Research on Machine Learning-Based Extraction and Classification of Crop Planting Information in Arid Irrigated Areas Using Sentinel-1 and Sentinel-2 Time-Series Data

Abstract

Irrigation areas in arid regions are vital production areas for grain and cash crops worldwide. Grasping the temporal and spatial evolution of planting configurations across several years is crucial for effective regional agricultural and resource management. In view of problems such as insufficient optical images caused by cloudy weather in arid regions and the unclear spatiotemporal evolution patterns of the planting structures in irrigation areas over the years, in this study, we took the Santun River Irrigation Area, a typical arid region in Xinjiang, China, as an example. By leveraging long time-series remote sensing images from Sentinel-1 and Sentinel-2, the spectral, index, texture, and polarization features of the ground objects in the study area were extracted. When analyzing the index characteristics, we considered several widely used global vegetation indices, including the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), and Global Environment Monitoring Index (GEMI). Additionally, we integrated the vertical–vertical and vertical–horizontal polarization data obtained from synthetic aperture radar (SAR) satellite systems. Machine learning algorithms, including the random forest algorithm (RF), Classification and Regression Trees (CART), and Support Vector Machines (SVM), were employed for planting structure classification. The optimal classification model selected was subjected to inter-annual transfer to obtain the planting structures over multiple years. The research findings are as follows: (1) The RF classification algorithm outperforms CART and SVM algorithms in terms of classification accuracy, achieving an overall accuracy (OA) of 0.84 and a kappa coefficient of 0.805. (2) The cropland area classified by the RF algorithm exhibited a high degree of consistency with statistical yearbook data (R² = 0.82–0.91). Significant differences are observed in the estimated planting areas of cotton, maize, tomatoes, and wheat, while differences in other crops are not statistically significant. (3) From 2019 to 2024, cotton remained the dominant crop, although its proportional area fluctuated considerably, while the areas of maize and wheat tended to remain stable, and those of tomato and melon showed relatively minor changes. Overall, the region demonstrates a cotton-dominated, stable cropping structure for other crops. The newly developed framework exhibits exceptional precision in categorization while maintaining impressive adaptability, offering crucial insights for optimizing agricultural operations and sustainable resource allocation in irrigation-dependent arid zones.

1. Introduction

In parched landscapes, water is a scarce and invaluable commodity. Therefore, it is absolutely critical to manage water resources wisely in irrigated agricultural zones if we want to ensure that agriculture can thrive in the long run [1]. By accurately extracting the crop planting structure in irrigated areas in arid regions, we can obtain detailed information about the types and planting areas of crops in each region. This not only helps to formulate more scientific irrigation plans, thereby avoiding the waste of precious water resources, but also enables reasonable assessment of the current planting structure, identification of existing problems, and the adoption of corresponding measures for optimization and adjustment, thus improving the agricultural production efficiency and resource utilization efficiency [2,3]. In addition, the food production potential and risks in irrigated areas in arid regions, as important global food production bases, can be predicted and evaluated through the extraction of crop planting structures, providing a solid guarantee for global food security [4]. Therefore, identifying crop planting patterns in arid irrigation zones is crucial for advancing sustainable farming and securing global food supplies.

Unlike conventional approaches, remote sensing has become a turning point in modern agriculture for mapping crop distribution across vast areas. Its ability to cover extensive regions quickly and cost-effectively makes it the primary solution for gathering spatial data on farmland usage [5]. Among them, multispectral remote sensing imagery possesses the capability for fine-scale classification. By comparing multispectral images with crop spectral library data, different crops within farmland can be classified [6,7]. For example, Liu et al. [8] employed Sentinel-2 remote sensing imagery to identify winter wheat in northern Henan Province, China, by generating spectral characteristics and vegetation indices at different growth stages. Qin et al. [9] established a method for extracting the cropping structure of maize at different growth stages by using time-series spectral index data derived from Sentinel-2 imagery. However, multispectral satellite monitoring is affected by lighting and weather conditions, leading to degraded image quality and data loss, especially in long-term time-series observations where frequent cloudy weather exacerbates data gaps, increasing the difficulty of crop classification. Compared with multispectral sensors, Synthetic Aperture Radar (SAR) has an all-weather capability, overcoming the limitations of cloud cover and illumination, enabling continuous monitoring under cloudy and adverse weather conditions, and significantly improving the accuracy and reliability of crop identification. Edyta et al. [10] utilized time-series Sentinel-1 imagery to characterize crop phenological stages and proposed a crop classification method based on radar polarization techniques. Pandžić et al. [11] conducted a classification of nine different crop types using time-series Sentinel-1 data for evaluation. Although SAR satellites have an all-weather advantage in crop classification in arid regions, they are limited by the lack of spectral information and rely solely on backscattering characteristics. As a result, surface cover types such as bare rocks and sandy soils can easily be confused with crops, leading to classification errors.

The integration of multi-source remote sensing and machine learning supports precision agricultural monitoring, particularly excelling in crop classification under complex environmental conditions. The combined application of Sentinel-1 SAR data and Sentinel-2 multispectral data has become mainstream [12,13,14]. Pageot et al. [15] successfully distinguished between irrigated and rainfed crops in southwestern France based on Sentinel-1 and Sentinel-2 data. Tang et al. [16] integrated Sentinel-2 imagery with time-series Sentinel-1 SAR features and, using RF, achieved high-accuracy extraction of citrus plantation areas in southern Ganzhou. Sun et al. [17] further investigated the applicability of Sentinel-1 and Sentinel-2 data for crop classification in farmland plots located in cloudy and rainy mountainous areas. However, most existing studies focus on humid or topographically homogeneous regions, and there remain significant limitations in the exploration of crop classification in irrigated areas of arid regions. In addition, research targeting such regions generally faces the following challenges: (1) Cloudy weather and sparse vegetation coverage result in the insufficient continuity of optical data, necessitating deep integration of SAR and optical data. (2) The high degree of spectral similarity among various crop types (e.g., tomato, melon, and sugar beet) requires the integration of multidimensional features such as phenology, texture, and morphology to enhance separability. (3) The performance differences of various machine learning algorithms in multi-crop scenarios in arid areas have not been systematically evaluated, and the applicability of mainstream algorithms such as RF, CART, and SVM urgently needs to be validated. (4) Limited by data continuity, long-term time-series dynamic monitoring studies of cropping patterns in arid regions are particularly scarce, making it difficult to reveal the interannual variation patterns of planting structures.

To address the aforementioned issues, this study takes the Santun River Irrigation District in Xinjiang, China, as a representative case of arid regions, integrating time-series Sentinel-1/2 data and proposing a crop classification framework based on multi-feature fusion and algorithm optimization. Firstly, continuous six-year (2019–2024) datasets of backscattering coefficients (VV, VH) and multiple vegetation indices (NDVI, EVI, SAVI, GEMI) were constructed based on the Savitzky–Golay filtering technique. Simultaneously, phenological features, spectral responses, texture, and shape features were integrated, and a sample augmentation strategy was introduced to address the scarcity of training data. Secondly, the classification performance of three algorithms—RF, CART, and SVM—was systematically compared to select the optimal model for the high-accuracy identification of cotton, maize, wheat, and characteristic orchard crops. Finally, the spatiotemporal evolution patterns of cropping structures in the irrigation district were analyzed, revealing the adaptive mechanisms of agriculture to climate change in arid regions. This study not only provides a theoretical basis for multi-algorithm optimization in the remote sensing monitoring of crops in irrigated areas of arid regions, but its long-term classification results also offer valuable data support for regional agricultural structure adjustment and optimized water resource allocation, contributing practical significance toward achieving the United Nations Sustainable Development Goals of food security and ecological sustainability.

2. Materials and Methods

2.1. Overview of the Study Area

The Santun River Irrigation District (86°24′33″–87°37′ E, 43°6′30″–45°20′ N) is located in the central section of the northern foot of the Tianshan Mountains in Xinjiang, on the southern margin of the Junggar Basin. Its geographic location ranges from the north latitude. It borders the Toutunhe Irrigation District to the east, adjoins the Hutubi River Irrigation District to the west, is bound by Askedaban Mountain in the Tianshan Mountains and Hejing County to the south, and is adjacent to the Gurbantünggüt Desert, Hoboksar County in the Tacheng Prefecture, and Fuhai County in the Altay Prefecture to the north. The district extends 260 km from north to south, is approximately 31 km wide from east to west, and has a drainage area of 4466 km².

The Santun River Basin is located in the hinterland of Eurasia, on the southern margin of the Junggar Basin, far from the ocean, and has a mid-temperate continental arid climate. Influenced by the temperate weather system and the cold air from the Arctic Ocean, cold air accumulation occurs in winter, and the basin has a heat-gathering effect in summer. The annual, monthly, and daily temperature differences vary significantly, exhibiting typical continental climate characteristics.

The planting structure in the Santun River Irrigation District is mainly dry-land farming, including crops such as wheat, corn, cotton, tomatoes, vegetables, melons, orchards, and alfalfa. Cotton is the main cash crop in this area, while corn and wheat are the main food crops. The study area, an arid zone with diverse crops, provides an ideal research setting. Its location is depicted in Figure 1.

2.2. Data Sources

2.2.1. Sentinel-1 Remote Sensing Image Data

The Sentinel-1 radar imaging system comprises two polar-orbiting satellites (A and B), offering 6-day revisit cycles. The sensor it carries is a C-band-based SAR. The radar imaging system is divided into four scanning imaging modes, namely, interferometric wide swath (IW), extra-wide swath (EW), wave (WV), and StripMap (SM) modes [18]. The study’s data were sourced from the European Space Agency’s official website (https://dataspace.copernicus.eu/explore-data/data-collections/sentinel-data/sentinel-1, accessed on 16 February 2025), covering the time period from 2019 to 2024. The data were collected in IW mode using Level-1 GRD products, featuring 10 m resolution and dual-polarization (VV/VH). The specific acquisition dates of Sentinel-1 SAR imagery used in this study are detailed in

Table 1. Sentinel-1 image acquisition dates.

Year	Month	Day	Year	Month	Day
2019	March	16/21/28	2022	March	10/17/24/31
	April	2/9/14/21/26		April	5/15/17/22/29
	May	3/8/15/20/27		May	6/11/18/23/28
	June	1/8/13/20/25		June	2/9/16/21/28
	July	2/7/14/19/26/31		July	3/8/15/22
	August	7/12/19/24/31		August	3/10/15/22/27
	September	5/12/17/24/29		September	1/6/13/18/25/30
	October	6/11/18/23/30		October	2/7/14/21/26/31
	November	4		November	2
2020	March	22/27	2023	March	7/19/24/31
	April	3/8/15/20/27		April	5/12/17/24/29
	May	9/14/21/26		May	6/11/18/23/30
	June	2/7/14/19/26		June	4/16/23/28
	July	1/8/13/20/25		July	5/10/17/22/29
	August	1/6/13/18/25/30		August	3/10/15/22/27
	September	6/11/18/23/28/30		September	3/8/15/20/27
	October	5/12/17/24/29		October	2/9/14/21/26
	November	3/5		November	2
2021	March	27/29	2024	March	30
	April	3/8/15/20/27		April	6/18/23/30
	May	2/9/14/21/26		May	5/12/17/24/29
	June	2/7/14/19/26		June	5/10/17/22/29
	July	1/8/13/20/25		July	4/11/16/23/28
	August	1/6/13/18/25		August	4/9/16/21/28
	September	6/11/18/23/30		September	2/9/14/21/26
	October	5/12/17/24/29		October	8/15/20/27
	November	5		November	1

. The data span a long temporal range each year (from March to November), totaling 233 observation days, with consistent spatial resolution, providing a reliable time-series observation basis for subsequent analysis.

2.2.2. Sentinel-2 Remote Sensing Image Data

The Sentinel-2 satellite, part of the European Space Agency’s (ESA) Global Monitoring for Environment and Security initiative, represents a cutting-edge advancement in Earth observation technology. Launched on 23 June 2015, the Sentinel-2A spacecraft operates as a medium-resolution multispectral imaging platform. Fitted with a sophisticated Multispectral Instrument (MSI), it orbits at an altitude of 786 km. The satellite boasts an impressive array of 13 spectral bands, offering spatial resolution ranging from 10 to 60 m, while capturing imagery across a generous 290 km-wide swath. This advanced system provides invaluable data for environmental monitoring and security applications [19].

The Level-2A imagery used in this study underwent preprocessing including radiometric calibration, orthorectification, and atmospheric correction. The acquisition period spans the vegetation growing seasons (March to November) from 2019 to 2024, and the data were downloaded from the Copernicus Data Space Platform of ESA (https://dataspace.copernicus.eu/explore-data/data-collections/sentinel-data/sentinel-2, accessed on 16 February 2025). For each temporal phase, three scenes were required to fully cover the study area. A total of 156 valid time-series images with cloud cover of less than 5% were selected, with good data quality and a spatial reference system set to GCS-WGS-84. The original Sentinel-2 data provided only 52 effective observation days during the vegetation growing season, with a time-series completeness ratio of 3.07%. After integration with Sentinel-1 data, the number of effective observation days increased to 233, representing an increase of 348.08%, and the time-series completeness improved to 20.23%. The multi-source collaborative observation improved the spatiotemporal continuity by 560%, significantly enhancing the coverage of the constructed phenological time-series dataset, thereby providing a highly reliable data foundation for continuous monitoring of surface processes. The acquisition dates of the Sentinel-2 data used in this study are listed in

Table 2. Sentinel-2 image acquisition dates.

Year	Date	Year	Date	Year	Date
2019	19 April 28 June 28 July 27 August 21 September 6 October 16 October 31 October	2021	13 April 23 May 2 July 1 August 11 August 31 August 5 September 15 September 25 September 5 October 10 October	2023	24 March 28 April 18 May 27 June 12 July 16 August 5 October 25 October
2020	8 April 28 April 8 May 28 May 2 June 22 June 17 July 16 August 31 August 30 September 5 October 15 October 30 October	2022	18 April 18 May 22 July 5 September 15 November	2024	18 March 22 April 6 June 21 June 5 August 4 September 19 September

.

2.2.3. Sample Data

(1) Field survey data

In this study, we conducted a systematic field investigation in the study area during the crop growth and maturation seasons from April to November 2024. This included accurately positioning sample points using a global positioning system (GPS) and obtaining ground information through unmanned aerial vehicle (UAV) imagery to collect sample data for 2024. In addition, high-resolution imagery from Google Earth and Amap was utilized to select samples following the principle of stratified random sampling.

Considering the actual spread of the diverse crops and their corresponding growth stages within the study region, we meticulously examined the spectral patterns and the distinctive indices of the 2024 field sample points relative to other historical years. Afterward, visual interpretation markers were scattered randomly across the study area. Figure 2a illustrates their spatial distribution in 2024, while Figure 2b displays the corresponding sample counts. For analytical purposes, the dataset was split randomly into training and validation subsets using a 70/30 ratio.

(2) Interannual migration of sample points

Based on the characteristic information of various crops obtained from the ground survey sample points in 2024, the time-series feature curves corresponding to different phenological stages were used as a reference. A systematic comparative analysis was then conducted on crop information from randomly selected farmland plots during 2019–2023. By comparing the feature set curves, shape features, and texture features of farmland crops across different years, the similarities and differences in spectral and texture characteristics of randomly generated points were analyzed, thereby identifying the crop sample points for the years 2019–2023. The distribution of these sample points is illustrated in Figure 3.

2.2.4. Statistical Data

The statistical data for the crop planting areas in the Santun River Irrigation District from 2019 to 2024 (a six-year period) were collected from the Watershed Management Office of the Santun River Irrigation District in Changji City (https://www.xjsth.cn, accessed on 16 February 2025). In addition, to ensure the comprehensiveness and accuracy of the data, the data for the sown areas of the major crops in the Xinjiang region and the crop planting structures during the study period provided by the Bureau of Statistics of Changji Prefecture (https://www.cj.gov.cn/, accessed on 16 February 2025) were referred to (https://tjj.xinjiang.gov.cn/tjj/nyypu/list_nj1.shtml, accessed on 16 February 2025). In addition, the annual reports on government information disclosure published on the official website of the Changji City People’s Government (https://www.cj.gov.cn/, accessed on 16 February 2025) (https://www.cj.gov.cn/p1/zfxxgknb.html, accessed on 16 February 2025) were consulted to determine the policy orientation for use in analyzing the causes of the crop planting structure and for data verification.

2.3. Software Used in This Study

In this study, a variety of professional software and algorithms were employed for the pre-processing and subsequent analysis of the remote sensing images to ensure the efficiency and accuracy of the data processing.

Preprocessing of Sentinel-1 data was performed using the SNAP software 12.0.0 (https://step.esa.int/main/download/snap-download/, accessed on 16 February 2025) provided by the European Space Agency (ESA). The Sentinel-1 data were processed through a multi-step workflow: orbital correction was first conducted, followed by noise removal via filtering algorithms. Subsequently, radiometric calibration was completed, and then filtering processing was applied to smooth the data. A DEM-based terrain correction was applied, following which, the data transitioned to a dB measurement scale. Finally, mosaicking and cropping were carried out. For Sentinel-2 data, the resampling process was also completed in the SNAP software, followed by band fusion, mosaicking, and cropping of the data using the ENVI 5.6 remote sensing image processing platform.

During the multiscale segmentation stage of the images, the eCognition software (https://geospatial.trimble.com/zh-cn/products/software/trimble-ecognition, accessed on 16 February 2025) was selected, and the multiresolution segmentation algorithm within it was invoked for object-based segmentation. Regarding the selection of the classification algorithms, multiple machine learning algorithms were used for training and application. RF, CART, and SVM in the classifier algorithm were used. By training these algorithms, the research team was able to achieve the automatic classification of the remote sensing images, thereby improving the efficiency and accuracy of the classification. Finally, to verify the accuracy of the classification results, the error matrix was selected in the statistic type for verification.

In addition, in this study, the fresh extension package of Python 3.10 was used to calculate the characteristic parameters of the time-series curves of the backscattering coefficients in the VH and VV polarization modes.

2.4. Research Methods

The data used in this study were categorized into two types: multispectral Sentinel-2 data and Sentinel-1 SAR data. The Sentinel-1 SAR data were utilized to compensate for missing optical observations during critical crop growth periods in the study area. Initially, an object-based approach was applied to conduct multi-scale segmentation based on the optimal segmentation scale. Subsequently, Sentinel-1/2 data from 2024, combined with the phenological characteristics of various crops within the irrigation district, were used to build a dataset comprising shape, texture, and index features. Time-series curves of multiple vegetation indices and backscattering coefficients were extracted for different crops and smoothed using the Savitzky–Golay (S–G) filter. Three classification algorithms—CART, SVM, and RF—were then applied, and the optimal classification model was selected based on accuracy assessment via confusion matrices. Finally, the selected model was transferred interannually to acquire crop planting structure distribution information in the study area from 2019 to 2023. The technical framework of this study is presented in Figure 4.

2.4.1. Multiscale Segmentation and Optimal Segmentation Scale Preference

Multiscale image segmentation is of great significance in the object-oriented classification method. The method’s performance depends on multiple variables, including band count, shape settings, and scaling factors—any of which can throw off segmentation precision, resulting in outcomes that are either too coarse or overly refined. To pinpoint the perfect segmentation scale, one typically has to delve deep into the attributes of the land features and the distinctive qualities of the imagery within the research domain, fine-tuning the outcomes through a series of iterative tests and simulations [20].

The ESP tool, a region-merging method, derives optimal segmentation parameters through iterative processing. This research employed the ESP tool to assess optimal segmentation parameters for the study area’s imagery. During the iteration process, the shape heterogeneity weight (shape) and compactness heterogeneity weight (compactness) were varied from 0.1 to 0.9 in increments of 0.1, and multiple segmentation tests were conducted using all possible value combinations. When determining the optimal segmentation scale, based on the local variance (LV) of the homogeneity of the segmented objects and its rate of change (ROC), the optimal segmentation scale can be more accurately identified and selected, thereby improving the accuracy and reliability of land use information extraction [21]. The calculation formulas for the LV and ROC are as follows:

$$LV={\displaystyle \frac{1}{m}}\times {\displaystyle \sum _1^m{\left(C_L-{\overline{C}}_L\right)}^2},$$

(1)

$$ROC={\displaystyle \frac{L_i-L_{i-1}}{L_{i-1}}}\times 100\%,$$

(2)

where C_L is the brightness value of a single image in the Lth band; ${\overline{C}}_L$ is the average brightness value of all of the objects in the image in the Lth band; m is the total number of image objects; ROC is the rate of change of the LV (%); L_i is the average standard deviation of the ith object layer in the target layer; and L_i-1 is the average standard deviation of object layer i-1 in the target layer.

2.4.2. Feature Dataset Construction

(1) Climate Characteristics

Based on the results of the field exploration and investigation, the phenological calendars of the seven major crop types in the study area, namely, cotton, corn, wheat, tomatoes, sugar beets, melons, and seed gourds, were established. Through systematic data collection and analysis, a deep understanding of the phenological information of the different crops was obtained. This information is presented in an intuitive way in Figure 5.

Within the study area, the seven major crops generally exhibited the most vigorous growth from August to September each year. This is a crucial stage in their life cycle and a period with good spectral reflectance. After the spring sowing and summer growth processes, the harvesting was usually basically completed before November—i.e., when winter began—thus preparing for a new round of farming in the coming year.

(2) Characteristic Variable Set

To fine-tune the precision of identifying various crops’ planting patterns in the area under investigation, we employed VV and VH polarimetric characteristics from the Sentinel-1 imagery, alongside the spectral, index, texture, and geometric attributes from the Sentinel-2 images. This integration led to a comprehensive set of 29 distinct features. Polarization features are important and unique attributes of radar data and can be used to obtain information about ground objects. Different ground objects contain different information under different polarization modes [22,23]. Therefore, the VV and VH polarization features of the Sentinel-1 images were introduced to complete the feature set.

Texture features are another important attribute for obtaining information about ground objects from images. Different ground objects have unique textures. Image texture features were computed using eCognition’s GLCM-based texture analysis function. This characteristic describes the gray-level correlation between two pixels separated by a fixed distance in the image space. To avoid redundancy due to excessive texture features when training the classification model, four texture features, namely, homogeneity, dissimilarity, autocorrelation, and entropy, were selected. The detailed calculation formulas are presented in

Table 3. Texture feature calculation formulas in the gray-level co-occurrence matrix (GLCM).

Texture Feature	Calculation Formulas
Homogeneity	${\displaystyle \sum _i{\displaystyle \sum _{j}p}}(i,j)\times {\displaystyle \frac{1}{1+{(i-j)}^2}}$
Dissimilarity	${\displaystyle \sum _i{\displaystyle \sum _{j}p}}(i,j)\times |i-j|$
Entropy	${\displaystyle \sum _i{\displaystyle \sum _{j}p}}(i,j)\times \ln p(i,j)$
Correlation	${\displaystyle \sum _i{\displaystyle \sum _j\frac{(i- \mathrm{Mean} )\times (j- \mathrm{Mean} )\times p{(i,j)}^2}{ \mathrm{Variance} }}}$

Note: in the gray-level co-occurrence matrix, P(i, j) represents the frequency with which two pixels, having gray levels i and j, occur adjacent to each other in image p.

.

Spectral characteristics serve as the fundamental criteria for identifying and categorizing diverse ground objects within remote sensing imagery [24]. Thus, in this study, 12 bands of Sentinel-2 data were used as the original bands, and the EVI, NDVI, and SAVI were introduced. In addition, red-edge bands that help distinguish different crops were added, including the NDVI red-edge 1 (NDVIre1), NDVI red-edge 2 (NDVIre2), and NDVI red-edge 3 (NDVIre3). To achieve applicability to different arid and semi-arid regions globally, in this study, we also introduced the GEMI for reference and comparison [25].

Since Sentinel-2 remote sensing images with a resolution of 10 m, which is a relatively high resolution, were used to extract more image information, two shape features, i.e., compactness and density, were used to jointly construct the dataset. The detailed feature dataset is shown in

Table 4. Classification feature set.

Feature Name	Characteristic Variable	Description/Calculation Formula
Spectral band	B1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B10, B11, B12	Sentinel-2 original full-band
Polarization characteristics	VV, VH	Sentinel-1 raw polarization characteristics
Shape characteristics	Compactness	The tightness feature describes the compactness of an image object
	Density	Density features describe the distribution of image objects in pixel space
Texture	Homogeneity	Reflects the consistency or smooth texture in the image. A higher homogeneity index means that the texture of an image is more evenly distributed in space and the details change less
	Dissimilarity	Used to measure the dissimilarity of textures in an image. Reflects the contrast or grayscale difference between pairs of pixels in an image
	Entropy	A property that reflects the organizational arrangement of an object’s surface and is used as a measure of the degree of uncertainty or confusion at the pixel gray level in an image
	Correlation	A statistic that describes the degree of correlation of gray levels between pixels and measures the linear correlation between the gray levels of pixels in an image, i.e., the trend of gray levels
Vegetation index	SAVI	Soil-Adjusted Vegetation Index (B8 − B4) $\times$ (1 + L)/(B8 + B4 + L)
	EVI	Enhanced Vegetation Index 2.5 $\times$ (B8 − B3)/(B8 + 6 $\times$ B4 − 7.5 $\times$ B2 + 1)
	NDVI	Normalized Difference Vegetation Index (B8 − B3)/(B8 + B3)
	GEMI	Global Environment Monitoring Index eta $\times$ (1 − 0.25 $\times$ eta) − ((B4 − 0.125)/(1 − B4)) eta = (2 $\times$ (B82 − B42) + 1.5 $\times$ B8 + 0.5 $\times$ B4)/(B8 + B4 + 0.5)
Red-edge index	NDVIre1	Normalized vegetation red-edge 1 (B8A − B5)/(B8A + B5)
	NDVIre2	Normalized vegetation red-edge 2 (B8A − B6)/(B8A + B6)
	NDVIre3	Normalized vegetation red-edge 3 (B8A − B7)(B8A + B7)

Note: the Sentinel-2 band was used in all of the calculations of the vegetation/red-edge indexes.

.

The S–G filtering technique was used to smooth the time-series curves of the backscattering coefficients and vegetation indices. This can significantly reduce the interference of noise and effectively remove jagged fluctuations and sudden abnormal values, making the change trends smoother and easier to identify. The post-filtering and smoothing vegetation index time-series curves effectively reveal and depict the distinct temporal patterns of various crops. In view of this, in this study, we utilized the S–G filter to perform filtering on the time-series curves, with the aim of obtaining smoother cross-annual time-series curves. The default settings were used for the S–G filter, which is a filtering method that performs least squares fitting in the time domain based on a moving window. In the filtering process, five points were selected on each side of the filter kernel, and the degree of the smoothing polynomial was set to two for the experiment.

2.4.3. Object-Oriented Classification Algorithms

This research utilized three algorithms—RF, CART, and SVM—to categorize crops within the study region and to computationally ascertain the object-oriented agricultural planting patterns. As a result, tailored to the unique aspects of our research area, the ideal classification technique was chosen for cross-year analysis to sift through and extract data on the crop structure for the region between 2019 and 2024.

Crop classification data in the study area typically exhibit high-dimensional features, with a potentially limited sample size. RF and CART perform well in handling high-dimensional and small-sample datasets, while SVM is also effective in high-dimensional spaces [26,27].

In contrast, deep learning models such as Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks generally require large training datasets and substantial computational resources, which are not available in the study area to support the training of such complex models. Furthermore, although deep learning models have demonstrated superior performance in certain domains, their training processes are complex and computationally intensive. RF and CART offer significant advantages in terms of training speed and computational efficiency [28].

Moreover, model interpretability is of particular importance in the field of agricultural remote sensing. CART and RF provide feature importance analysis, which helps in identifying the most influential features for classification outcomes. SVM also offers a certain degree of interpretability through the selection of kernel functions [29].

2.4.4. Classification Accuracy Evaluation

The overall accuracy represents the ratio of correctly classified pixels to the total pixel count. A standard metric for assessing change detection accuracy, the kappa coefficient offers a more precise measure. It offers a superior evaluation of categorization precision, accurately mirroring the congruence between predicted and observed classes.

The producer’s accuracy (PA) quantifies the percentage of correctly identified pixels along the diagonal of the confusion matrix. It reflects the classifier’s ability to recognize the objects being classified. The user’s accuracy (UA) measures the proportion of correctly classified pixels relative to the actual category’s total pixels. It reflects the extent to which the classification result meets the user’s needs [30].

The classification performance of the three algorithms—RF, CART, and SVM—was evaluated using several key metrics. Given that both overall accuracy and the kappa coefficient are influenced by producer and user accuracy, this study calculated these measures for each crop type within the research area. Specifically, the analysis included the overall accuracy (OA), the user’s accuracy (UA), the producer’s accuracy (PA), and the kappa coefficient to thoroughly assess the algorithms’ effectiveness. These four indicators were used as the basis for the quantitative evaluation. The calculation formulas are as follows:

$$\mathrm{PA}={\displaystyle \frac{X_{ii}}{X_{+i}}}\times 100\%,$$

(3)

$$\mathrm{UA}={\displaystyle \frac{X_{ii}}{X_{i+}}}\times 100\%,$$

(4)

$$\mathrm{OA}={\displaystyle \frac{{\displaystyle \sum _{i=1}^k{X}_{ii}}}{X}},$$

(5)

$$\mathrm{Kappa} ={\displaystyle \frac{N{\displaystyle \sum _{i=1}^k{X}_{ii}}-{\displaystyle \sum _{i=1}^k\left(X_{i+}{X}_{+i}\right)}}{N^2-{\displaystyle \sum _{i=1}^k\left(X_{i+}{X}_{+i}\right)}}}$$

(6)

where N is the total number of pixels, X_ii is the number of correctly classified pixels of a certain category i, X_i+ is the total number of pixels of this category, and X_+i is the total number of pixels of this category in the reference data.

3. Results

3.1. Optimal Scale Preference Results

To optimize multiscale segmentation via rational and reliable parameter selection, segmentation trials were performed. Comparative analyses were carried out by setting different parameter combinations. Regarding the segmentation scale parameter, four ROC peaks were obtained using the ESP tool (Figure 6), with values of 70, 123, 127, and 136. The larger the segmentation scale is, the larger the patches after segmentation are, and the under-segmentation phenomenon is likely to occur. Conversely, the over-segmentation phenomenon exists. Through experiments on the four groups of segmentation scales, the segmentation effects of the four groups of segmentation scales were determined. As shown in Figure 6, when the segmentation scale was 70 (Figure 7a), the segmented objects in the figure were too fragmented, and the patches were dense and fractional. When the segmentation scales were 123 (Figure 7b) and 127 (Figure 7c), the results of the two segmentation scales were only slightly different, and relatively complete field information was retained. When the segmentation scale was 136 (Figure 7d), there was a connection between some ground objects and fields. Therefore, after a comparison of the experimental results, a segmentation scale of 127 was selected.

In addition, the shape factor and compactness factor were adjusted. After experiments and adjustments, in the image object layer with a segmentation scale of 127, it was found that when the compactness factor was set to 0.5 and the shape factor was set to 0.3 (Figure 7e) or when the compactness factor was set to 0.5 and the shape factor was set to 0.8 (Figure 7f), the edges of the figure were more detailed, the segmentation effect between the fields and other ground objects was clear, and the segmentation of the vegetation edges was relatively ideal. Therefore, a segmentation scale of 127, a compactness factor of 0.5, and a shape factor of 0.8 were selected as the optimal scale parameters, and the segmentation result was used as the object for the next step of the feature extraction.

3.2. Time-Series Characteristics of Backscattering Coefficients for Typical Crops

Since the backscattering coefficient curves of different crops in the VH-polarization and VV-polarization modes are generally similar, taking the VH-polarization mode as an example, the annual time-series curve results of the backscattering coefficients of the different crops (Figure 8) were analyzed.

According to the backscattering coefficient curves of the different crops in Figure 7, the growth period of the wheat was concentrated from March to April. During this period, its backscattering coefficient reached a peak value, i.e., it reached approximately −24 dB in the first ten days of April, which was significantly higher than those of the other crops in the same period. The backscattering coefficients of the other crops in the same period were all lower than −20 dB. In addition, the low peak values of the backscattering coefficients of the different crops coincided with their sowing and growth periods. When crops are in the flourishing stage, the amplitudes of their backscattering curves are relatively small. For example, during the flourishing periods of the corn (from May to August) and tomatoes (from July to the first ten days of September), their backscattering coefficients remained stable at around −20 dB and −17 dB, respectively.

In addition, during the flourishing stages of the crops, the amplitudes of the backscattering curves of the gourds, melons, and cotton exhibited similarities. However, due to significant differences in their phenological periods, their backscattering coefficient curves also exhibited different characteristics. The main differences appeared in two key periods: the sowing period and the growth period. Specifically, the peak value of the backscattering coefficient of the melon crops occurred the earliest, usually in the last ten days of April. In contrast, both the seed gourds and cotton reached their peak values in the first ten days of May. In addition, the peak-value range of the seed gourds was the largest among the three. Finally, because the sugar beets needed to go through a long sugar accumulation period during their growth and development, their entire growth and development cycle was relatively long, and the oscillation amplitude of their backscattering coefficient curve was the smallest among the seven crops.

3.3. Temporal Characteristics of Typical Crop Spectral Profiles

Using Sentinel-2 time-series data, four vegetation indices—NDVI, EVI, SAVI, and GEMI—were computed. The S–G filtering method was employed to make the change trends of these vegetation indices smoother and clearer. As shown in Figure 9, the overall change trends of the NDVI, EVI, and SAVI on the spectral reflection curves of the seven different crops are relatively consistent. This indicates that these three indices have good discrimination abilities for the different crops and that the spectral reflectance of the seven different crops in the different phenological periods is relatively stable. However, there are differences in the dynamic ranges of the spectral reflection curves of these three indices. Among them, the EVI in Figure 9b has the largest dynamic range and the highest rate of change, followed by the NDVI in Figure 9a, while the SAVI curve in Figure 9c has the smallest rate of change. This implies that the recognition effect of the EVI may be the most significant during the crop maturity period. In contrast, the spectral curve of the GEMI in Figure 9d exhibits a trend that is quite different from the previous three. During the flourishing stage of the crops, the spectral index of the GEMI decreases, while during the growth and harvest periods of the crops, it increases. This indicates that the GEMI is highly sensitive to the spectral reflectance of the soil, but its performance deteriorates when there is vegetation cover.

Based on the phenological periods and spectral characteristic curves of the different crops, from April to June, wheat has the highest spectral reflectance and matures the earliest. During this stage, the spectral reflectance of the melons and watermelons remains at a relatively low level throughout their growth period. The sugar beets enter the vigorous growth stage in mid-June and maintain a high value thereafter, which is related to its long sugar-accumulation period. In June and July, the melons, seed gourds, and cotton begin to enter a vigorous growth stage, and their spectral reflectance reaches the highest values. However, the time from the flowering stage to the fruiting stage of the melons and watermelons is very short, so their spectral reflectance values decrease rapidly, while the cotton maintains a high value. From July to August, both the corn and tomatoes enter the vigorous growth stage. The harvest period of the tomatoes is relatively short, while that of the corn is relatively long. Therefore, using both the crop phenological calendar and spectral reflection curves is helpful for identifying and classifying the planting structures of the different crops in the study area.

3.4. Typical Annual Crop Classification Results and Accuracy Evaluation

To thoroughly evaluate and contrast the effectiveness of three distinct classification algorithms—RF, CART, and SVM—within the research zone, we segmented the area into four representative regions labeled I, II, III, and IV. This division allowed for a more nuanced assessment of each algorithm’s performance across varying conditions. Detailed enlarged views are shown in Figure 10.

In Area I, according to a comparison with the actual images, the RF algorithm yielded the best classification effect for the cotton fields. The decision tree algorithm performed second best, but it still misclassified a small number of cotton fields as melon fields. The SVM algorithm achieved the worst classification effect, and a large number of cotton fields were classified as melon fields. In Area II, although the classification results of RF and CART for the cotton fields were similar, SVM classified cotton fields into other categories. It should be noted that there were a large number of fragmented plots in Area II. RF tended to classify a large area of these plots as wheat, while the CART and SVM algorithms classified more of these plots as melon fields and corn fields. In Area III, the classification effects of the three algorithms were relatively close. In particular, the results of RF and CART were highly consistent. In Area IV, the classification results of the CART and SVM algorithms were similar, while the RF algorithm was more inclined to classify the plots as cotton.

Figure 11 demonstrates the classification accuracy metrics of three algorithms. The RF algorithm outperformed the other two methods across all four evaluation indices: producer’s accuracy, user’s accuracy, overall accuracy (OA), and kappa coefficient. CART ranked second, while SVM exhibited the lowest performance, with RF achieving a peak OA of 84% compared to SVM’s minimum OA of 61.5%.

An integrated analysis of confusion matrices (Figure 12), quantitative accuracy metrics, and visual interpretation of actual imagery revealed consistent effectiveness across all algorithms for tomato and wheat classification. However, significant misclassification occurred in cotton and melon discrimination, particularly SVM’s substantial misclassification of cotton as melons. Through a comprehensive evaluation of the classification results and accuracy metrics, the RF algorithm was ultimately selected as the optimal method for crop planting structure extraction in the study area.

3.5. Planting Structure Area Extraction for 2019–2024

To verify the accuracy of the classification results of the planting structure in the study area, we systematically statistically analyzed and compared the crop area data obtained from the RF classification for six consecutive years from 2019 to 2024 with the officially reported crop area data. The specific data for each year are presented in

Table 5. Stochastic random forest classification area and statistical area results.

Crops		2019	2020	2021	2022	2023	2024
Wheat	RFC (hm2)	1891.41	2952.41	2015.71	3040.03	6412.66	3498
	Statistical Area (hm2)	1933.33	1986.67	2266.67	3706.67	6066.67	3886.67
Maize	RFC (hm2)	6006.67	6821.29	7211.86	6849.96	8285.35	8516.66
	Statistical Area (hm2)	6506.67	5766.67	6700	7013.33	7786.67	8333.33
Cotton	RFC (hm2)	20949.5	22,894.77	19,541.52	25,390.07	22,240.45	25,190
	Statistical Area (hm2)	20,746.67	22,426.67	18,133.33	24,626.67	21,246.67	22,900
Sugar beet	RFC (hm2)	662.42	674.46	231.13	527.05	391.2	816
	Statistical Area (hm2)	706.67	766.67	213.33	613.33	833.33	906.67
Tomato	RFC (hm2)	3599.11	2008.03	2899.34	2941.84	2576.09	6120.99
	Statistical Area (hm2)	3120	3320	2513.33	1833.33	3926.67	5533.33
Melon	RFC (hm2)	2085.94	1893.65	1023.51	690.98	1001.83	1091.99
	Statistical Area (hm2)	1693.33	2152	1726.67	1306.67	926.67	1213.33
Seed gourd	RFC (hm2)	1907.03	1699.18	1612.76	737.12	926.67	630
	Statistical Area (hm2)	2536	3228	2893.33	1060	618.67	700

.

To further explore the accuracy of the classification results, in this study, we conducted a fitting analysis of the area data classified by RF and the officially reported area data over six consecutive years from 2019 to 2024. The results are shown in Figure 13. For the classification results of the planting structure over these six years, the R² values of the classified area and the statistical area are all higher than 0.8, indicating a good degree of fit. Specifically, the R² values of the classified areas of cotton and wheat compared with the officially reported areas are as high as 0.9329 and 0.9436, respectively, exhibiting an excellent fitting effect. The fitting effects of corn and seed gourds are slightly lower, with R² values of 0.8792 and 0.8956, respectively. The fitting results of tomatoes, melons, and sugar beets are lower.

A comparative analysis of cultivated areas for seven major crops was conducted between the RF method and statistical data, as shown in Figure 14. The results revealed significant discrepancies in area estimation across crop types (denoted by *, p < 0.05), with statistical significance confirmed through hypothesis testing. The RFC method demonstrated substantially higher median values than the statistical regional approach for cotton, maize, tomatoes, and wheat cultivation area estimations. In contrast, melon and sugar beet showed comparable median values between datasets with no significant discrepancy (p > 0.05). Although seed pumpkin exhibited statistically significant differences between methods (p < 0.05), the minimal median divergence (Δ = 1.2 ha) suggests limited practical implications.

3.6. Spatial Distribution of the Planting Structure During 2019–2024

During 2019–2024, the spatial distribution of the various crops exhibited certain regularities and characteristics (Figure 15). Specifically, the planting distribution of cotton was relatively fixed and concentrated, it was mainly located in the upper part of the study area, and its planting area was relatively large. In addition, the planting range of corn was also relatively stable, and its planting area was second only to that of cotton. In contrast, the planting area of tomatoes was mainly concentrated in the lower part of the study area, and it was often intercropped with corn. The planting distribution of wheat was relatively scattered, but its planting range was extensive, covering the entire irrigation area. In addition, the distributions of melons, seed gourds, and sugar beets were relatively dispersed, and their planting areas were relatively small—i.e., smaller than the planting scales of the cotton and corn.

During the period from 2019 to 2024, the spatial distribution of various crops exhibited distinct patterns and characteristics, as shown in Figure 16. The proportions of different crops demonstrated significant fluctuations during this period. Cotton accounted for a relatively large proportion from 2019 to 2022, declined in 2023, but rebounded in 2024. The proportion of maize gradually increased from 2019 to 2021, peaked in 2022, slightly decreased in 2023, and remained stable in 2024. Wheat showed an increase in proportion during 2020 and 2021 but stabilized after 2022. The proportion of tomatoes remained relatively stable throughout the period, generally ranging between 10% and 15%. The proportion of melons fluctuated between 5% and 10%, with minimal changes. Based on this development trend, cotton is expected to remain the dominant crop, though its proportion may stabilize or fluctuate slightly. The proportions of maize and wheat are likely to remain stable or undergo minor changes. The proportions of tomato, melon, seed gourd, and sugar beet, with smaller variations, are projected to maintain their current levels.

4. Discussion

4.1. Comparative Analysis of Different Algorithms for Crop Cultivation Extraction

To enhance the precision of crop classification and explore the distinct performance of various algorithms, this research employed three models—RF, SVM, and CART—for comparative analysis. The findings showed notable variations in crop classification performance among the three algorithms. Among them, the RF algorithm performed best, with an OA index of 84% and a kappa coefficient of 0.805. The kappa coefficients of RF were 0.165 and 0.255 higher than those of the CART and SVM algorithms, respectively, thus clearly demonstrating the advantages of the RF algorithm in terms of crop classification accuracy.

RF integrates multiple decision trees and uses the bagging (bootstrap aggregating) technique to sample the data with replacement, generating multiple different training subsets. Remote sensing image data usually have a very high dimension, including spectral features, texture features, and polarization features. RF can randomly select features for splitting in each tree, avoiding the use of all of the features, thereby effectively dealing with high-dimensional data and reducing the influence of redundant information. Consequently, the random forest method sees extensive application in analyzing remotely sensed information. Mousavi et al. [31] and Guo et al. [32] used RF in grassland mapping and dominant tree species classification, respectively.

In contrast, decision trees are prone to capture details and noise in deep learning data, leading to overfitting, especially when the data volume is limited or there is noise. In our study, the decision tree algorithm achieved an OA of 69.7% and a kappa coefficient of 0.64. Its accuracy ranked just below RF but outperformed SVM. Through comparative analysis of RF and CART, we found that RF effectively reduced the risk of the model falling into overfitting by introducing multiple randomization mechanisms such as sample sampling and feature sampling. In their research on the extraction of crop planting structure information, Zhao et al. [33] also confirmed that the results of RF were better than those of CART, which is highly consistent with the conclusion of this study.

In addition, in this study, we also found that RF was superior to SVM in terms of both efficiency and results, especially when dealing with large-scale remote sensing images. Liao et al. [34] conducted a comparative analysis of the RF algorithm and SVM in their research on crop mapping. They observed that RF outperformed SVM in accuracy, aligning with this study’s findings.

In practical applications, remote sensing image data often have high dimensions and limited sample sizes. RF can avoid the issue of dimensionality caused by high dimensional sparsity by randomly selecting some features. SVM may lead to an overly complex model due to the use of high dimensional data, while decision tree is easily affected by dimensional redundancy. There may be interference factors such as cloud shadows and cross-interference of ground objects in remote sensing images. The robustness and noise resistance of RF enable it to handle these problems more stably. CART and SVM may not perform as well as RF when facing unbalanced data or interference factors.

4.2. Reliability of Model Migration

In this study, by constructing RF based on phenological differences and a multidimensional feature dataset, high-precision crop distribution mapping for six consecutive years from 2019 to 2024 was achieved for the Santun River Irrigation Area in Xinjiang, a typical irrigation area in an arid region. In addition, by fully utilizing the polarization characteristics of the Sentinel-1 SAR satellite and the band information of the Sentinel-2 multispectral satellite, the temporal resolution was significantly enhanced and the classification feature set was enriched, thereby improving the accuracy and comprehensiveness of the crop identification.

Regarding the core challenge in model transfer research, namely, the uneven spatial distribution of sample points, in this study, we improved two previous strategies to optimize the sample set. Zhang et al. [35] effectively increased the number of sample points by fitting the characteristics of the time-series NDVI curves during the crop growing season, alleviating the problem of an uneven sample spatial distribution to a certain extent. Xu et al. [36] analyzed the spectral curves and index features of sample points from the measured data years and then randomly introduced visually interpreted points based on these features and screened them according to the spectral characteristics as supplementary samples, further enriching the diversity and representativeness of the sample set.

When conducting model transfer, the quantity and quality of the sample points are crucial for the successful transfer of the model. Therefore, in this study, in addition to using multiple index features such as the NDVI, SAVI, EVI, and GEMI, we also considered the VV and VH polarization features of the Sentinel-1 SAR satellite, thus enhancing the information content and reliability of the sample points.

Through comparison of the crop area estimation results and the officially reported area data for 2019–2024, we found that there was a high degree of consistency between the crop areas calculated by the transferred model and the officially reported areas, and the coefficient of determination (R²) values were all greater than 0.8. In particular, the R² values for cotton and wheat exceeded 0.9. This study confirms the high stability of the model transfer process and demonstrates the wide application potential of this method in crop mapping in typical irrigation areas in arid regions.

Although the model demonstrated robust interannual transferability over the six-year period, ongoing vigilance is required regarding potential temporal performance degradation. Model stability is primarily constrained by three interrelated factors: (1) phenological shifts induced by climate change; (2) alterations in spectral response patterns resulting from evolving agricultural management practices; and (3) systematic remote sensing feature deviations caused by sensor performance degradation or data acquisition inconsistencies.

4.3. Analysis of the Effects of Vegetation Indices on Crop Classification

To date, numerous types of vegetation indices have been developed. The selection of vegetation indices varies depending on the band characteristics of the spectral satellite sensors used to collect the data and the monitoring conditions [37,38]. Accurately selecting appropriate indices for extracting crop planting structures has become a key challenge.

Past research has frequently utilized the NDVI as a key vegetation indicator. Bhatti et al. [39] used the NDVI values of different crops over three seasons to identify and monitor crops in the Sargodha region. Sang et al. [40] studied the spatiotemporal variation characteristics and evolution trends of vegetation in Hunan Province using NDVI time-series data. The NDVI is derived from red and near-infrared band reflectance. Chlorophyll absorbs the red band intensely, whereas vegetation reflects the near-infrared band effectively. Due to the differences in the chlorophyll absorption and near-infrared reflection of vegetation, the NDVI can well reflect the greenness, density, and growth status of vegetation.

With the development of vegetation indices, the EVI emerged. The EVI adds the blue band on the basis of the NDVI and compensates for atmospheric influence, thus achieving stronger sensitivity in areas with high vegetation cover and more accurately reflecting the growth dynamics of crops. Guan et al. [41] reported that in wheat crop identification, EVI time-series data could better describe the phenology of crops than the NDVI, and the extraction accuracy achieved using the EVI is higher than that achieved using the NDVI. This result is consistent with the findings of this study, that is, the EVI is superior to the NDVI in identifying the critical growth stages of crops.

In irrigation districts in arid regions, vegetation is often sparse, and soil background reflection may interfere with vegetation indices. The SAVI uses an L factor to minimize soil reflectance effects. Bera et al. [42] evaluated the land use value using SAVI values according to the types of planted crops, demonstrating the advantages of the SAVI under strong soil background reflection conditions. In areas with sparse vegetation and more exposed soil, soil background reflection may interfere with NDVI results. Hence, SAVI enhances monitoring precision in scant vegetation zones via soil background effect correction. In this study, we focused on an irrigation district in an arid region, and our results verify that the SAVI has a monitoring scope that the NDVI does not have in arid areas. In particular, during the crop sowing period, during which the soil reflectance is high, the SAVI has certain advantages.

In addition, as a comprehensive vegetation index, the GEMI integrates reflectance data from multiple bands and uses various models to evaluate factors such as global-scale vegetation biomass and climate change. Saeid et al. [43] and Varnosfaderani et al. [44] demonstrated the large-scale advantages of the GEMI in crop and forest identification and classification, respectively, indicating that the GEMI focuses more on large-scale and multidimensional ecological monitoring.

In terms of scale differences and regional adaptability, the NDVI and SAVI are usually effective for small-scale agricultural crop monitoring. In contrast, the GEMI is more suitable for large-scale ecological monitoring and crop biomass assessment, providing data support for the macroscale adjustment of crop planting structures. In this study, we comprehensively considered the characteristics of the NDVI, EVI, SAVI, and GEMI, with the aim of improving the accuracy of crop classification in irrigation areas in arid regions and providing a reference model framework for irrigation areas in arid regions globally.

4.4. Factors Influencing Spatiotemporal Changes in the Planting Structure in the Study Area During 2019–2024

There are significant regional differences in the factors influencing planting structure changes, and these factors can be generally divided into two major categories: technological development factors and market factors.

From a technological standpoint, saline–alkali soil remediation is critically important. Through a series of comprehensive treatment measures, the saline–alkali land in Xinjiang has been significantly improved, laying a solid foundation for the expansion of agricultural production [45,46]. Green and water-saving agricultural technologies are also important driving forces. Guided by the concept of green agriculture, highly efficient water-saving technologies such as drip irrigation under plastic film have been widely applied, greatly improving irrigation efficiency and strongly promoting the continuous growth of planting areas [47]. Moreover, the application of modern agricultural technologies has further optimized the agricultural production model. The popularization of technologies such as drip irrigation under plastic film has effectively improved the utilization efficiency of water resources, providing direct support for the large-scale development of agriculture [48].

From the perspective of market factors, the strategic position of the cotton industry is of great significance. The cotton planting area in Xinjiang continuously expanded from 2019 to 2024. Taking the Santun River Irrigation Area in Xinjiang as an example, the total cotton planting area increased by 4240.5 hm² from 2019 to 2024. Except for 2021, the planting area exceeded 20,000 hm² in the other five years. With its excellent quality and large-scale advantages, Xinjiang cotton effectively ensures the security of China’s cotton industrial chain and fully meets the needs of both domestic and international markets [49,50].

The demand-driven factor of food crops is also important. The planting areas of food crops such as corn and wheat in arid regions are increasing. In the Santun River Irrigation Area in Xinjiang, the planting areas of major food crops such as corn and wheat reached a peak value of 14,698 hm² in 2023 and exhibited an overall increasing trend from 2019 to 2024. This fully reflects the strategic layout for ensuring food security and a positive response to market demand. Furthermore, the economic advantages of characteristic crops are prominent. Relying on the unique climate resources in Xinjiang, characteristic crops such as fruits, melons, and tomatoes have formed strong market competitiveness and become important growth points for the agricultural economy.

4.5. Uncertainty Analysis and Prospects

In research on the extraction of planting structures, there are still several uncertain factors. First, the multiple-cropping situation of fields was not considered in this study. In reality, due to human intervention or natural factors, multiple-cropping may occur in some fields. This omission may lead to inaccuracies in the extraction results.

Second, the quantity and spatial distribution of the sample points and their fitting sample points also significantly affect the extraction effect. This necessitates a more rigorous and scientific approach to sample selection.

Moreover, although Sentinel-2 images have a high resolution and can accurately depict ground features, the problem of mixed pixels is still inevitable in the inter-cropping mode, posing a challenge to achieving a high extraction accuracy.

Finally, in this study, we only analyzed two polarization modes (VV and VH) of microwave data. The potential of integrating fully polarized microwave and multi-source remote sensing data for ground object classification requires further exploration.

In conclusion, in future research on planting structure extraction, various factors should be comprehensively considered to improve the accuracy and reliability of the results.

5. Conclusions

(1) By constructing the inter-annual variation curves of the backscattering coefficients of ground objects using time-series Sentinel-1 data and integrating Sentinel-2 multispectral data, the growth information of crops can be well obtained. Through similarity matching of time-series curves, in combination with the statistical characteristic parameters of the curves, the accuracy of extracting complex planting information among different crops in irrigation districts in arid regions can be improved. This study establishes a transferable technical framework for multi-source remote sensing synergy monitoring in global arid regions.

(2) In the identification and classification of crops in remote sensing images, RF has a strong robustness against high-dimensional data and noise. It performs more stably in multi-classification tasks, is easy to implement, and has a high training efficiency, making it suitable for large-scale remote-sensing data processing. In this study, the overall accuracy of the crop classification achieved using the RF algorithm reached a maximum of 84%, the kappa coefficient reached 0.805, and the accuracy of the confusion matrix was superior to that of CART and SVM. In addition, when comparing the classified area obtained using RF with the statistical area, the R² values were all higher than 0.8. This study demonstrates that the proposed methodology provides reliable geospatial decision-making support for agricultural water resource management in arid regions along the Belt and Road Initiative (BRI), revealing its practical value in transboundary agricultural management.

(3) From 2019 to 2024, in the study area, the total crop planting area exhibited an increasing trend, with a total increase of 8961 hm² over the six-year period. Among the crops, the annual average planting area of cotton was the largest with an average planting area of 22,701 hm² during the six-year period, and its planting area was relatively concentrated and fixed. The planting area of corn ranked second, with an average planting area of 7281.9 hm² during the six-year period. The planting areas of wheat, tomatoes, melons, and sugar beets were relatively small, and their planting distributions were also more scattered.