Soil Salizanation Information in the Yellow River Delta Based on Feature Surface Models Using Landsat 8 OLI Data

Soil salinization as one of the major eco-environmental problems, has greatly restricted the regional development of Yellow River Delta. In this paper, five parameters that derived from Landsat images, including MSAVI, Albedo, SI, <inline-formula> <tex-math notation="LaTeX">$I_{\mathrm {Fe2O3}}$ </tex-math></inline-formula>, and WI have been utilized to establish ten feature spaces, and four categories of salinization detection model have been proposed. After analysis and comparison of inversion accuracy, the three typical surface parameters, including SI (<inline-formula> <tex-math notation="LaTeX">$R^{2}=0.85$ </tex-math></inline-formula>), <inline-formula> <tex-math notation="LaTeX">$\text{I}_{\mathrm {Fe2O3}}$ </tex-math></inline-formula> (<inline-formula> <tex-math notation="LaTeX">$R^{2}=0.83$ </tex-math></inline-formula>), and WI (<inline-formula> <tex-math notation="LaTeX">$R^{2}=0.83$ </tex-math></inline-formula>) are better indices to retrieve the salinization information and the WI-SI point to line (soil line) model(<inline-formula> <tex-math notation="LaTeX">$R^{2}=0.88$ </tex-math></inline-formula>), the Albedo–SI point to line (wet line) model(<inline-formula> <tex-math notation="LaTeX">$R^{2}=0.87$ </tex-math></inline-formula>) and the Albedo–<inline-formula> <tex-math notation="LaTeX">$I_{\mathrm {Fe2O3}}$ </tex-math></inline-formula> point to point model(<inline-formula> <tex-math notation="LaTeX">$R^{2}=0.86$ </tex-math></inline-formula>) have better applicability to monitor the salinization condition in the Yellow River Delta. The research results can provide technical method reference for salinization monitoring of other regions.


I. INTRODUCTION
Yellow River Delta has the most integrated new-born salinization system in China [1], [2]. It has prominent soil salinization and fragile ecological environment because of less precipitation, large evaporation, serious uneven drought and flood, and significant seasonal dry and wet alternation [2]- [4]. Soil salinization has critically restricted the development of regional agricultural economy in the Yellow River Delta [5]- [7]. Thus, it is urgent to obtain the soil salinization information, which can provide decision supports for the prevention and governances of soil salinization [2], [8]- [11].
Traditional studies on soil salinization are mostly based on field measurements, which are characterized by expensive cost, time-consuming and small spatial scale [1], [2], [12]- [16]. During the past decades, remote sensing and Geographic Information System technology have provided a fast and accurate new approach to obtain the salinization information of large-scale regions [2], [17]- [22].
The associate editor coordinating the review of this manuscript and approving it for publication was Le Hoang Son .
Scudiero et al. [23] applied the linear modeling of remote sensing vegetation indices to explore the potentials of assessing soil salinity. Ivushkin et al. [3] produced a global soil salinity map by combining thermal infrared images, soil texture data and field observations with machine learning method. Wang et al. [24] conducted the detailed comparison in monitoring salinization between Sentinel-2 MSI and Landsat 8 image and found that MSI image performed better than Landsat 8 image. In recent years, the feature space models that constructed by surface parameters have been applied to monitor the soil salinization [25], [26]. Wang et al. [10] utilized the biophysical parameters, such as vegetation coverage, soil moisture and surface evapotranspiration, to propose a ''Triangle methods'' to analyze the process of soil salinization. Ha et al. [4] constructed the SI-Albedo feature space model and found that the change of soil salinity and soil moisture content had significant relations with soil salinization. Ding et al. [27] had proposed the MSAVI-VI feature space model to detect the regional salinization information and found that there existed significant relationships between MWI index and surface soil salinity. Zhang et al. [28] had explored the relationship between modified soil adjusted vegetation index (MSAVI) and salinity index (SI) based on the Landsat OLI image and field survey data and then proposed the salinization remote sensing information extraction model. However, only the linear relations between different feature space parameters has been considered in most previous relative studies, which has ignored the complicated impacts of biological and abiotic factors on salinization process [2], [29], [30].
Therefore, after fully considering the landscapes types of Yellow River Delta, five parameters that derived from Landsat 8 OLI images have been applied to establish ten monitoring indices based on four categories of feature space models for Yellow River Detla. Finally, the optimization detection model of soil salinization has been proposed.

A. DATA COLLECTION AND PREPROCESSING
In this study, two kinds of data have been utilized to construct the models:(1) Satellite images;(2) field observed data. The Landsat 8 OLI data (Oct 26, 2016; 122/34) was used to retrieve the salinity index with spatia-temporall resolutions of 30m and 16 days, respectively. This above dataset can be obtained from USGS. The FLAASH models have been utilized to remove the influence of atmosphere and light on the reflection of ground objects [2], [7]. Figure 1 showed that 32 surface soil samples (30m×30m) with a depth of 0-10 cm were chosen from zones with diverse landscapes during Oct [25][26]2016. Each sample was composed of five observed data with plum blossom shape. The 32 soil samples were crushed passed through a 2-mm mesh and then the suspensions of soil and water with 1:5 ratio were utilized to obtain the salt content [2], [14]. Previous studies showed that vegetation coverage would became become sparse with the deterioration of soil salinization [2], [10], [15], [22]. Therefore, the MSAVI is a better indicator to reflect the salinization process, especially in the arid and semi-arid zones [2], [4]. Surface albedo (Albedo) is an important parameter to indicate the short wave radiation from the earth's surface to the sun. Its value would be greatly influenced by the soil surface conditions [2], [11], [17]. With the aggravation of salinization, the surface albedo increased correspondingly. As shown in Figure 2, there was a negative correlation between Albedo and MSAVI. The upper boundary A-D can reflects the severe salinization condition, which is the maximum Albedo corresponding [7], [19]. The lower boundary B-C represents the low Albedo line and it can reflects the slight salinization condition [2].

C. INDEX OF THE FEATURE SPACE MODEL
Albedo can reflect the short wave radiation from the land surface [2], [11], [15], [23], [31]- [33]. The vegetation growth and albedo can be indirectly affected by the soil salinization process. Thus, both albedo and vegetation indices (such as MSAVI) are better surface parameters to monitor the soil salinity [2], [15]. Salt in seawater is another dominant factor of soil salinity. Salinity index (SI) can directly reflect the salinization condition. Wetness index (WI) has significant negative effects on soil salinization [2], [7], [15], [27]. The soil salinization process will not only affect the vegetation growth, but also affect chemical substances contained in the soil. Therefore, the iron oxide index (I Fe2O3 ) is also a better indicator for the soil salinization process [2], [9]. The five indices that derived from Landsat 8 OLI image, including Albedo, WI, MSAVI, SI, and I Fe2O3 are calculated as follows [2], [9], [11], [15]   where Blue, Green, Red, Nir, SW 1 , and SW 2 are the reflectance of blue band, green band, red band, mid-infrared band, and near-infrared band, respectively.

D. INDEX STANDARDIZATION
There existed great difference among parameters, and the data standardization should be conducted to eliminate these 94396 VOLUME 8, 2020 differences [7], [11], [15], [34], [35]: where A refers to the standardized surface albedo index; Al refers to the surface albedo index; Al min and Al max refer to the minimum and maximum value of surface albedo index, respectively.
where W refers to the standardized wetness index; WI refers to the wetness index; WI min andWI max refer to the minimum and maximum value of wetness index, respectively.
where M refers to the standardized vegetation index; VI is the vegetation index; VI min andVI max refer to the minimum and maximum value, respectively.

S =
Si − Si min Si max −Si min (9) where S refers to the standardized salinity index of i; Si refers to the salinity index; Si min and Si max refer to the minimum and maximum value of salinity index, respectively.
VOLUME 8, 2020 where I is the standardized iron oxide index; I Fe2O3 is the iron oxide index; I Fe2O3min and I Fe2O3max are the minimum and maximum value, respectively.

A. CONSTRUCTION OF FEATURE SPACES
The zones with water and artificial impervious surface were extracted firstly with ENVI 5.3 [2]. Then, five indexes including Albedo, MSAVI, SI, WI, and I Fe2O3 were applied to establish ten feature spaces. According to the spatial change patterns of soil salinization process in feature space, ten feature spaces could be divided into four categories: (1) point to point feature space, includes Albedo-

B. SPATIAL DISTRIBUTION SALINIZATION MODELS
As was shown in Figure 4, we chose four point clusters to indicate different levels of soil salinization according to the distance to the point (0.8, 0) in Albedo-I Fe2O3 feature space. The results showed that the spatial distributions of different categories of soil salinization (non salinization, mild salinization, moderate salinization, and severe salinization) differed greatly in the Albedo-I Fe2O3 feature space. Figure 5 showed that the distance to the L line that is parallel to the ''wet line'' could be utilized to reflect the soil salinization process in Albedo-SI feature space. The soil salinization condition would become more severe with the increasing distance that from the L-line [2].
As was shown in Figure 6, this fit curve could better reflect the change process of soil salinization in I Fe2O3 -MSAVI feature space. These spatial distribution laws could help to distinguish levels of soil salinization.  Figure 7, the distance from any point to L line that was parallel to the soil line could better reflect the degrees of salinization condition in Albedo-SI feature space. The salinization would be more severe when the distance from the L-line became larger [2].

Similar to
As was shown in Figure 8, the distance to D (0.8, 0) could better explain the soil salinization process in Albedo-I Fe2O3 feature space. The further away any point from point D, the more severe soil salinization. The salinization detection function (SDI 1 ) was developed as follows: As shown in Figure 9, the distance to L line could indicate the soil salinization process in Albedo-SI feature space. The further away from the L line, the more severe soil salinization. VOLUME 8, 2020 The salinization monitoring function (SDI 2 ) was proposed and M referred to the slope of ''wet line'': Figure 10 showed that the I Fe2O3 -MSAVI feature space was divided into different parts in the vertical direction of the curve, which could reflect the salinization process. And then different levels of soil salinization could be better distinguished. The region perpendicular to the I Fe2O3 -MSAVI feature space can be determined by a binary linear function considering the linear relationship between the above factors.
where a refers to the slope of the linear equation. Similar to Figure 11, the distance to L line could be utilized to distinguish different levels of soil salinization in WI-SI feature space. The soil salinization would become more severe with the increasing distance to the L line (parallel to the soil line). The salinization monitoring function of soil salinization (SDI 4 ) was established.
where p is the slope of ''soil line''.

IV. RESULTS AND DISCUSSIONS
Ten salinization detection indices were obtained utilized the above four categories of models, then the inversion accuracies have been analyzed and compared based on 32 field observation samples( [2], Table 1 and Figure 1). The results ( Table 2) showed that WI-SI point to line (soil line) model had the largest inversion accuracy with R 2 = 0.88, followed by Albedo-SI point to line (wet line) model (R 2 = 0.87) and Albedo-I Fe2O3 point to point model (R 2 = 0.86). On the contrary, the Albedo-MSAVI point to point model had the smallest inversion accuracy with R 2 = 0.77. This result was consistent with that of Wang et al. [10] and Guo et al. [34]. The reasons lied in the fact that the point to line (wet line or soil line) models had fully considered the non-linear relations among different variables. It could better eliminate the effects of soil background or vegetation saturation effect [2], [35]- [38]. As shown in Figure 12, the average precision of feature space models that contained SI was the best with R 2 = 0.85, followed by that of I Fe2O3 (R 2 = 0.83) and WI (R 2 = 0.83). On the contrary, the average precision of feature space model that contained MSAVI was the worst with R 2 = 0.80. This above conclusions were consistent with the studies of Ivushkin et al. [3] and Guo and Wen [7]. The reason was that the salinity would precipitate out of the surface soil, which apparently affected the surface albedo [2], [34], [32]- [36]. Moreover, iron oxide was an important factor that affected the spectral characteristics of the soil [9]. Many absorption characteristics of soil in the visible band had significant relations with iron oxides. And the reflectivity of soil would decrease with the presence of iron oxides [2], [9], [40]. Moreover, the chemical weathering played an important role in the soil salinization process [2], [41]. Humidity index derives from the third component of Hat Transformation with Landsat8 OLI images, which can better reflects the condition of soil humidity in the study area [2], [34], [42]. There existed better relations between soil moisture and soil salinization process [42]. Salt in seawater was another dominant factor of soil salinity [2]. Soil humidity would be significantly affected by seawater, so that the soil salinization occurred commonly in the study area [27], [42]. Although, MSAVI could better reflect the vegetation condition, there were many types of salt-tolerant plants. And vegetation coverage was not the smallest in some region with severe salinization [2], [15], [41], [43]. Therefore, the MSAVI was not the optimal indices to reflect the condition of soil salinization for this study area.
In conclusion, three typical surface parameters, including SI, I Fe2O3 , and WI were better indices to retrieve the salinization information. And for the Yellow River Delta, the WI-SI point to line (soil line) model, the Albedo-SI point to line (wet line) model and the Albedo-I Fe2O3 point to point model have better applicability to detect the salinization conditions.

V. CONCLUSION
In this paper, five inversed parameters that derived from Landsat 8 OLI image, including MSAVI, Albedo, WI, I Fe2O3 , and SI have been applied to construct ten feature spaces. And then four categories of models have been established. After analysis and comparison, three typical surface parameters, including SI (R 2 = 0.85), I Fe2O3 (R 2 = 0.83), and WI (R 2 = 0.83) are better indices to retrieve the salinization information and the WI-SI point to line (soil line) model(R 2 = 0.88), the Albedo-SI point to line (wet line) model(R 2 = 0.87) and the Albedo-I Fe2O3 point to point model(R 2 = 0.86) have better applicability to monitor the salinization condition in Yellow River Delta. However, further research is needed to clarify the comprehensive and interaction effects of different factors (climate and human activities) on soil salinization process.
WENQIAN ZANG was born in Xinxiang, Henan, in 1984. He received the B.S. and M.S degrees in resources environment and the management of urban and rural planning from Henan Polytechnic University, Jiaozuo, in 2006, and the Ph.D. degree in geography from the University of Chinese Academy of Sciences, Beijing, in 2013.
From 2013 to 2019, he was an Assistant Professor with the Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences. He is the author of more than 20 articles and two Chinese National Invention Patents. His research interests include remote sensing image processing, cluster computing, and natural hazards.
Dr. Zang was a member of the China Association of Remote Sensing Application and won the first prize of science and technology progress award of Xinjiang Uygur Autonomous Region, in 2017.
RUI ZHANG received the Ph.D. degree in cartography and geography information system from the Institute of Remote Sensing and Digital Earth (RADI), in 2019. His scientific interests are in the fields of land use classification, remote sensing monitoring of disaster environment based on machine learning, and spatial information science including remote sensing, GIS, and their integration. VOLUME 8, 2020