ( Figure 13 ). From the descriptive statistical analysis conducted by EC_1:5, the measured EC_1:5 of the soil ranged from 2.26 dS m^-1 to 20.09 dS m^-1, and the optimal model predicted soil salinity ranging from 1.15 dS m^-1 to 20.34 dS m^-1, and the spatial distribution of soil salinity was in accordance with the results of the field survey. Sub-membrane drip irrigation technology provides a breakthrough for the sustainable development of agro-ecology in this region. Sub-membrane drip irrigation technology can achieve the purpose of water-saving irrigation, but the small amount of irrigation water is difficult to adequately wash the salts in the soil, and the strong evapotranspiration in this region has been continuous, which makes the salts in the soil will always be migrated to the surface of the soil, so that the salts in the surface layer of the soil accumulates year by year, especially when the soil is not covered with a membrane it is very obvious to see that there are soil salts higher than the plant salinity, so that the salts in the surface layer of the soil are higher than the plant salts. The soil salinity is higher than that of the vegetated areas, especially in the case of non-mulched areas.

Preprocessing of hyperspectral data is the most important step before building quantitative prediction models (Minu et al., 2016). In this study, the hyperspectral data were first subjected to S-G smoothing and SNV processing, which clearly represented the spectral characteristics of cotton affected by salinity with abnormal changes at the level of spectral reflectance, and removed and amplified the spectral characteristics by removing the spectral signal variance.

However, hyperspectral data are complex and contain a large amount of intricate information and noise between consecutive and short spaced bands, which requires further fractional order differentiation processing of hyperspectral data (Sun et al., 2021). The results show that FOD is more helpful to reduce the sharp peak aberrations than integer-order differencing, which maintains the structure of the original spectral curves and preserves the original information, but also makes the baseline vary within a moderate curvature range, which makes more useful information of the spectral data to be highlighted (Hong et al., 2020). In this study, we differentiated the hyperspectral data by 20 orders at 0.1 intervals and found that the FOD spectra could reveal hidden information related to the vegetation spectral information. A correlation heat map of the 0 to 2.0 order spectra with soil EC_1:5 was plotted ( Figure 14 ), with wavelength on the horizontal axis, order on the vertical axis, and the color representing the magnitude of the correlation coefficient. The correlation heat map shows a nearly symmetrical pattern, when the FOD order is between 0.1 and 1, the reflectance in the wavelength range of 400-700 nm is basically positively correlated with soil EC_1:5, and the reflectance in the wavelength range of 700-1000 nm is basically negatively correlated with soil EC_1:5; when the FOD order is between 1.0 and 2.0, the soil reflectance in the range of 400-700 nm shows a weak correlation with soil EC_1:5. When the FOD order is between 1.0 and 2.0, the reflectance of soil in the range of 400-700nm shows a weak positive or negative correlation; the reflectance of soil in the range of 700-1000nm basically shows a positive correlation. With the increase of FOD order, the correlation of reflectance at the blue light wavelength is gradually weakened, the negative correlation of reflectance at the red light and near-infrared wavelengths is changed to a positive correlation with the division of 1.0 order, but the overall correlation is strong, and the weak correlation of reflectance at the green light wavelengths tends to be stabilized. The core purpose of FOD processing is to bring out the useful information from the complex and noisy hyperspectral data, and the subsequent processing of the FOD processed data will help to improve the accuracy of the model, but for different research areas and different data, the FOD processing may present different effects and even results, and the method needs to be further investigated and improved in a variety of different research areas and using different data sets.

Extracting appropriate feature variables from hyperspectral data containing a large amount of information is a key step in constructing a stable model (Lao et al., 2021), and the three modeling strategies can be compared to derive the degree of sensitivity of the three feature indicators.

Modeling strategy I used only the single-band feature variables screened by CARS to participate in the modeling, and the results showed that the 1.9 order was the optimal differential order which was the same as the optimal differential order for region I in the study of salinity content estimation in different saline zones by (Fu C. et al., 2021). The characteristic bands were scattered in the visible (408-777 nm), but mainly concentrated in the near-infrared (NIR) shortwave (800-1000 nm), which indicated that the NIR bands performed better in the characterization of soil salinity using vegetation spectra, which was also confirmed in the results of (Tian et al., 2021) and (Nguyen et al., 2020). However, since the single-band variables can only reflect the spectral properties from a one-dimensional perspective, the R² of modeling strategy I is only 0.9107, with the lowest coefficient of determination among the three modeling strategies, and the RPD is 1.8736, which is also the weakest in terms of stability.

Modeling strategy II adds 25 texture features to the single-band variables, and the results show that modeling strategy II improves R² by 0.0129 and RPD by 0.8118.The overall improvement of R² is not significant, which may be due to the fact that texture features are also statistics of single-band information (Zhang et al., 2018), and although texture features increase the spatial distribution describing spatial dimensionality information, it lacks the characterization of interconnected information between bands (Fu Y. et al., 2021). The RPD enhancement is more obvious compared to R², and the stability of the model is significantly improved by adding texture features. Overall, the combination of single-band indices with image texture parameters can increase the dimensionality of the information, which can more accurately characterize yield variations, which is consistent with the findings of (Wang et al., 2021) in the performance of spectral indices alone and in combination with spectral indices based on texture measurements of UAV-based hyperspectral images in seed yield estimation. However, the inclusion of texture features provided less overall improvement to the prediction model.

Modeling strategy three adds six three-band indices to strategy two, and the results show an improvement of 0.0369 in R² and 1.6478 in RPD compared to strategy one, which is a very significant overall improvement ( Figure 15 ). The three-band index improves the model performance more than the single-band index, which is consistent with the conclusion that the three-band index constructed by OBCA in the study of (Zhu et al., 2022) improves the relationship between spectra and soil salinity, which is of high importance in modeling, and this study is the first time that texture features are added into the modeling strategy to compare with the three-band index, and it is seen that the three-band index is also more sensitive than the texture features are more sensitive. The correlations between the six tri-band indices and soil EC_1:5 were similar in magnitude, with TBI4 showing the best performance with an R of 0.912, and it can also be found that all six tri-band index band combinations contain bands within 777~1000 nm, which shows that for the tri-band indices, the near-infrared band is still the most sensitive band to soil EC_1:5, which is also consistent with the findings of (Wang S. et al., 2019) study in which the selected characteristic bands using the SPA method showed that different spectral transformation forms characterize soil salinity differences, mainly in the NIR band in agreement.

The screening and construction method of feature indices used in this study can extract effective spectral variables from highly redundant hyperspectral data while ignoring most irrelevant variables, thus substantially improving the accuracy of the model. However, there may be different results for different crop types and regions, and further research on the generalizability of the method is needed.

There are many studies that use machine learning methods for soil salinity prediction (Zhu et al., 2022), utilized RF model R² to reach 0.815 in a study of soil salinity inversion of Abbey Lake Oasis based on UAV hyperspectral imagery; (Jia et al., 2022) used an RF model with an R² of 0.93 in a study of salinity inversion for different cultivated soil types based on hyperspectral data and machine learning; (Hu et al., 2019) reached an R² of 0.94 in quantitative soil salinity estimation using RF model predictions based on UAV hyperspectral data; (Cui et al., 2023) used unmanned aerial multispectral remote sensing and three machine learning algorithms to invert the soil salinity of farmland at different depths under crop cover, and the optimal prediction model had an R² of 0.775. After comparison, the accuracy of the SOA-RF model used in this study exceeds that of the previous models used (Zhao et al., 2022). used SOA algorithm to optimize the RF model in the study of vulnerability of agricultural soil and water resources system, and the R² reached 0.9999; (Li et al., 2023) in the prediction of concrete compressive strength showed that the SOA-RF model outperformed ANN, ELM and empirical models; (Zhou et al., 2023) used the SOA-RF model in a study of maximum surface settlement prediction with an R² of 0.9372. In this study, the SOA-RF model was innovatively used for the inversion of soil EC_1:5, and the parameters of the RF model were optimized using the SOA algorithm, and there was a great improvement in the accuracy. A comparison of the outputs of the traditional RF and the SOA-RF model is shown in Figure 16 , and although the traditional RF showed better results in the training set, the performance in the test set was very poor, and the suppression of the overfitting problem can be seen in the SOA-RF.

To cope with the complex scenarios of vegetation and film cover during the cotton growing season, a method of quantitatively extracting soil salinity from cotton fields covered with film has been developed by combining spectral processing, band selection, spectral indices, texture features and machine learning algorithms, with the advantages of speed, simplicity, and non-destructive nature, with the main theme of “Salinity Information Extraction in Cotton Fields”, using hyperspectral remote sensing by unmanned aerial vehicle (UAV) as a means to enhance soil salinity monitoring in cotton fields, and providing a useful reference to optimize the irrigation technology in cotton fields.

Remote sensing by drones offers the possibility of obtaining crop growth information on a large scale and continuously. However, the area of remote sensing by drones is still limited, and future work needs to be carried out in conjunction with satellite remote sensing. And in the subsequent research, based on the UAV hyperspectral data, applying the developed optimal prediction model to map the soil salinity distribution in different fertility periods of cotton, we can try to explore the distribution law of soil salinity in different fertility periods of cotton, grasp the health status of cotton in different periods, put forward countermeasures to deal with excessive accumulation of soil salinity, and safeguard the cotton yield from the impact of soil salinity.