Hyperspectral Models and Forcasting of Physico-Chemical Properties for Salinized Soils in Northwest China

XIAO Zhen-zhen, LI Yi,2*, FENG Hao

1.College of Water Resources and Architecture Engineering, Northwest A&F University, Yangling 712100, China 2.Institute of Water-Saving Agriculture in Arid Areas of China, Northwest A&F University, Yangling 712100, China 3.National Engineering Research Center for Water Saving Irrigation at Yangling,Northwest A&F University, Yangling 712100, China

Hyperspectral Models and Forcasting of Physico-Chemical Properties for Salinized Soils in Northwest China

XIAO Zhen-zhen1, LI Yi1,2*, FENG Hao2,3

1.College of Water Resources and Architecture Engineering, Northwest A&F University, Yangling 712100, China 2.Institute of Water-Saving Agriculture in Arid Areas of China, Northwest A&F University, Yangling 712100, China 3.National Engineering Research Center for Water Saving Irrigation at Yangling,Northwest A&F University, Yangling 712100, China

Hyperspectral remote sensing data have special advantages, i.e., they have high spectral resolution and strong band continuity, and a great number of spectral information could be widely used in soil properties monitoring research.Using hyperspectral remote sensing technique to analyze saline soil properties makes great significance for the crop growth in the irrigation district and agricultural sustainable development.221 soil samples were collected from Manasi River Basin to measure soil electrical conductivity (EC), soil organic matter (SOM) and 3 kinds of cation concentrations including Na+, Ca2+and Mg2+, which were used to obtain sodium adsorption ration value (SAR).The soil hyperspectral curves were also measured.EC, SOM and SAR models were established based on the six spectral-related indices, including raw reflectance (R), standard normal variable (SNV), normalized difference vegetation index (NDVI), logarithm of the reciprocal (LR), the first derivative reflectance (FDR) and continuum-removal reflectance (CR) by the stepwise linear regression method.The results showed that, compared to the other five models, the model of log (EC)～Rhad the highest accuracy withrvalue of 0.782 and RMSE value of 0.256.The model of SOM vs.NDVI had the highest accuracy withrvalue of 0.670 and RMSE value of 5.352.The model of SAR vs.FDR had the highest accuracy withrvalue of 0.647 and RMSE value of 1.932.As to the model accuracy of the studied soil physico-chemical properties, the log(Ec) model was the most effective one, followed by the SOM model, the SAR model was the most inaccurate.The sensitive wavelengths for EC, SOM and SAR distributed in 395～1 801 nm, 352～1 144 nm and 394～1 011 nm, respectively.Since soil physico-chemical properties were highly spatially variable, there were large differences for the model establishment and validation of the soil properties.This research could be a reference of hyperspectral remote sensing monitoring of salinized soils.

Soil electrical conductivity; Soil organic matter; Sodium adsorption ration; Hyperspectral model

Introduction

Soil salinization is a typical phenomenon of land degradation.It reduces soil permeability, constrains crop growth and production and restricts agricultural sustainable development[1].In Xinjiang Uygur Autonomous Region, northwest of China, the area of arable saline soil land account for 31.1% of the total arable land[2].Among the related indices to denoting soil salinization, soil electrical conductivity (EC) is an important soil quality assessment parameter[3-4].Besides, sodium adsorption ratio (SAR) reflects differences in salt ions and is an important parameter to characterize the degree of soil alkalization.On the other hand, soil organic matter (SOM) is not only an important material basis for soil fertility, but also an important part of the global carbon pool, it is a main source of plant nutrients.Although SOM occupies only a small part of total soils, it exerts significant influence on improving soil structure and texture, contributes much to harmonizing soil physical properties, such as soil water, temperature, nutrient and air[5].However, traditional methods of measuring soil physic-chemical properties are slowly in detection with pollution and poor real-time capability, and the sampling number is limited by human and material resources and other factors.Therefore, they cannot be applied to a large area of real-time dynamic monitoring and do not qualify the development of precision agriculture.Hyperspectral data are fast, time-saving with no pollution, and not limited by space-time and topography, so they have been applied widely to the monitoring of soil properties[6].

So far, many scholars have established a quantitative relationship between soil spectral reflectance and soil properties, such as soil moisture, pH, salinity (or EC) and SOM.The results showed that the established models between soil properties and spectral reflectance-related indices can be applied to estimate soil properties well[7-11].For the Manasi River basin, Liu et al.investigated Beiwucha town in north Manasi county, and they found that the models established for soil salinity vs.continuum-removal index and for SOM vs.reciprocal index were the best, and spectrum bands of red light, purple light and near-infrared light played very important roles in the prediction of soil salinity and SOM[12].

Although there were many researches on the spectral characteristics of soil physical and chemical properties, the criteria and parameters of soil spectrometry were not uniform, thus there were differences between spectral data even in the same study area.In addition, soil spectral reflectance was comprehensively affected by soil physical and chemical properties.Former research that studied the characteristics of hyperspectral characteristics of salinized soils had its regional limitations.The established model may have good ability to predict soil properties in a certain region, but its accuracy would be greatly reduced for the other regions, even weak modeling accuracy could be obtained when sampling in the same area but at different dates or when different number of samples were taken.Hyperspectral models for SAR are still rare.Manasi River basin is the largest artificial oasis in Xinjiang and the fourth largest irrigated area in China.In this paper, saline-alkaline soils in Manasi River basin were taken as the research subject.Based on the measured spectral reflectance data and its multi-shapes of transformations, the models of log(Ec), SOM and SAR with the obtained spectral indices will be established using stepwise linear regression (SLR) method, and the best spectral models for predicting the studied soil physico-chemical properties will be selected.This results will provide references for estimating physico-chemical properties of salinized soils.

1 Materials and Methods

1.1 Study area

Manasi River basin is located within 43°27′N—45°21′N and 85°01′E—86°32′E with a total area of about 3.14×104km2.It is located in the southwest of Junggar Basin in Xinjiang.Its western part is the edge of the Junggar Basin, and its south edge is the Tianshan Mountain.In topography, the south of is higher than the north part, and the area of plain take half of the total area, which is similar to the area of mountain.With the annual precipitation here is 110～200 mm, annual evaporation is 1 500～2 100 mm, and average annual temperature is 5.2 ℃[13].

1.2 Sampling collection and laboratory analysis

A total 221 soils were randomly collected with general interval of 2～3 km in seven irrigation areas including Mosuowan, Manasi county, Anjihai, Shihezi, 121 Regiment, 136 Regiment and 132 Regiment.Samples were air dried after the debris like gravels and grass roots were moved.And then they were ground to pass a 2mm-in-diameter sieve for the later use.

Soil water slurry was obtained by mixing 10 g soil and 50 mL distilled water.Samples were shaken for 5 min and settled for 24 h.EC was measured by a DDB-303A EC-meter.SOM was determined by the potassium dichromate heating method[18].Concentration of Na+, Ca2+and Mg2+were all measured by The Atomic Absorption Spectrophotometric method.SAR was calculated by the following equation[14]

(1)

Where SAR denotes the sodium adsorption ratio (mmol·L-1)1/2, [Na+], [Ca2+] and [Mg2+] denotes concentration of Na+, Ca2+and Mg2+(mmol·L-1), respectively.

1.3 Measurement of spectral reflectance and pretreatments

Soil spectral reflectance was measured for each soil sample by using an analytical spectral device (ASD FieldSpec FR Spectroradiometer) at wavelength scopes from 350 to 1 830 nm with a spectral sampling intervals of 1.4 and 2 nm in the wavelength ranges of 350～1 000 and 1 000～1 830 nm, respectively.The measurement was conducted outdoor at 10:00—14:00 in sunny cloudless days.The field of view angle of the transducer probe was 30°, and held vertically to the soil sample with a height of 5 cm to the soil surface.The soil samples were placed in the aluminum containers with 6 cm in diameter and 4 cm in depth.The dark current effects were removed and the white panel calibration was done before the measurements.After measuring 30 soil samples, the white panel calibration was conducted again.Each sample was scanned 10 times to obtain an actual spectral curve after the data were averaged arithmetically.

Spectral curves near 1 400 and 1 900 nm were affected by water absorption valley, which resulted to a big fluctuation.Thus these two water vapor absorption bands were excluded, and a parabolic splice was used to correct the gaps in 1 000 nm between detectors via ViewSpec Pro Version 6.0 software.

Spectral data pretreatment is essential to spectral analysis, which directly affects the prediction accuracy and stability of the established model.Thus, in addition to the original spectral reflectance index(R), standard normal variable (SNV) was obtained by standard normal transformation, spectral index normalized difference vegetation index (NDVI) was calculated by borrowing a mathematical algorithm from remote sensing data processing, index of continuum-removal (CR) was obtained by using the module Spectral in ENVI 4.7 software, in addition, logarithm of the reciprocal (LR) and the first derivative reflectance (FDR) were also calculated.Total six spectral indices including R were obtained to be used for the later model establishment.

1.4 Model establishment and performance assessment

2 Results and discussion

2.1 Descriptive statistical analysis of soil physico-chemical properties

Statistical analysis of physico-chemical properties of the soil samples in the total, the modeling subset and the validation subset are presented in Table 1.CV denotes coefficient of variation.

Table 1 shows that the soil physico-chemical properties ranged with a similar manner in the total, the modeling subset and the validation subset.As toCvvalues,Cvof EC for the modeling subset was largest and reached 1.1,Cvof SAR for both the modeling and validation subsets followed it and were both 1.0.Cvof SOM in the modeling subset was the smallest, i.e.0.6.In general, for all of the three soil properties-EC, SOM and SAR, the statistical parameters of the modeling subset were close to those of the validate subset, so the divided subsets can be used to establish spectral models.

Table 1 Statistics of soil physico-chemical properties in the total, the modeling subset and the validation subset

2.2 Analysis of soil spectral reflectance

Fig.re 1 demonstrates 10 typical soil spectral reflectance curves, of which the water vapor absorption band was removed.

In Figure 1, the spectral reflectance curves of the soil samples were basically similar as the wavelength changed, and there was no significant difference in their shapes.In the visible light band (350～760 nm), the reflectance increased significantly with a quick increase in wavelength especially within 350～577 nm.The highest soil sample reflectance curve corresponded to EC of 44.8 μs·cm-1, SOM of 5.06 g·kg-1and SAR of 0.83 (mmol·L-1)1/2, respectively.Followed by the curve with 62.5 μs·cm-1of EC, 11.08 g·kg-1of SOM, and 6.65 (mmol·L-1)1/2of SAR.The lowest soil sample spectral curve corresponded to 99.2 μs·cm-1of EC, 7.22 g·kg-1of SOM and 0.63 (mmol·L-1)1/2of SAR, respectively.

Fig.1 Spectral reflectance curves of ten soil samples

2.3 Spectral model of soil physico-chemical properties

Based on the modeling subsets, six spectral indices including R, SNV, NDVI, CR, LR, and FDR were used for establishing models of soil EC, SOM and SAR, respectively.Then the established models were validated by the validate subset, combining with the comparisons of the predicted and measured values to select the best model for the studied soil properties.The selection of best model of each soil property was described below.

2.3.1 Hyperspectral model of soil EC

Since the statistical test indicated the EC values were logarithmic normal distributed, logarithmic transformation was applied to obtain normal distributed EC data.Then the SLR analysis was conducted to establish models between logarithmic transformed EC values and the studied six spectral indices including R, SNV, NDVI, CR, LR and FDR at the 350～1 830 nm band where the water vapor absorption band were removed.The calibration and validation parameters for the established logarithmic models of EC using different spectral indices are shown in Table 2.

Table 2 Calibration parameters and validation results for the established models of logarithmic EC using different R-related indices

From the validation results,rvalues of the log(EC) models as functions ofRand FDR are both above 0.7, of whichrvalue of log (EC)～Rmodel was 0.782 and reached the highest, followed by the log (EC)～FDR model withrvalue of 0.723.While thervalues of the log (EC) models based on other four indices were lower than 0.54 with poor availability, and they cannot be the best model.In the two available models, RMSE values of the log(EC)～Rand the log(EC)～FDR models were 0.256 and 0.260, respectively.All of the established log(EC) models based on the sixR-related indices passedt-test.Combining with the comparisons ofr, RMSE andt-test values, the log(EC)～Rmodel was generally good.To further choose the best model for describing relationship between EC andR, direct comparison of the predicted and the measured log(EC) values are plotted in Figure 2.

Fig.re 2 shows that scatter plot of log(EC)～Rmodel concentrated more to 1∶1 line than that of the log(EC)～FDR model.TheR2value of the log(EC)～Rmodel was 0.772 and larger, therefore, log(EC)～Rmodel was selected as the best model in Manasi river basin.log(EC) value should be transformed to EC when the model is applied in practice.

2.3.2 Hyperspectral models of SOM

Similarly with the model establishment procedure of EC, SOM models were set up based onR, SNV, NDVI, CR, LR and FDR using the SLR method.The best SOM model was selected.The calibrated parameters and the validation results are shown in Table 3.

Fig.2 Comparison of the measured and the predicted log (EC) values for two models

Table 3 Calibration parameters and validation results for the established models of SOM using different R-related indices

In Figure 3, scatter plot based on the SOM～NDVI model was more close to 1∶1 line than that of the SOM～FDR model.R2of SOM～NDVI model was larger (0.709).Therefore the SOM～NDVI model was selected as the best spectral model to predict SOM in the Manasi River basin.

2.3.3 Hyperspectral model of SAR

Similar to the modeling procedure of EC and SOM, models were established between SAR and the spectral-related indices includingR, SNV, NDVI, CR, LR and FDR using the SLR method.The best performance of SAR model was selected, and the calibration and validation results are demonstrated in Table 4.

Fig.3 Comparison of the measured and predicted SOM values

Table 4 Calibration parameters and validation results for the established models of SAR using different R-related indices

From the validation results, models of SAR vs.R, NDVI and LR did not pass thet-test, thus these three models cannot be selected as the best model because of their weak availability.In the other three models, the SAR～FDR model had the highestrvalue of 0.647, followed by the SAR～SNV model withrvalue of 0.621 and the SAR～CR model withrvalue of 0.445, which was the smallest one and couldn’t be the best model.By comparingrvalues, RMSE values andt-test parameters of the established models comprehensively, the models of SAR～FDR and SAR～SNV both performed well, while SAR～FDR was better.In order to choose the best SAR model based on the six spectral reflectance indices, the predicted SAR values using SAR～FDR and SAR～SNV models and the measured values are plotted in Figure 4.

Fig.re 4 showed that the scatter plot of SAR～FDR model was more close to the 1∶1 line than that of SAR～SNV model.R2of SAR～FDR model was 0.557 and larger than that of SAR～SNV model.Therefore, the SAR～FDR model was chosen as the best model to predict SAR in the Manasi River basin.

3 Discussions

In this study, the models between log(EC) and the six spectral indices was established with sensitive wavelengths distributed in 350～1 801 nm.Currently, many scholars have done research on soil salinity improvement using hyperspectral remote sensing technology and they achieved good results[7-8,10-12,15-16].Nevertheless, for the spectral characteristics and sensitive bands of EC, there was no unified conclusion.Many studies showed that sensitive bands of SOM distributed in 400～1 100 nm and particularly focus on the 600～800 nm[17].Liu et al.suggested that sensitive wavelengths of SOM are 474, 636 and 1 632 nm, respectively[18].In this study, the most sensitive bands of SOM distributed in the visible and the near-infrared light bands, which was consistent with previous studies.Sensitivity wavelengths of SAR models vs.various indicators generally distributed in the visible light band.

4 Conclusions

(1)Statistical characteristics of the modeling subset for EC, SOM and SAR are close to those of the validate subset and total, so the divided subsets can be applied for the calibration and the validation of the established models and further prediction.

(3)By comparing the best hyperspectral models established for the three studied soil properties, the models for soil EC had the highest accuracy, the next was the SOM model, while the SAR model was of the most inaccurate.

[1] Mettemicht G I, Zinck J A.Remote Sensing of Environment, 2003, 85(1): 1.

[2] Tian C Y, Zhou H F, Liu G Q, et al.Arid Land Geography, 2000, 23(2): 177.

[3] Karlen D L, Tomer M D, Neppel J, et al.Soil Tillage Research, 2008, 99(2): 291.

[4] Meternicht G，Alfred Zinck J.Remote Sensing of Soil Salinization Impact on Land Management.New York: CRC Press, 2009.63.

[5] Manna M C, Swarup A.Soil and Tillage Research, 2007, 94(2): 397.

[6] Cambule A H, Rossiter D G, Stoorvogel J J, et al.Geoderma, 2012, 183-184: 41.

[7] Weng Y L.Pedosphere, 2010, 20(3): 378.

[8] Li Y, Liu S B, Liao Z H, et al.Canadian Journal of Soil Science, 2012, 92(6): 845.

[9] Marco N, Antoine S.Soil Biology and Biochemistry 2014, 68: 337.

[10] Aldabaa A A A, Weindorf D C, Chakraborty S, et al.Geoderma, 2015, 239: 34.

[11] Fan X W, Liu Y B, Tao J M, et al.Remote Sensing, 2015, 7: 488.

[12] Liu S B, Li Y, He C S.Soil Science, 2013, 178(3): 138.

[13] Feng Y, Luo G P, Zhou D C, et al.Acta Ecologica Sinica, 2010, 30(16): 4295.

[14] Hasheminejhad Y, Ghane F, Mazloom N.Communications in Soil Science and Plant Analysis, 2013, 44(18): 2666.

[15] Pang G, Wang T, Liao J, et al.Soil Science Society of American Journal.2014, 78: 546.

[16] Wang Q，Li P H，Maina J N, et al.Soil Science and Plant Analysis, 2013, 44(9): 1503.

[17] Takata Y, Funakawa S, Akshalov K, et al.Soil Science and Plant Nutrition, 2007, 53(3): 289.

[18] Liu Jiao, Li Yi, Liu Shibin.Spectroscopy and Spectral Analysis, 2013, 33(12): 3354.

[19] Zhang H, Li Y, Deng H W, et al.Journal of Northwest A&F University, 2013, 41(3): 153.

[20] Lü Z Z, Liu G M, Yang J S.Acta Pedologica Sinica, 2013, 50(2): 289.

[21] Mevik B H, Wehrens R.Journal of Statistical Software，2007, 18(2): 1.

*通訊聯(lián)系人

S127

A

西北鹽堿土理化性質(zhì)的高光譜建模及預測

肖珍珍1，李 毅1,2*, 馮 浩2,3

1.西北農(nóng)林科技大學水利與建筑工程學院，陜西 楊凌 712100 2.西北農(nóng)林科技大學中國旱區(qū)節(jié)水農(nóng)業(yè)研究院，陜西 楊凌 712100 3.西北農(nóng)林科技大學國家節(jié)水灌溉楊凌工程技術(shù)研究中心，陜西 楊凌 712100

高光譜數(shù)據(jù)具有光譜分辨率高、波段連續(xù)性強、信息豐富等特點，在土壤信息的監(jiān)測中得到廣泛應用。利用高光譜遙感技術(shù)測定鹽漬化土壤屬性對灌區(qū)農(nóng)作物的生長和農(nóng)業(yè)可持續(xù)發(fā)展具有重要意義。采集瑪納斯河流域221個土壤樣品，分別測定土壤電導率(EC)、有機質(zhì)(SOM)和Na+, Ca2+, Mg2+三種離子濃度含量等土壤理化性質(zhì)和光譜反射率曲線，并由三種離子含量得出鈉吸附比值(SAR)，采用逐步線性回歸方法建立EC，SOM和SAR與原始光譜反射率(R)、標準正態(tài)變量(SNV)、歸一化差異植被指數(shù)(NDVI)、倒數(shù)的對數(shù)(LR)、一階微分(FDR)和去包絡線(CR)等六種指標的模型。模型驗證結(jié)果表明，相較其他五種變量的模型，以R為自變量的EC對數(shù)模型精度最高，相關(guān)系數(shù)為0.782，均方根誤差為0.256。以NDVI為自變量的土SOM預測模型精度最高，相關(guān)系數(shù)為0.670，均方根誤差為5.352。以FDR為自變量的SAR預測模型精度最高，相關(guān)系數(shù)為0.647，均方根誤差為1.932。EC預測模型效果最好，SOM預測模型次之，SAR預測模型精度最低。最優(yōu)模型中EC，SOM和SAR的敏感波長分別分布于395～1 801，352～1 144和394～1 011 nm波段。由于土壤中各屬性的差異和不同成分空間分布的變異性，對于不同土壤性質(zhì)的建模和驗證結(jié)果差異較大。本研究可為鹽漬化土壤的高光譜遙感監(jiān)測提供依據(jù)。

電導率； 有機質(zhì)； 鈉吸附比； 高光譜模型

Foundation item：National High Technology Research and Development Program of China (SS2013AA100904), Natural Science Foundation of China (51579213), the China 111 Project (B12007), and China Scholarship Council for Studying Abroad (201506305014)

10.3964/j.issn.1000-0593(2016)05-1615-08

Received：2015-03-06; accepted：2015-07-10

Biography：XIAO Zhen-zhen, (1992—), Master Degree in College Water Resources and Architecture Engineering, Northwest A&F University e-mail: xiaozz0212@163.com *Corresponding author e-mail: liyikitty@126.com