Identification and Analysis of Long-Term Land Use and Planting Structure Dynamics in the Lower Yellow River Basin

Abstract

Accurate identification of the spatio-temporal planting structure and analysis of its driving factors in an irrigation district are the important bases for scientific and reasonable utilization of irrigation water resources. In pursuit of this goal, the training sample migration method combined with the random forest algorithm were used to classify land use and planting structure over 2001–2022 in the lower Yellow River Basin. Moreover, an econometric regression model was applied to quantify the driving factors of the change in the crop-planted area. The results illustrated that the identification method has relatively high accuracy in identifying historical periods of land use and planting structures, with the average kappa coefficient equating to 0.953. From 2001 to 2022, the area of cultivated land was the largest, with the proportion of the total area increasing from 45.72% to 58.12%. The planted area of winter wheat–summer maize rotation increased from 74.84% to 88.11% of the cultivated land. While the planted area of cotton declined by 96.36%, about 50% of cotton planting was converted to the winter wheat–summer maize rotation planting. The government policies about grain purchase and storage were the dominant factors for the change in the crop-planted area. This resulted in an increase of 63.32 × 10³ ha and 63.98 × 10³ ha in the planted area of winter wheat and summer maize, respectively. The findings are of great significance to the allocation of water resources in irrigation districts of the lower Yellow River Basin.

1. Introduction

The spatio-temporal distribution of land use and planting structure in a region can influence policymaker decisions in water resource management, urban planning, and agricultural production management [1]. The lower Yellow River Basin (YRB) is one of the important grain crops (including wheat and maize) production regions in China, and the grain production accounts for about 18.30% of the total grain production [2]. However, the spatio-temporal distribution of precipitation in the lower YRB is uneven, and extreme drought events occur frequently, resulting in a large reduction in grain crops. Especially in recent years, the precipitation in the lower YRB has gradually decreased, while the air temperature has gradually increased, which would aggravate the pressure on agricultural water in the lower YRB [3]. So, irrigation has become a key security feature for agricultural production in the lower YRB. As shown in the Yellow River Basin Statistical Yearbook, agriculture in the lower YRB accounts for more than 60% of the total water consumption [4,5]. Therefore, to realize the scientific allocation of water resources in this region, it is essential to identify the land use and planting structure in the lower YRB and analyze the driving factors affecting the change in planting structure.

The identification of land use and planting structure presents a massive challenge due to the influence of the climate, natural resources, human activities, and so on. Remote sensing has great potential for identifying the land use and planting structure [6]. With the improvement of temporal and spatial resolution of satellite images as well as computational power, remote sensing technology has been widely applied to identify cultivated-land-cover information over large basins [7]. In previous studies, researchers have used multispectral and multi-temporal remote sensing data for land use mapping [8]. However, there are still deficiencies in previous studies. For example, in areas with complex and diverse land use types, the satellite images with low spatial resolution will affect the recognition accuracy of land use mapping. Moreover, unsuitable images make it difficult to identify the spectral characteristics of the key growth period in performing planting structure recognition. These factors lead to higher requirements for the quality and accuracy of remote sensing data. Moreover, different classification algorithms would affect the accuracy of land use information extraction. In previous studies, multiple classification algorithms (e.g., classification and regression trees, minimum distance method, random forest, support vector machine, etc.) have been applied to land use classification and crop growth detection [9,10,11]. The random forest (RF) algorithm is a combinatorial classification algorithm for learning from ensembles, which has been an effective way for long-term land use and planting structure identification because of the fast operation speed and high accuracy. Additionally, it has been demonstrated that the RF algorithm can be applied to map the spatio-temporal dynamics of an agricultural planting structure [12,13].

The selection of characteristic band sets may also be an important base for the accurate identification of land use and planting pattern. In previous studies, various characteristic bands have been proposed, such as normalized vegetation index (NDVI), normalized water body index (NDWI), and ratio vegetation index (RVI). It was found that the accuracy of land use identification can be effectively improved with the inclusion of these indicators [14,15]. For example, Luo et al. (2023) identified the planting structure in the North China Plain (NCP) using NDVI as the characteristic band [16]. Zhang et al. (2021) analyzed the spatial and temporal dynamics of winter wheat growing areas in the NCP using EVI as the characteristic band [17]. Therefore, except for the utilization of appropriate classification algorithms, it is necessary to incorporate some appropriate feature indicators in order to classify land use and planting structure accurately.

The number and quality of samples are the prerequisite for land use and planting structure identification. For a single or special growing season, it can be easy to use remote sensing image data to extract the planting structure in a basin or irrigated district on the basis of field survey samples. However, the lack of sufficient training samples limited the long-term dynamic identification of the land use and cropping pattern. Currently, there are relatively few studies focused on the long-term dynamics of planting structures in watersheds or irrigation districts. Among the existing studies, the threshold method had been commonly applied to study the long time series dynamics of planting structures [18,19]. This method determines the threshold range by analyzing the band characteristics and vegetation index of these samples in remote sensing images. In other years, different crops were identified by the threshold range. However, the determination of the threshold range usually depends on a subjective judgement, which ultimately affects the identification results of the planting structure. In addition, many studies have selected land use samples using intuitive interpretation of high-resolution images and then carry out long time series identification of land use and planting structure [20,21]. This method is also affected by the subjective judgement of researchers, resulting in the selection of land use samples that may deviate from the actual situation. More recently, the training sample migration (TSM) method has been applied to obtain sample data from the historical period [22] to overcome the disadvantage of insufficient samples. The TSM method screens for unchanged land use samples by comparing the changes in spectral angle between the same points in other years and reference years. This method is not affected by a subjective judgement and can objectively obtain the land use samples in the historical period.

Although the analysis of land use and planting structure is a simple process by traditional remote sensing processing methods, it is difficult to efficiently analyze land use and planting structure changes over a long time and large area due to the limitation of local hardware facilities [23]. The Google Earth Engine (GEE) platform has been widely applied in natural resource monitoring, land cover classification, and crop-planting area extraction in recent years because of the powerful computing and storage capabilities [24,25]. The GEE platform database can be used by online programing, which cannot be used by ordinary local software. Moreover, GEE can well meet the computational requirements of remote sensing data to extract planting structures with a large computational volume. Therefore, it can be used for long time series identification and extraction of land use and planting structure.

The analysis of driving factors affecting planting structure is of great significance for optimizing the crop industry structure and realizing the reasonable distribution of agricultural water resources in the lower YRB. The factors affecting planting structure change can be divided into natural factors and social factors. The natural factors include air temperature, precipitation, topography, soil type, etc. The social factors include government policies, crop purchase price, planting cost, etc. In most previous studies, the analysis of the driving factors influencing planting structure is the quantitative analysis of natural factors [26] and qualitative analysis of social factors [27]. In general, the effects of natural factors on the changes of land use and planting structure were analyzed by empirical statistical methods, such as the analytic hierarchy process [28], geographically weighted regression [29], the logistic regression model [30], and principal component analysis [31]. However, these methods cannot be used to quantify the impact of the special one or more social factors because of the heterogenicity of these factors. An econometric regression model has been applied to analyze the correlation between different variables [32]. It can provide the influence degree among the corresponding variables [33]. Therefore, it might provide an efficient tool for analyzing the effects of drivers on land use and planting structures.

To our knowledge, the long-term land use and planting structures and their driving factors in the lower YRB over the last 22 years have not been well investigated. Therefore, the aims were (1) to evaluate the classification accuracy of the TSM combined with the RF method in identifying historical periods of land use and planting structure in the lower YRB; (2) to analyze the spatio-temporal variation in land use and planting structure over the last 22 years in the lower YRB; and (3) to analyze the significant factors affecting the crop-planted area.

2. Materials and Methods

2.1. Study Site Description

In this research, the study site was selected to be in the Weishan Irrigation District (WID), Liaocheng, Shandong Province, China (36°07′–37°10′N, 115°16′–116°33′E) (Figure 1). WID covers an area about 5700 km² and is one of China’s six major irrigation areas, with the Yellow River as one of its water supply sources. The study area consists of a diverse range of land use types, including farmland, forest, watersheds, and urban areas, and is situated in a flat plain with a relatively low elevation. The main grain crops grown in the study site are winter wheat and summer maize; cotton and garlic are also grown (

Table 1. Distribution of main crop-growth periods in the Weishan Irrigation District.

Crops	Growth Period			Phenophase a (Days)
Winter wheat	October–June b			220–240
Garlic	October–May c			220–240
Summer maize			June–October d	90–110
Cotton		April–October e		150–180

^a The phenological period refers to the period from sowing to complete harvest. ^b Winter wheat was sown in October of the first year and harvested in June of the second year. ^c Garlic was sown in October of the first year and harvested in May of the second year. ^d Summer maize was sown in June and harvested in October of the same year. ^e Cotton was planted in April and harvested in October of the same year.

). Due to its location in the temperate monsoon climate zone, the annual rainfall is uneven, annual precipitation is 60% and concentrated in June to August, and the average annual rainfall is 550 mm.

2.2. Data Collection

2.2.1. Sample Collection

According to land use types, the samples were classified into cultivated land, construction, forest, and water. In addition, the cultivated land was reclassified into winter wheat, summer maize, garlic, cotton, and other crops. Moreover, crop rotation is the main planting pattern in WID, including winter wheat–summer maize rotation (WMR) and garlic–summer maize rotation (GMR). So, the crop rotation was also considered in this research. The sample points were collected from June to July 2022 using field research methods. The geographic location of the samples was obtained using a global positioning system tool. The percentage of each type of sample (WMR, GMR, cotton, other crops, construction, forests and water) used for classification was 18.04%, 14.01%, 6.01%, 14.25%, 18.04%, 17.62%, and 12.03%, respectively. All survey samples were uniformly distributed in the study area. The distribution of samples collected by field research is shown in Figure 2.

2.2.2. Remote Sensing Data

In this study, the remote sensing images utilized were sourced from the GEE, covering the time period from 2001 to 2022. To meet the temporal and spatial resolution requirements of crop spatio-temporal distribution classification, Landsat products that have undergone 2A-level standard geometric correction and atmospheric correction processing were selected. In our research, land use type and planting structure of WID from 2001 to 2022 were identified. While the time range of availability dataset for Landsat 5, 7 and 8 were from 16 March 1984 to 5 May 2012, from 28 May 1999 to present, and from 18 March 2013 to present, respectively (

Table 2. Remote sensing product information and characteristic from GEE.

Product	GEE ID	Dataset Availability	Usable Images
LANDSAT 5 ETM	LANDSAT/LT05/ C01/T1_SR	16 March 1984– 5 May 2012	111
LANDSAT 7 ETM+	LANDSAT/LE07/ C01/T1_SR	28 May 1999–	178
LANDSAT 8 OLI/TIRS	LANDSAT/LC08/ C01/T1_SR	18 March 2013–	174

). Because the scan lines corrector of Landsat 7 failed on 31 May 2003, all subsequent data became abnormal. Therefore, three remote sensing image data products were used in our research. The spectral reflectance for each band among three different Landsat products was different. To address this issue, the spectral reflectance conversion equations proposed by Roy et al. (2016) for Landsat 5 (L5), Landsat (L7), and Landsat (L8) [34] were used. Band parameters of these conversion equations can be found in

Table 3. Band parameter information concerning Landsat series satellites from GEE and eliminating spectral differences.

Band	Satellite Data			Transform a
	Landsat 5	Landsat 7	Landsat 8
Blue	B1	SR_B1	SR_B2	L7 = 0.0183 + 0.8850 × L8
Green	B2	SR_B2	SR_B3	L7 = 0.0123 + 0.9317 × L8
Red	B3	SR_B3	SR_B4	L7 = 0.0123 + 0.9372 × L8
Near infrared (NIR)	B4	SR_B4	SR_B5	L7 = 0.0448 + 0.8339 × L8
Shortwave infrared (SWIR) 1	B5	SR_B5	SR_B6	L7 = 0.0306 + 0.8639 × L8
Shortwave infrared (SWIR) 2	B7	SR_B7	SR_B7	L7 = 0.0116 + 0.9165 × L8

^a The band range of Landsat 5 and Landsat 7 is the same, but the band range of Landsat 8 and Landsat 7 is different, and the difference between the data of Landsat 8 and Landsat 7 needs to be transformed and eliminated.

. In this study, the cloud coverage of the satellite images was less than 20%. The cloud masking was performed on each satellite image to exclude the effects of clouds and snow.

2.2.3. Statistical Data Collection

The statistical data on the crop-planted area (winter wheat and summer maize) used for validation was obtained from the Weishan Irrigation District Administration in Liaocheng City (No. 2 Chenkou Road, Dongchangfu District, Liaocheng City, China).

The planting structure was influenced by various factors, which can be divided into natural factors and social factors. As for natural factors, average annual temperature (AT) and precipitation (Pre) were considered (Figure S1), which were derived from the China Meteorological Data of National Meteorological Center (http://data.cma.cn/, accessed on 1 January 2021). As for social factors, social subjective factors and social objective factors were included. There are seven social objective factors studied in this research, namely the price index of agricultural products (API), the price index of farm hand tools (FHTI), the price index of semi-mechanized farm implements (SMFI), the price index of mechanized farm implements (MFI), the price index of chemical fertilizer (CFI), the price index of pesticides and pesticide instruments (PAPI), and the price index of multipurpose tractor oil (MPTOI). The corresponding statistics data of the social objective factors were derived from the Rural Statistical Yearbook of China and the Chinese Statistical Yearbook. The social subjective factor is the government policy (GP), including the minimum purchase price policy for wheat and the temporary purchase and storage for maize, respectively (https://www.ndrc.gov.cn/, accessed on 15 December 2023). The minimum purchase price policy requires that wheat be purchased at the minimum purchase price, whereas the temporary purchase and storage are required by China Grain Reserve Management Group Co., Ltd. (Building 8, No.16, West Fourth Ring Road, Haidian District, Beijing, China) for temporary storage of maize. The dummy variables 0 and 1 were constructed to quantify the implementation of the policy (GP value was considered to be 1 or 0 when the policy was implemented or unimplemented, respectively). Since the data about PPI, FHT, and SMFI, etc. can only be found in the Rural Statistical Yearbook of China and the Chinese Statistical Yearbook over 2001–2020, the analysis was only carried out for 2001–2020.

Table 4. Information concerning the statistical data collection.

Statistical Data	Year	Sources
Crop-planted area	2001–2020	Weishan Irrigation District Administration in Liaocheng City
Average temperature	2001–2020	China Meteorological Data of National Meteorological Center http://data.cma.cn/, accessed on 1 January 2021
Precipitation	2001–2020	China Meteorological Data of National Meteorological Center http://data.cma.cn/, 1 January 2021
Producer price index of agricultural products	2001–2020	Chinese Statistical Yearbook https://www.stats.gov.cn/, accessed on 25 December 2023
Price index of farm hand tools	2001–2020	Rural Statistical Yearbook of China https://www.stats.gov.cn/, accessed on 25 December 2023
Price index of semi-mechanized farm implements	2001–2020	Rural Statistical Yearbook of China https://www.stats.gov.cn/, accessed on 25 December 2023
Price index of mechanized farm implements	2001–2020	Rural Statistical Yearbook of China https://www.stats.gov.cn/, accessed on 25 December 2023
Price index of chemical fertilizer	2001–2020	Rural Statistical Yearbook of China https://www.stats.gov.cn/, accessed on 25 December 2023
Price index of pesticides and pesticide instruments	2001–2020	Rural Statistical Yearbook of China https://www.stats.gov.cn/, accessed on 25 December 2023
Price index of multipurpose tractor oil	2001–2020	Rural Statistical Yearbook of China https://www.stats.gov.cn/, accessed on 25 December 2023
Government policy	2001–2020	National Development and Reform Commission, PRC https://www.ndrc.gov.cn/, accessed on 15 December 2023

provides a summary of the information concerning the above statistics.

2.3. Study Methods

2.3.1. Training Sample Migration Method

It is directly affected by the quantity and quality of the sample data as to the accuracy of the classification results. However, the traditional sample collection method (field surveys) cannot meet the requirements of long-term series planting structure identification of historical periods.

The training sample migration (TSM) method is performed by measuring the change in spectral angle between the same points in the goal year and the reference year using spectral angle mapping (SAM) [35] and obtaining unchanged samples. In this study, the sample sites collected in 2022 were used as the reference sites; the TSM method was used to obtain the high-quality sample sites over 2001–2021. To ensure enough high-quality reference samples were migrated, the threshold value was set to 0.2 (after several trial and errors) in this research. The high-quality samples were filtered according to the optimal threshold.

2.3.2. Classification Method

In this study, the random forest method (RF) was chosen to undertake land use and planting structure classification. This method performs well when dealing with high-dimensional data and large numbers of features by multiple decision trees to complete training and prediction. It has good generalization ability and resistance to overfitting [36]. In this study, the classification of land use and planting structure was carried out in GEE. When using the RF algorithm in GEE, the parameters to be set included the number of trees, variables per split, min leaf population, bag fraction, max nodes, and seed. The most important parameter is the number of trees. With an increase in the number of decision trees, the accuracy will also increase. The number of decision trees was 400, and the other parameters are the default. In order to avoid overfitting in the classification process, the number of training samples and test samples was 70% and 30% of the total sample, respectively.

The original bands that were utilized in this study include blue, green, red, near infrared (NIR), shortwave infrared 1 (SWIR-1), and shortwave infrared 2 (SWIR-2). The band parameter information of L5, L7, and L8 is shown in Table 3.

Moreover, the vegetation indexes are also used for distinguishing the different land use types to enhance classification accuracy, including NDVI, NDWI, the modified normalized difference water index (MNDWI) and the normalized difference built-up index (NDBI) [37,38,39]. In order to make the classification results more accurate, the original bands were combined with the vegetation index as the feature band’s scheme. It was used as an input index for RF and helped to improve the classification accuracy of land use and planting structures.

2.3.3. Econometric Regression Analysis

The regression analysis method was adopted to quantitatively analyze the driving factors affecting the planted area. This method is applied to describe the variables in relation to each other. In an econometric regression model, it is usually assumed that there is a linear relationship between the dependent and independent variables, and the parameters are estimated by the least-squares method so as to establish a mathematical model to explain the relationship between the variables. The calculation equation of the econometric regression model is as follows:

$$Y_t={\mathsf{\beta }}_0+{\mathsf{\beta }}_1{x}_{1,t}+{\mathsf{\beta }}_2{x}_{2,t}+\cdots +{\mathsf{\beta }}_9{x}_{9,t}+{\mathsf{\beta }}_{10}{x}_{10,t}+{\mathsf{\epsilon }}_t$$

(1)

where Y is the planted area of crops; β₀, β₁, β₂, …, β₁₀ are the parameters of the model; x₁, x₂, …, x₁₀ are the AT, Pre, GP, API, FHTI, SMFI, MFI, CFI, PAPI, and MPTOI, respectively; ε is the error term; and t is the year.

The goal of the model is to estimate the parameters β₁, β₂, …, β₁₀ to best fit the data and explain the relationship between the variables.

The regression coefficients of the respective variables were obtained by the least-squares method for all years. The least-squares method can be expressed as follows:

$$\mathsf{\beta }={(X^{\prime }·X)}^{-1}·X^{\prime }·Y$$

(2)

in which, β is the matrix of regression coefficient, $\mathsf{\beta }=\left(\begin{array}{l}{\mathsf{\beta }}_0\\ {\mathsf{\beta }}_1\\ ⋮\\ {\mathsf{\beta }}_{10}\end{array}\right)$. X is the vector of the independent variable, $X=\left(\begin{array}{l}1\\ ⋮\\ 1\end{array}\right.\left.\begin{array}{ccc}{x}_{1,2001}& \dots & x_{10,2001}\\ ⋮& \ddots & ⋮\\ x_{1,2020}& \cdots & x_{10,2020}\end{array}\right)$, and Y is the vector of the dependent variable $Y=\left(\begin{array}{l}{Y}_{2001}\\ Y_{2002}\\ ⋮\\ Y_{2020}\end{array}\right)$.

2.4. Classification Performance Criteria and Statistical Analysis

The accuracy evaluation was performed using overall accuracy (OA) and the kappa coefficient. OA is obtained by calculating the number of correctly classified samples as a proportion of all samples. The kappa coefficient represents the reduction in errors compared to a completely random classification. Three accuracy evaluations were conducted, and the average was taken as the final classification accuracy to reduce the randomness when extracting training samples.

Three statistical indices were used to evaluate the classification results, mean absolute error (MAE), root mean squared error (RMSE), and normalized root mean square error (NRMSE):

$$MAE=\frac{{\displaystyle \sum _{i=1}^k\left|P_i-O_i\right|}}{k}$$

(3)

$$RMSE=\sqrt{{\displaystyle \sum _{i=1}^k\frac{{(P_i-O_i)}^2}{k}}}$$

(4)

$$NRMSE=\frac{1}{O_{AVG}}\sqrt{{\displaystyle \sum _{i=1}^k\frac{{(P_i-O_i)}^2}{k}}}\times 100\%$$

(5)

where P_i is the inversion results of the i, O_i is the statistical findings of the i, k is the number of data samples, P_AVG is the average of the inversion results, and O_AVG is the average of the statistical findings.

The regression analysis method was applied to analyze the significant factors (p < 0.05) of the planting structure in this study. Stata 16.0 software was applied to conduct regression analyses.

3. Results and Discussion

3.1. Evaluation of Training Sample Migration Method

Figure 3a shows the evaluations of the sample retained rate for the winter wheat growing period over 2001–2022 in WID. The sample retention results show that the average retention rate from 2001 to 2021 was 90.48%, and the longer the period between the reference year and the goal year, the lower the retention rate of the sample. The lowest sample retention rate was only 77.84% in 2003. Figure 3b shows the evaluations of the sample retained rate for the summer maize growing period over 2001–2022 in WID. The sample retention results show that the average retention rate from 2001 to 2021 was 94.69%. The sample retention rate in the summer maize growth period was higher than that in the winter wheat growth period, which was higher by 4.66%. The lowest sample retention rate was only 82.77% in 2003.

The effect of crop classification can be affected by a variety of factors. The classification results are directly affected by the quality of the training samples. Fekri et al. (2021) found that the TSM method based on SAM can better obtain samples of the same type of features between different years [40]. In this study, we also obtained crop classification samples for 2001–2021 by the TSM method and passed the accuracy evaluation. According to this result, we found that the retention rate of the sample was affected by time and that as more time passes between the recommended year and the goal year, the retention rate of the sample decreases. The sample retention is affected by land use changes, and there is a significant decrease in sample retention when land use changes are large. The sample retention is also affected by the remote sensing images, and when the image quality of the target date is poor (e.g., too much cloud included, etc.) (Figure S2), the sample retention decreases substantially. However, in this study, we had screened the remote sensing images and then performed sample migration and crop classification work for years with high image quality. In our study, the TSM method has been demonstrated with high applicability in regions with relatively stable land use and planting structure, such as WID in the lower YRB, which had significant meaning for obtaining samples for long time classification in similar regions (i.e., where land use and planting structure are relatively stable).

3.2. Evaluation of the Classification Accuracy

Figure 4a shows the evaluations of the classification accuracy for the winter wheat growing period 2001–2022 in WID. The evaluation results show that the average OA from 2001 to 2022 was 0.936, and the OA was relatively stable between different years. Among them, the OA of 2021 was the highest, which was higher than that of 2011 (which had the lowest OA) by 7.47%. The results of kappa coefficients were similar to the OA. The average kappa coefficient was 0.953 from 2001 to 2022, and the kappa coefficient was relatively stable between different years. Among them, the kappa coefficient of 2021 was the highest, which was higher than that of 2011 (which had the lowest kappa coefficient) by 10.43%.

Figure 4b shows the evaluations of the classification accuracy for the summer maize growing period 2001–2022 in WID. The evaluation results show that the average OA from 2001 to 2022 was 0.941, and the OA was relatively stable between different years. Among them, the OA of 2021 was the highest, which was higher than that of 2011 (which had the lowest OA) by 8.72%. The results of kappa coefficients were different from the OA. The average kappa coefficient was 0.951 from 2001–2022, and the kappa coefficient showed a significant decrease from 2015–2017. Among them, the kappa coefficient of 2021 was the highest, which was higher than that of 2016 (which had the lowest kappa coefficient) by 10.72%. The kappa coefficient for the 2016 crop classification was lower, but it was still within the acceptable range.

Figure 5a indicates the comparison between inversion results and statistical findings of winter wheat from 2001 to 2020. In this study, the inversion result of winter wheat was better, and the statistical evaluation indices were MAE = 11.82, RMSE = 21.99 × 10³ ha, and NRMSE = 5.42%. The average relative error between inversion results and statistical findings of the winter wheat-planted area in WID during the past 20 years (2001–2020) was 2.75%. As shown in Figure 5b, the inversion result of the summer maize-planted area was close to the statistical findings, and the statistical evaluation indices were MAE = 14.23, RMSE = 20.60 × 10³ ha, and NRMSE = 5.75%. The average relative error between inversion results and statistical findings of summer maize-planted area in WID during the past 20 years (2001–2020) was 4.56%.

The WMR is the main crop cultivation pattern in the lower YRB [41]. Winter wheat and summer maize are the crucial grain crops in northern China and the two largest food crops grown within WID. The inversion of the planted area for the two crops is particularly critical. By comparing the planting area of major crops obtained by inversion results with the statistical data, we found that the crop classification method and training samples used in this study can obtain the planted area of crops more accurately. The inversion results of the crop-planted area are influenced by several factors, including the classification method, the selection and representativeness of the sample collection, the pixel scale and spatial distribution, etc. [42,43]. Since the crop-planting structure in the study site was relatively simple, it was only necessary to differentiate between crops with similar growing periods based on their growing periods, including winter wheat and garlic, winter wheat and cotton, and cotton and summer maize (Table 1). Then relatively accurate planted structures and planted areas of crops can be obtained.

Identification accuracy is influenced by many factors, such as the classification method, image quality, and sample quality and quantity. In this research, the average accuracy of crop classification over 2001–2022 was 93.85%, which demonstrated the good performance of the RF method in WID. Similar conclusions were reached by other researchers. For example, Niu et al. (2022) have found that there was a significant improvement in identification accuracy by using the RF method with a combination of multiple feature bands [44]. Luo et al. (2023) have demonstrated that the RF classifier could effectively use Sentinel-2 data to map the planting structure in the North China Plain [16]. However, the long-term land use and planting structure have not been identified by a machine learning algorithm (such as RF methods). In our study, the RF algorithm proved to be useful for long-time land use and planting structure classification in the lower YRB with high accuracy. In addition, the image quality was also be reported as one of the main factors affecting the accuracy of crop classification [45]. In addition, poor image quality (including missing images, severe banding, and excessive image cloudiness) could directly reduce the crop classification accuracy. In this research, the quality of the L5 and L7 images of 2011 was lower compared with those of the other years, in which the L5 image was seriously missing and the L7 image banding was missing. These then led to lower crop classification accuracy. For crop classification after 2015, the images of L8 were used in this research. The image quality of L8 was significantly improved compared to the other two qualities, so the accuracies after 2015 were higher than those in 2001–2011. In this study, the long-time and high-quality training samples were obtained by the TSM method, which overcomes the shortcomings of time-consuming and low-quality traditional sample collection methods. Compared with other classification algorithms, the RF algorithm has the advantages of high accuracy and stability. Therefore, there was higher accuracy in identifying long-time land use and planting structures in the lower YRB. This had a significant meaning for identifying long time series of land use and planting structures in similar regions.

3.3. Spatial and Temporal Variation in the Land Use and Planting Structure in the Irrigation District

Figure 6 indicates the spatial distribution of the land use and planting structure in WID over 2001–2022. From the figures, winter wheat and summer maize are mainly planted in the eastern and southern parts of the irrigation district. Cotton was mainly planted in the northern part of the irrigation district until 2010, and then the area of cotton planting was significantly reduced. The area planted with cotton had declined by 96.36%. Forest land was mainly distributed in the west of WID. The construction land and forest land were basically stable after 2015. The area of construction land increased in the central part of the irrigation area after 2010.

Figure 7a indicates the inversion of the planted area of winter wheat from 2001 to 2022. According to the inversion results, the planted area of WID had increased from 2001 to 2022, with an increase rate of 4.74 × 10³ ha per year. In the last 22 years, there has been a massive increase in winter wheat planting, where the maximum value of area was 325.57 × 10³ ha (2022) and the minimum value was 217.53 × 10³ ha (2001), which was a 49.66% increase. Figure 7b indicates the inversion results of the summer maize-planted area from 2001 to 2022. According to the results, over the past 22 years (2001–2022), there has been an increasing trend in the area under summer maize planting in WID. The increasing rate was 8.57 × 10³ ha per year. In the first decade (2001–2010), the summer maize followed a visible growth trend, with a rapid increase of 59.19%. In the last 22 years, the maximum value of area was 300.50 × 10³ ha (2022), and the minimum value was 174.28 × 10³ ha (2001), which was a 91.11% increase.

Figure 8 reveals the land use and planting structure change in WID for the period 2001–2022. According to the results of the land use and planting structure in 2001, the cultivated land was the largest, accounting for 45.72% of the total area. The WMR-planting area accounts for 74.84% of the cultivated land area, and the cotton planting area accounts for 16.27% of the cultivated land area. According to the results of the land use and planting structure in 2022, the cultivated land area accounted for 58.12% of the total area. The WMR-planting area accounted for 88.11% of the cultivated land area, and the cotton planting area accounted for 0.50% of the cultivated land area. Over the past 22 years, there has been an upward trend in the area of WMR. The construction land and water remained relatively stable. The planted area of cotton was relatively stable from 2001 to 2010, while it declined significantly after 2010. About 50.05% of the cotton-growing area was converted into the WMR-planting area during the 22 years.

In previous studies, researchers have focused on land use change or mapping the distribution of single crops, such as winter wheat, in the lower YRB [16,46]. There have been few studies analyzing the changes of planting structure. It will provide implications for the reasonable allocation of agricultural water resources in WID by clarifying the planting structure change. According to the results from analyzing the spatial and temporal variation in land use, we found that the land use changes in the northern part of the irrigation area were larger. This is probably attributed to the lower soil nutrients in the northern part of the irrigation area from 2001 to 2010. The content of soil organic matter in the north of Liaocheng City (which is the north of the WID) was lower than that in other areas in 2005 [47]. This means that soils in this region were not in the best growing environment for grain crops. According to the 12th Five-Year Plan of Water Conservancy Development in Liaocheng City (Liaocheng Municipal Water Resources Bureau, 2011, http://slj.liaocheng.gov.cn, accessed on 30 December 2023), the government had planned the project to divert the Yellow River to Taochengpu. This would solve the serious water shortage problem in the west and north of Liaocheng and alleviate the water supply pressure of the west trunk canal in WID. Before 2015, WID had completed the lining works of 37 main and branch canals with 378.1 km, 870 supporting buildings, 7333.33 ha of field water-saving irrigation, management engineering, and information construction, etc. With the construction and improvement of water conservancy projects in WID, irrigation and drainage conditions and the soil environment have been improved [48]. For example, soils in the northern part of the irrigation districts have been significantly improved over the past 22 years [47,49]. Moreover, part of the natural land was changed to farmland over the past 22 years, which increased the area of cultivated land in WID. These changes will affect the allocation of irrigation water resources.

3.4. The Driving Factors of the Variation in Crop Planted Area

Table 5. Baseline regression results of influencing factors on the historical planted area of main crops.

Factors	Winter Wheat		Summer Maize
	Coefficient	Robust	Coefficient	Robust
Natural factor a
AT	12.91	15.08	21.48	14.21
Pre	−0.01	0.04	−0.01	0.08
Social subjective factor b
GP	63.32 ***	17.40	63.98 **	20.26
Social objective factors c
API	1.24	0.88	1.00	0.66
FHTI	−0.81	2.13	−1.60	2.68
SMFI	1.64	5.42	−2.13	4.17
MFI	−1.93	5.20	3.53	5.20
CFI	0.13	0.60	0.04	1.07
PAPI	0.88	2.61	0.34	1.93
MPTOI	0.08	0.80	0.50	2.90
N	20		20
R2	0.79		0.87

^a AT means average temperature and Pre means precipitation. ^b GP means government policy: the government policy for winter wheat and the government policy for summer maize. ^c API means producer price index of agricultural products, FHTI means farm hand tools, SMFI means semi-mechanized farm implements, MFI means mechanized farm implements, CFI means chemical fertilizer, PAPI means pesticides and pesticide instruments, and MPTOI means multipurpose tractor oil. ** indicates significant differences at the 0.05 level, and *** indicates significant differences at the 0.01 level.

shows the baseline regression results of influencing factors on historical planted areas of main crops from 2001 to 2020. The results show that the GP significantly affected the planted area of winter wheat (p < 0.01), and the GP also significantly affected the planted area of summer maize (p < 0.05). When government policies that have a positive impact on crop production were implemented, the planted area of winter wheat and summer maize increased by 63.32 × 10³ ha and 63.98 × 10³ ha, respectively. However, the natural factors (AT and Pre) and the social objective factors (PPI, FHT, and SMFI, etc.) had no significant effect on the planted area of winter wheat and summer maize.

Winter wheat and summer maize are important grain crops. It is of great significance to study the change in their planting area and the corresponding driving factors. In previous studies, researchers mostly focused on the influence of natural factors (i.e., temperature, precipitation) and social factors (i.e., price and cost) on a crop-planting area. Only a few focused on the effect of government policies on a crop-planting area. Besides the natural and social factors, the government policies and their impacts on crop-planting areas were also considered in our current research. It was found that government policies were the major driving factors for promoting agricultural production. For example, the minimum purchase price policy for wheat had been implemented by the government (i.e., Henan and Shandong provinces) to ensure the grain crop-planting area from 2006 [50]. Moreover, in order to stabilize the maize market and solve the problem of falling prices in China, the government decided to implement a temporary policy of maize purchase and storage [51]. Therefore, the influence of government agency polices on land use and planting structure need to be further considered in the future.

4. Conclusions

In order to investigate the long-term land use and planting structure and its driving factors in the lower YRB, samples were collected from June to July 2022, and the 2001–2021 samples were obtained through the TSM method in GEE. The RF method was used to identify the land use and planting structure over 2001–2022. The driving factors that may affect the changes in winter wheat- and summer maize-planting areas were collected in WID. The econometric regression model can analyze the influence of driving factors (i.e., temperature, precipitation, prices and government policies, etc.) on the planting structure. The main conclusions are as follows:

(1)

The TSM method combined with the RF method has high classification accuracy in identifying historical periods of land use and planting structures in the lower YRB, and the average accuracy of classification over 2001–2022 was 93.85%. This classification method helps to map the long-term land use and planting structure in the lower YRB. It is critical to achieve high-quality development in the region;

(2)

The WMR-planted area has increased over the past 22 years; the rates of increase for winter wheat and summer maize were 4.74 × 10³ ha and 8.57 × 10³ ha per year, respectively. The cotton-planted area has decreased in WID;

(3)

Government policies were the driving factors affecting the change in crops planted in WID. The natural and social objective factors also affected the planted area of winter wheat and summer maize, but the effect was not significant.

The findings are expected to provide implications for the reasonable allocation of agricultural water resources in the irrigation district of the lower YRB and provide a basic dataset for future studies of the agricultural water cycle in the region.