**1. Introduction**

The Yellow River beach area is a vast, relatively flat area between the main channel of the Yellow River and the river's flood control levee. The beach area has been flooded and filled with silt from the river several times [1]. Since the completion of the Yellow River Xiaolangdi Dam, the floods have been controlled effectively and much of the beach area has been used as fertile cultivated land [2]. Since 2019, when the ecological protection and high-quality development strategy of the Yellow River basin were proposed in China, the local governments have increased the protection of the beach area [3–5]. To support the government's strategy, studies of the beaches and the cultivated land are needed to provide basic land use information, which will form the space basis for improving the safe protection and governance of the water and land resources in the beach area.

There have been some previous studies on the Yellow River beach area [6,7]. Some scholars have considered how the water body and beach land have developed and transformed spatially and temporally [8], and what the ecological benefits accrued from developing the cultivated land was [9–13]. The cultivated land and the water area were closely related, and the progress of water area means the retreat of cultivated land, and vice versa [14]. As various flood control projects and engineering repairs have been successfully implemented, both the speed and flow of the river have decreased, resulting in a decrease

**Citation:** Run, Y.; Li, M.; Qin, Y.; Shi, Z.; Li, Q.; Cui, Y. Dynamics of Land and Water Resources and Utilization of Cultivated Land in the Yellow River Beach Area of China. *Water* **2022**, *14*, 305. https://doi.org/ 10.3390/w14030305

Academic Editors: Chris Bradley and Jan Wesseling

Received: 12 November 2021 Accepted: 18 January 2022 Published: 20 January 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

in the water area and an increase in the area of exposed beaches [15]. However, the relationship between cultivated land and water body in the Yellow River beach has not been fully explored. It thus is a primary task to identify the balance between water, land, and cultivated land structure.

Recently, remote sensing images have been used increasingly to support land use analysis and research [16–18]. Satellite data like Landsat, Sentinel images can be used free of charge to support land use mapping and cultivated land monitoring [19–24]. Traditional methods for extracting ground objects mainly consider the gray value of pixels; however, this does not adequately use information about the image space and texture and may result in a low classification accuracy [25]. In the Third National Land Survey of China, based on the information from remote sensing images, the land use types are classified into 13 primary categories, such as wetland, cultivated land, and forest land [26]. However, the data of the Third National Land Survey cannot distinguish crop types, nor does it take into account the unique geographical environment, which means that there is insufficient information about crop growing activity in the beach area. Additionally, based on the support of high-resolution remote sensing data like Sentinel, mapping crop types have been widely explored in many studies [27,28]. Therefore, Sentinel data has great potential to accurately extract the information of cultivated land.

The purpose of this study is to study the dynamics of water and soil resources and the utilization of cultivated land in the Yellow River beach area. In this study, the area of cultivated land in the beach area was extracted from the Third National Land Survey. The area of continuous beach land over the last 34 years was extracted and the spatial and temporal changes from beach land to cultivated land were discussed. Object-Oriented Feature Extraction was used to extract staple crops from Sentinel-2A/B images. The results can provide a reference for the management and protection of cultivated land resources in this special area.

#### **2. Data and Methods**

#### *2.1. Study Area*

The study area is the section of the Yellow River that flows through Henan Province (Figure 1), which has the greatest extent of cultivated beach along the river [29]. The length of the river channel involved in the beach is 464 km, flowing through the middle reaches of the Yellow River (19.35%) and the lower reaches (80.65%), with a total area of 2673 km<sup>2</sup> [30]. The Taohuayu section is characterized by hilly terrain in the west and an alluvial plain in the east, with sediment deposits and an elevated river bed caused by frequently changing flow [31]. About a million people live here and the beach area is a semi-natural farming environment managed by humans. [31,32].

#### *2.2. Data*

We used the Landsat and Sentinel-2A/B images to extract the continuous beach land and staple crops. Landsat images with a long-time span were used to identify the continuous beach land on a scale of pixels. Sentinel-2A/B images have a spatial resolution of 10 m and numerous studies have used Sentinel images to identify crop types [33,34]. Therefore, Sentinel-2A/B images were used to identify crop types in this study.

Google Earth Engine (GEE) (https://earthengine.google.org (accessed on 7 March 2021)) is a cloud computing platform that provides a wealth of geospatial data [35,36]. Landsat Thematic Mapper (TM), Enhanced Thematic Mapper + (ETM+), and Operational Land Imager (OLI) for the Yellow River beach area from 1987 to 2020 were acquired from the GEE platform. The surface reflectance Landsat images from the GEE platform were geometrically and atmospherically corrected and were cross-calibrated between different sensors [37], and images with 0–20% cloud cover have been selected. A total of 4722 valid Landsat images (Figures 1 and 2) have been obtained. For each image, clouds, cloud shadows and snow pixels were removed using the data quality layer from a cloud masking method, CF-Mask, which helped to prepare the Landsat data for change detection [38–40].

*Water* **2022**, *14*, x FOR PEER REVIEW 3 of 15

**Figure 1.** Yellow River beach area in Henan Province. **Figure 1.** Yellow River beach area in Henan Province. method, CF-Mask, which helped to prepare the Landsat data for change detection [38–40].

sensors [37], and images with 0–20% cloud cover have been selected. A total of 4722 valid Landsat images (Figures 1 and 2) have been obtained. For each image, clouds, cloud shadows and snow pixels were removed using the data quality layer from a cloud masking method, CF-Mask, which helped to prepare the Landsat data for change detection [38–40]. The Sentinel 2A/B mission of Copernicus Europe comprises a multispectral and highresolution sub-satellite that employs a Multi-Spectral Imager (MSI) with 13 bands [41,42]. In this study, 72 Level 1C images of Sentinel-2A/B with sub-1% cloudiness for the Yellow River bank area from 2019 to 2020 were downloaded from the European Space Agency website (https://scihub.copernicus.eu (accessed on 15 March 2021)), which had undergone orthorectification and sub-pixel geometric fine correction [43]. The Sen2cor plug-in (http://step.esa.int (accessed on 3 April 2021)) was used to radiometrically calibrate and geometrically correct the images, and finally Level 2A atmospheric bottom reflectivity data

**Figure 2.** Statistics of the Landsat images of the Yellow River beach.

*2.3. Methods* 

was obtained [41,43]. The range of the Yellow River beach area, the data of the type of cultivated land within the beach area were from the Third National Land Survey. 2.3.1. Technical Framework

The Sentinel 2A/B mission of Copernicus Europe comprises a multispectral and highresolution sub-satellite that employs a Multi-Spectral Imager (MSI) with 13 bands [41,42]. In this study, 72 Level 1C images of Sentinel-2A/B with sub-1% cloudiness for the Yellow River bank area from 2019 to 2020 were downloaded from the European Space Agency website (https://scihub.copernicus.eu (accessed on 15 March 2021)), which had undergone orthorectification and sub-pixel geometric fine correction [43]. The Sen2cor plug-in (http://step.esa.int (accessed on 3 April 2021)) was used to radiometrically calibrate and geometrically correct the images, and finally Level 2A atmospheric bottom reflectivity

#### *2.3. Methods* Based on Landsat images downloaded from the GEE platform, water bodies were

#### 2.3.1. Technical Framework first identified using the relationship between the water body and vegetation indexes.

*Water* **2022**, *14*, x FOR PEER REVIEW 4 of 15

Based on Landsat images downloaded from the GEE platform, water bodies were first identified using the relationship between the water body and vegetation indexes. Then water body frequency maps were created and continuous beach lands were extracted based on thresholds. Using the data of the Third National Land Survey, the cultivated land information within the beach area was extracted. Finally, Object-Oriented Feature Extraction was used to extract crops (Figure 3). Then water body frequency maps were created and continuous beach lands were extracted based on thresholds. Using the data of the Third National Land Survey, the cultivated land information within the beach area was extracted. Finally, Object-Oriented Feature Extraction was used to extract crops (Figure 3).

**Figure 3.** Technical framework. **Figure 3.** Technical framework.

2.3.2. Water Bodies and Continuous Beach Land Identification 2.3.2. Water Bodies and Continuous Beach Land Identification

(1) Water body identification. The relationship between water body and vegetation indexes was used to identify water body, following the approach adopted in previous studies [44–46]. The modified Normalized Difference Water Index (*mNDWI*), Normalized Difference Vegetation Index (*NDVI*) and Enhanced Vegetation Index (*EVI*) were used. The (1) Water body identification. The relationship between water body and vegetation indexes was used to identify water body, following the approach adopted in previous studies [44–46]. The modified Normalized Difference Water Index (*mNDWI*), Normalized Difference Vegetation Index (*NDVI*) and Enhanced Vegetation Index (*EVI*) were used. The formulas for these water and vegetation indices are as follows:

$$m\text{NDVI} = \frac{\rho\_{Green} - \rho\_{Swi1}}{\rho\_{Green} + \rho\_{Swi1}} \tag{1}$$

$$\rho\_{Ni-} = \rho\_{To}$$

$$NDVI = \frac{\rho\_{Nir} - \rho\_{Red}}{\rho\_{Nir} + \rho\_{Red}} \tag{2}$$

$$EVI = 2.5 \times \frac{\rho\_{Nir} - \rho\_{Red}}{1.0 + \rho\_{Nir} + 6.0 \rho\_{Red} + 7.5 \rho\_{Blue}} \tag{3}$$

where *ρBlue*, *ρGreen*, *ρRed*, *ρNir*, and *ρSwir1* are the surface reflectance values of bands Blue, Green, Red, Near-infrared and Shortwave-infrared-1 in the Landsat images.

When *mNDWI* > *EVI* or *mNDWI* > *NDVI*, the signal of water body in this area was stronger than that of vegetation. In order to further remove the noise, the mixed pixels of water and vegetation were removed using *EVI* < 0.1. Only the pixels that met the conditions (i.e., (*mNDWI* > *EVI* or *mNDWI* > *NDVI*) and (*EVI* < 0.1)) were classified into water body, while others were non-water pixels.

(2) Annual Water Frequency Calculation [40]:

$$F(y) = \frac{1}{N\_y} \sum\_{i=1}^{N\_y} \mathcal{W}\_{y,i} \times 100\% \tag{4}$$

where *F* is the water frequency of the pixel, *y* is the specified year, *N<sup>y</sup>* represents the total numbers of Landsat satellite observations for the pixel in that year and *Wy,i* represents whether a pixel is water in that year, where 1 is water and 0 is non-water. The water frequency map shows the annual presence or absence of water in each pixel since 1987.

(3) Continuous Beach Land

Most of the noise caused by poor data quality can be eliminated by choosing an appropriate water frequency threshold [30]. The area of beach with 0–25% of simultaneous water frequency from 1987 to 2020 was extracted and recognized as continuous beach land. The pixels with continuous beach land indicated stable non-water-covered areas were considered to be available beach land.

#### 2.3.3. Object-Oriented Feature Extraction

In this study, Object-Oriented Feature Extraction was used to extract crops. Pixels with similar internal features were composed into new objects, and staple crops were extracted according to specific rules. These were mainly divided into two processes, namely multiscale segmentation and object feature extraction [47,48] (Figure 3).

#### (1) Multi-Scale Segmentation [49]

In this process, a boundary-based segmentation algorithm was used for multi-scale segmentation, where pixels with similar spectral features were combined, and the average value was calculated and compared. Each segmentation object with similar internal feature information was filled in this scale. Each segmented object has a rich spectrum, texture, space and other information. Each segmentation object with similar internal feature information was filled in this scale. The object was the basic unit of object-oriented image processing based on rule classification. During image segmentation, some objects will be misclassified, and some objects may be divided into many parts, these problems can be solved by merging merge scale.

#### (2) Object Feature Extraction

The Object-Oriented Feature Extraction model is based on the principle of decision tree classification. The decision tree classification method is carried out layer by layer [50]. In this process, the areas were classified as vegetation-covered areas and non-vegetationcovered areas using *NDVI* and spectral values of the crop. When *NDVI* was greater than 0.35, it was a vegetation-covered area and when *NDVI* was less than 0.35, it was a nonvegetation-covered area [51]. The vegetation-covered areas were then further distinguished using the spectral values of winter wheat and summer corn as the rules.

#### 2.3.4. Accuracy Verification

This process follows the concept of an error matrix [52]. the producer's accuracy, the user's accuracy, and the overall accuracy, were used in the accuracy assessment.

Producer's accuracy includes producer accuracy (*Pa*) of goal region and producer accuracy (*Pn*) of non-goal region. They are defined as follows:

$$Pa = TP / (TP + FN) \tag{5}$$

$$Pt = TN/(FP + TN) \tag{6}$$

**3. Results** 

3% of the cultivated land, respectively.

and Wuzhi, had no areas of dry land.

User's accuracy includes user accuracy (*Ua*) in goal region and user accuracy (*Un*) in non-goal region. They are defined as follows:

$$
\Box a = \text{TP} / (\text{TP} + \text{FP}) \tag{7}
$$

$$\text{Ult} = \text{TN} / (\text{TN} + \text{FN}) \tag{8}$$

where *TP* (true positive) is the number of goal pixels correctly extracted; *FN* (false negative) is the number of goal pixels extracted as non-goal; *FP* (false positive) is the number of non-goal pixels extracted as a goal, and *TN* (true negative) is the number of non-goal pixels correctly extracted.

Overall accuracy (*Oa*) is:

$$\text{Oa} = (TP + TN) / (TP + TN + FP + FN) \tag{9}$$

where *TP + TN* is the number of correctly extracted true goal and non-goal pixels; *TP + TN + FP + FN* is equal to the number of total pixels in the image. The overall accuracy can be used to evaluate the correctness percentage of the detection algorithm.

For each land cover type, based on different phases of Sentinel-2 high-resolution images, sample points were randomly generated by GEE for accuracy verification, here 400 sample points were generated for each water body, winter wheat and summer corn, respectively. Water bodies were verified using the method of Wang et al. [53], a water body with water frequency ≥25% in 2020 water frequency image was used as verification. Some sampling points were shown in Figure 4. The overall accuracy of water body (96%), winter wheat (97.1%), and summer corn (94.55%) met the needs of this study (Table 1). *Water* **2022**, *14*, x FOR PEER REVIEW 7 of 15

**Figure 4.** Spatial extraction accuracy. (**a**–**c**) compare the extracted accuracy with high-resolution RGB images of winter wheat, summer corn, and water bodies, respectively. **Figure 4.** Spatial extraction accuracy. (**a**–**c**) compare the extracted accuracy with high-resolution RGB images of winter wheat, summer corn, and water bodies, respectively.

The statistics from the Third National Land Survey indicated that a large proportion of the Yellow River beach area was cultivated land. Cultivated land, garden land and other agricultural land accounted for 66.04% of the total beach area. Among them, cultivated land accounted for the largest proportion (58.26%) and covered an area of 1557.19 km2 (Figure 5). Irrigable land, paddy field and dry land accounted for 97% and

The area of dry land accounted for 2.20% of the cultivated land, which was unevenly

distributed between the north and south banks of the Yellow River, counties and cities along the Yellow River beach were shown in Figure 5. Of the dry land, 82.84% distributed throughout seven counties on the south bank and 17.16% distributed through six counties on the north bank. Gongyi county had the largest area of dry land (9.95 km2), followed by Xiangfu district (8.68 km2). Six counties, including Lankao, Puyang, Longting, Mengzhou


**Table 1.** Error matrix for accurate evaluation.

#### **3. Results**
