A hierarchical framework for cropland and crop classification in the Qingtongxia irrigation district using support vector machine and random forest models

Abstract

Accurate classification of cropland and crop types is essential for agricultural management in arid irrigation districts but remains challenging due to spectral similarity among crops and interference from heterogeneous non-cropland backgrounds. To address these challenges, this study develops a hierarchical framework for cropland and crop-type mapping using Sentinel-2 time-series data on the Google Earth Engine platform. A two-stage strategy is adopted, in which cropland is first delineated using a support vector machine with a radial basis function kernel (SVM_RBF) based on multi-temporal spectral features and indices, followed by crop classification within the cropland mask using a random forest (RF) model incorporating phenology-driven temporal features. Using Sentinel-2 imagery (10 m resolution) and ground-truth samples from the Qingtongxia Irrigation District, the proposed approach achieves an overall accuracy of 94.61% for cropland classification. The overestimation of cropland area is reduced to 9.7%, compared with 17–19% in existing high-resolution land-cover products. For major crops such as maize and rice, classification accuracies exceed 86%, representing an improvement of 5–10% over single-stage classification approaches. This study also provides a systematic comparison between the derived cropland map and existing high-resolution land-cover products, highlighting the challenges that such products may encounter when applied to heterogeneous irrigation-district environments. Overall, this study provides an interpretable and practical framework for improving crop mapping in complex irrigation environments.

Introduction

Cropland extent and cropping structure are foundational data for agricultural management and water-resource allocation^1,2. In agriculturally intensive regions, particularly irrigation districts, crop types are diverse and vary markedly across space and time. Accurate mapping of crop distributions is essential for food security, irrigation scheduling, and evidence-based agricultural policy^3–5. Consequently, producing high-accuracy maps of cropland and cropping structure remains a central challenge in agricultural remote sensing.

Several publicly available land use and crop structure datasets have been released globally and in China. FROM-GLC10 is a global 10 m resolution land cover product derived from Sentinel-2 imagery, characterized by a cropland user accuracy of 58.14% and a producer accuracy of 61.06%, with coverage limited to 2017⁶. ESRI-GLC10 provides global land use/land cover (LULC) maps for 2017–2023 at 10 m resolution, with cropland user accuracy exceeding 80%⁷. ESA-GLC10, generated from Sentinel-1 and Sentinel-2 data, reports a global overall accuracy of 76.7% and 82.1% in Asia⁸. The China Land Cover Dataset (CLCD), the first annual land cover product for China, is based on Landsat imagery at 30 m resolution for 1990–2019, with cropland user accuracy of 75%⁹. SinoLC-1, the first nationwide 1 m resolution dataset, integrates December 2021 Google Earth imagery with three global 10 m products, achieving 67.86% cropland user accuracy¹⁰. ChinaCropland10m combines Sentinel-1 SAR and Sentinel-2 MSI imagery to produce 10 m cropping intensity maps for 2018–2023, with overall accuracy of 90.88%¹¹. Other datasets, such as the Cropland Data Layer (CDL) and WorldCereal 2021, offer crop structure data for the United States and globally^12,13, though WorldCereal 2021 shows lower accuracy in Asia due to complex agricultural landscapes and uneven training sample distribution. In China, 30 m maize and winter wheat maps for 2001–2020, derived from MODIS and Landsat data, achieve overall accuracies of 80.06% and 91.6%, respectively^14,15, but cover only selected major crops.

Crop classification based on NDVI time series is a very common approach in remote sensing and agricultural monitoring. This method leverages the unique phenological patterns (e.g., growth peaks and cycles) captured in NDVI profiles over time, allowing differentiation between crop types like rice, wheat, and maize through temporal features. Wardlow et al.¹⁶, and Wardlow and Egbert¹⁷used MODIS 250 m 16-day composite NDVI/EVI data with decision tree classification to map major crops in the U.S. Central Great Plains and distinguish irrigated from non-irrigated areas. Song et al.¹⁸, applied 16 m GF-1/WFV imagery, combining object-based segmentation, random forest classification, and NDVI time series with texture features to produce multi-stage crop maps during the growing season with high accuracy. Xu et al.¹⁹, integrated Landsat NDVI time series with AtLSTM and Transformer models to identify soybean in the U.S. Corn Belt, highlighting the importance of temporal growth characteristics in NDVI curves. Li et al.⁴⁵, fused MODIS and Landsat data to extract NDVI time series via an improved spatiotemporal data fusion (IFSDAF) model for crop classification. More recently, combining Sentinel-1 and Sentinel-2 imagery with NDVI and other vegetation indices has proven effective for national-scale crop mapping^20,21. Deep learning approaches have further advanced this field, including lightweight CNN-Transformer networks for pixel-based mapping²², long-term dataset generation using attention-based architectures²³, and sample-transfer strategies for in-season mapping without current-year labels²⁴. Explainable AI studies have confirmed that NDVI remains a critical feature even when multi-modal data are integrated²⁵. Collectively, these studies underscore the importance of NDVI time series for capturing phenological characteristics essential to crop classification.

Although recent studies have demonstrated strong performance in crop mapping using deep learning models such as CNNs, LSTMs, and Transformer-based architectures, these approaches typically require large volumes of high-quality training samples and dense temporal observations. In heterogeneous irrigation districts, where field sizes are small, cropping patterns are fragmented, and labeled samples are limited, such requirements are often difficult to meet. In addition, the limited interpretability of deep learning models constrains their applicability in studies that aim to understand crop-specific phenological mechanisms. Moreover, while global cropland and crop products offer broad coverage, their accuracy at the regional scale is often inadequate for irrigation-district applications²⁶.

To address these challenges, this study develops a conceptually driven hierarchical framework that explicitly targets two key sources of classification error in irrigation districts: (1) spectral confusion between cropland and non-cropland, and (2) spectral similarity among crop types with overlapping phenological patterns. Unlike conventional hierarchical approaches that treat classification stages as loosely connected processes, the proposed framework explicitly integrates crop-specific phenological signals into the second-stage classification. In particular, phenology-guided features such as flooding-sensitive indices and near-infrared reflectance during critical growth stages are incorporated to enhance separability among spectrally similar crops. Therefore, the novelty of this study lies not in the development of a new classifier, but in a problem-oriented hierarchical design combined with phenology-driven feature integration, providing an interpretable and practical solution for crop mapping in complex irrigation systems.

Study area and data

Study area

The Qingtongxia Irrigation District, located in the mid-northern Yellow River alluvial plain of Ningxia (105°37’−106°39’E, 37°49’−39°23’N), spans Yinchuan, Shizuishan, and Wuzhong (Fig. 1). Covering approximately 650,000 ha, of which 330,000 ha are irrigated, it ranks among China’s six largest irrigation districts. The region is characterized by a temperate arid climate, with a mean annual temperature of 11 °C, annual precipitation of 180–220 mm (70% occurring between July and September), annual evaporation of 1,825 mm, and a frost-free period of 160–180 days. The average annual water diversion in the Qingtongxia Irrigation District is about 5.0 billion m³. In 2023, the diverted volume was 32.48 billion m³, a decrease of 1.95 billion m³ compared with the previous year. Paddy rice requires approximately 18,000–22,500 m³/ha annually, and agricultural production consumes nearly 80% of the total water use. In 2023, the irrigation water use coefficient was 0.58, with a mean application rate of 7,725 m³/ha. The total drainage volume reached 1.1533 billion m³, down 0.2843 billion m³ from the previous year, which also played a role in regulating groundwater levels. The mean groundwater depth was 2.31 m, decreasing to 1.76 m during winter irrigation and increasing to 2.82 m before spring irrigation (Ningxia Hui Autonomous Region Water Resources Department²⁷. As “granary beyond the Great Wall” of Ningxia, the district produces 43% of the region’s grain. Major crops include maize, rice, wheat, and soybean (Table 1).

Fig. 1.

Overview of the qingtongxia irrigation district (map created using QGIS version 3.34; https://qgis.org/).

Table 1.

Growing seasons of major crops in the Qingtongxia Irrigation District.

Major Crops	March	April	May	June	July	August	September	October
Wheat	Sowing		Heading		Maturity
Maize		Sowing		Jointing			Maturity
Rice			Flooding	Tillering	Jointing	Grain filling	Maturity
Soybean		Sowing			Podding		Maturity

Satellite remote sensing data

Sentinel-2 A/B Level-2 A imagery for 2023 was obtained from Google Earth Engine for the Qingtongxia Irrigation District (275 scenes, mid-March to early November), covering the entire crop growing season. The dataset includes 10 m and 20 m bands, comprising visible (B02-B04), red-edge (B05-B07, B8A), NIR (B08) and shortwave-infrared (B11, B12) wavelengths, with the latter two being particularly valuable for vegetation assessment and crop classification.

Ground truth data

Ground truth sample data were obtained from field surveys and visual interpretation of high-resolution Google Earth imagery. Field surveys conducted in late July to early August 2023 collected 723 points: 443 croplands (wheat, n = 80; maize, n = 198; rice, n = 75; soybean, n = 35; other crops, n = 55), 45 water samples, 30 built-up samples, 148 forest samples, and 57 fallow land samples (including grassland and bare-land). An additional 663 points were interpreted from high-resolution Google Earth imagery using spatially uniform random sampling: 318 cropland samples, 35 water samples, 35 built-up samples, 175 forest samples, and 100 fallow land samples. In total, 1,386 samples were obtained and split into training and validation sets in a 7:3 ratio. The validation samples were used for accuracy assessment of both the classification results and the interpretation-based reference map. The spatial distribution of samples is shown in Fig. 2.

Fig. 2.

Distribution of sample points across the Qingtongxia irrigation district (map created using QGIS version 3.34; https://qgis.org/).

To evaluate the spatial independence of samples, nearest-neighbor distance statistics and spatial autocorrelation analyses were conducted. The nearest-neighbor analysis shows that the median distance between samples is 878.76 m, with the 25th and 75th percentiles being 470.69 m and 1405.73 m, respectively, indicating that most samples are spatially well separated. Spatial autocorrelation was further evaluated using Moran’s I (Cliff²⁸), based on NDVI values derived from Sentinel-2 imagery (tile T48SXH) acquired on 25 July 2023. The result shows a Moran’s I value of 0.167 (p < 0.01), indicating weak but statistically significant spatial autocorrelation. In addition, a semivariogram analysis was performed, yielding a spatial autocorrelation range of approximately 496 m (Fig. 3). Since the majority of sample pairs are separated by distances greater than this range, spatial dependence among samples is considered limited. Therefore, the random division of samples into training and validation sets is unlikely to introduce significant bias in accuracy assessment.

Fig. 3.

Semivariogram of spatial autocorrelation for NDVI at sample points.

Land cover products for comparison

Several high-resolution land-use/land-cover (LULC) datasets have been released both domestically and internationally. In this study, three products, ESRI-GLC10 (2023), ChinaCropland10m, and SinoLC-1, were selected for accuracy comparison with the cropland map generated in this study. The three selected datasets were produced using different remote-sensing imagery and classification methods, leading to variations in spatial resolution, thematic schemes, and accuracy specifications^7,10,11. Detailed information is provided in Table 2. To ensure spatial consistency and comparability across datasets, uniform preprocessing was performed, including projection transformation, co-registration, resampling and spatial clipping. All products reprojected to the UTM Zone 48 N (WGS 84) coordinate system and standardized to a 10 m spatial resolution as the cropland map derived from the SVM_RBF. Simultaneously, the three selected products were reclassified into a binary scheme (cropland vs. non-cropland), where only the cropland class was retained as “cropland”, and all other categories (e.g., forest, grassland, water, and built-up areas) were uniformly merged into a single “non-cropland” class across all products.

Table 2.

Basic information on land-cover products.

Dataset	Producer/Institution	Primary data sources	Classification method	Resolution
ESRI-GLC10	Impact Observatory and Esri7	Sentinel-2 optical imagery and manually dense markup	U-Net CNN	10 m
ChinaCropland10m	Fuzhou University11	Sentinel-1 SAR data and Sentinel-2 optical imagery	VSPS algorithm and MILS method	10 m
SinoLC-1	Wuhan University10	1 m Google imagery; three 10 m land-cover maps (weak labels); OpenStreetMap	L2HNet framework	1 m

Note: Abbreviations used in this Table 2 include U-Net CNN (U-Net Convolutional Neural Network), VSPS (Vegetation–Soil–Pigment–SAR), MILS (Mapping cropping Intensity based on combined consideration of vegetative and productive Stages), and L2HNet (Low-to-High Network).

Method

This study proposes a hierarchical method for mapping cropland and cropping structure from Sentinel-2 time-series imagery. The workflow consists of image preprocessing, phenology-driven feature extraction, parameter calibration, cropland extraction using support vector machine with radial-basis-function (SVM_RBF), major-crop classification with random forest model (RF).

Image preprocessing

To ensure the consistency and validity of imagery for crop classification, Sentinel-2 L2A data were screened and preprocessed on the Google Earth Engine platform. Images with < 20% cloud cover during the primary crop growing season (March-November) were retained, with clouds and shadows removed using the QA60 band and the s2cloudless algorithm (Gao et al., 2024). The 20 m bands (B06, B11) were resampled to 10 m via bilinear interpolation to match the other 10-m bands. The resulting cloud- and shadow-filtered scenes were used as inputs for subsequent phenology-driven feature extraction and classification.

Hierarchical classification framework and classifier

This study adopts a hierarchical classification strategy in which cropland is identified in a first stage and crop types are discriminated in a subsequent stage. In the first stage, a Support Vector Machine (SVM) is used to perform binary classification to separate cropland from non-cropland. Conceptually, SVM finds a maximum-margin hyperplane—formulated as a quadratic programming problem—in a (possibly) high-dimensional feature space, thereby promoting robust separation between classes^29–32. Nonlinear decision boundaries can be achieved by applying kernel functions; common choices include the linear, radial basis function (RBF), and polynomial kernels³³. While the linear kernel is computationally efficient, it may be inadequate for complex, nonlinear distributions; by contrast, the RBF kernel is widely used in land-cover and crop classification owing to its strong generalization properties^34,35. Therefore, SVM with RBF kernel (SVM_RBF) was used to map cropland in the study area.

The classification performance of the SVM_RBF model is governed primarily by two key hyperparameters: the penalty parameter and the kernel parameter . The penalty parameter controls the degree of penalization for misclassified samples, balancing the trade-off between maximizing the margin and minimizing training errors. The kernel parameter determines the influence range of each training sample in the feature space, thus regulating the model’s nonlinear mapping capacity³⁶. The kernel function is defined as:

1

Where and denote input feature vectors. A larger reduces the effective influence range of the kernel, producing more localized decision boundaries that may fit the training set more closely but risk overfitting. Conversely, a smaller expands the influence range, potentially yielding overly smooth boundaries and underfitting.

The second stage addresses multi-class crop classification and employs a Random Forest (RF) classifier. RF is an ensemble method based on bootstrap aggregating (bagging) that is robust to high-dimensional and noisy data and is computationally efficient. It has been widely validated for crop classification^37,38. Using Classification and Regression Trees (CART)³⁹ as base learners, RF builds an ensemble of decision trees from bootstrap-resampled training subsets and selects the split variable at each node from a randomly drawn subset of features⁴⁰. Trees are typically grown without pruning to maximize ensemble diversity, and final predictions are determined by majority voting. RF also provides an internal accuracy estimate via out-of-bag (OOB) samples, which are observations excluded from a given tree’s bootstrap sample; this feature often obviates the need for a separate validation set⁴¹. Key hyperparameters include the number of trees (_tree) and the number of features considered at each split (_try). In classification tasks, _try is commonly set to , where is the total number of features. _tree should be chosen large enough to stabilize predictions. Model variance and generalization can be controlled by limiting maximum tree depth or increasing the minimum number of samples per leaf, and hyperparameters can be tuned using OOB error or cross-validation.

This hierarchical framework integrates the strengths of the SVM_RBF and RF classifiers to address classification tasks at different levels. Implemented on the Google Earth Engine platform, it achieves a balance between computational efficiency and classification performance.

Accuracy evaluation and production of reference data

In this study, a confusion matrix was adopted as the evaluation framework to derive overall accuracy (OA), producer’s accuracy (PA), user’s accuracy (UA), F1-score and the Kappa coefficient. The Kappa coefficient was applied to correct for chance agreement, thereby providing a more objective assessment of classification performance.

To support accuracy assessment, high-quality reference maps were generated through a combination of object-based analysis of 2023 Sentinel-2 Level-2 A imagery and high-resolution visual interpretation using Google Earth, supported by prior field survey experience. Importantly, reference maps were constructed independently and did not rely on the collected sample points. Specifically, multi-scale segmentation was first applied to the Sentinel-2 scenes to generate homogeneous image objects, from which spectral and shape features were extracted. A Classification and Regression Tree (CART) was then employed for preliminary discrimination between cropland and other land cover types, as well as for initial crop-type labeling. Subsequently, each image object was visually interpreted and manually refined using resolution Google Earth imagery (sub-meter resolution) according to predefined interpretation criteria. The interpretation process was further supported by the prior field survey experience and knowledge of local crop types, which enhanced the reliability of the reference data. The accuracy of the reference map was evaluated using validation samples, yielding a Kappa coefficient of 0.92 and an overall accuracy (OA) of 96.64% for cropland mapping, while crop-specific reference maps achieved a mean Kappa of 0.89 and a mean OA of 91.24%, meeting the predefined accuracy requirements. It should be noted that the reference map is considered a high-quality interpretation-based dataset rather than absolute ground truth.

Results

Temporal NDVI profiles and characteristic spectral signatures of land covers

Temporal NDVI profiles provide an effective means of characterizing the seasonal dynamics of different crop and land-cover types. Using Sentinel-2 imagery acquired in 2023, NDVI values were extracted from 723 field survey sample points and plotted against acquisition dates to produce temporal profiles (Fig. 4). For wheat, NDVI rose sharply in early April, peaked between late May and early June, and declined rapidly from late June to mid-July, corresponding to the maturation and harvest period. Maize exhibited an increase in NDVI beginning in mid-May, maintaining high values from mid-June to late August—indicating peak vegetation cover and photosynthetic activity—before falling below 0.35 in mid- to late September, marking the onset of maturation, leaf senescence, and harvest. Rice displayed a similar pattern, with NDVI rising in early June, sustaining high values from mid-June to late August during vigorous growth, and dropping below 0.35 in mid- to late September as it entered maturity and harvest. Soybean followed a comparable trajectory, with NDVI increasing from mid-May, remaining elevated from mid-June to late August—reflecting rapid biomass accumulation—and declining below 0.35 in mid- to late September during maturation and harvest. The NDVI profiles of maize, rice, and soybean were highly similar in overall shape.

Fig. 4.

Variation of NDVI values for major crops and land-cover types in 2023 derived from Sentinel-2 imagery.

Forest land exhibited an NDVI increase from early April, reaching its annual maximum in July–August, followed by a gradual decline from September onward, with values consistently higher than those of major crops throughout the year. Built-up areas maintained low and stable NDVI values year-round, indicative of sparse vegetation cover. Water bodies showed NDVI values consistently below zero, reflecting their spectral properties of strong absorption in the red band and low reflectance in the near-infrared. The “other” category, which includes vegetables, alfalfa, and other multi-cropping systems, exhibited substantial intra-annual variability. Fallow land maintained generally low NDVI values throughout the year, with only minor fluctuations. To efficiently separate cropland from other land covers, a suite of spectral indices was incorporated into the cropland classification procedure: vegetation indices (NDVI, EVI), built-up indices (NDBI), and a water index (MNDWI). These indices were chosen for their complementary information on canopy greenness, surface moisture and impervious surfaces, thereby providing stable and representative inputs for the classification model.

During the paddy rice flooding period (May–July), the NDVI profiles of rice, maize and soybean are highly similar, making discrimination based on NDVI alone difficult. To improve separability, water index (MNDWI) was used computed from the green band (B03) and shortwave infrared (SWIR, B11). Water surfaces typically show relatively high green reflectance but very low SWIR reflectance, so the index effectively highlights inundation. In addition, B03 captures canopy greenness (for example, the deep-green peak of maize), while B11 is sensitive to leaf and soil moisture, which helps distinguish flooded rice paddies from upland crops.

To differentiate soybean from maize, multispectral reflectance values were extracted for sample points of both crops at each phenological stage. The mean spectral difference between the two crops was computed for each acquisition date, negative values were converted to their absolute values for comparison (Fig. 5). The results show that maize and soybean exhibit relatively small spectral differences in the blue (B02), green (B03), red (B04), and red-edge (B05) bands, whereas more pronounced differences occur in the red-edge (B06), near-infrared (B08), and narrow near-infrared (B8A) bands, particularly in imagery acquired between 20 July and 15 August. Considering image spatial resolution and classification accuracy requirements, the 10 m resolution near-infrared band (B08) from late July to mid-August was ultimately selected as the primary feature band for distinguishing soybean from maize.

Fig. 5.

Time series of mean spectral differences between maize and soybean at different phenological stages in 2023.

Classification model parameters and feature inputs

Cropland classification using SVM_RBF

Cropland classification using the SVM_RBF model requires tuning the kernel parameter and the penalty parameter to achieve optimal classification performance. As shown in Fig. 6, classification accuracy varies significantly across different and combinations, with cropland accuracy consistently exceeding the overall accuracy for cropland-non-cropland classification. Within the parameter ranges ∈[18,21] and ∈[10,20], the model exhibited stable performance, with cropland accuracy mostly above 0.94 and overall accuracy rarely below 0.90. After comparing all candidate combinations, =20 and =10 was selected as the final SVM_RBF parameters, yielding a cropland accuracy of 0.9461 and an overall accuracy of 0.9044, both approaching the respective optima. Neighboring settings (e.g., =18, =10; =20, =20) achieved similarly high accuracies, indicating robust performance. Compared with =21, =20, the choice of =10 also reduced overfitting risk and computational cost. Accordingly, all classification experiments based on the SVM_RBF model were conducted using the parameter configuration of =20 and =10; the corresponding model input features and parameter values are listed in Table 3. For a control experiment, a Random Forest (RF) model was separately trained and tested (features and parameters listed in Table 4) to compare its classification results with those of SVM_RBF.

Fig. 6.

Parameter calibration of the SVM_RBF classifier. OA represents the overall classification accuracy for cropland and non-cropland, while CA represents the accuracy specific to cropland.

Table 3.

Parameters and selected features for the SVM_RBF model.

Category	G	C
Parameters	20	10
Features	B02, B03, B04, B05, B06, B08, B11 NDVI, NDBI, MNDWI, EVI

Table 4.

Parameters and selected features for the RF model.

Category	Number of Tree	Bag Fraction	Features per Node	Min Leaf	Max Nodes
Parameters	390	0.57	3	1	80
Features	B02, B03, B04, B05, B06, B08, B11 NDVI, NDBI, MNDWI, EVI

Crop classification using RF

Within the hierarchical classification framework, after obtaining the optimal cropland distribution, crop classification was conducted within the cropland mask. Differences in phenological profiles and spectral signatures of wheat, maize, rice, and soybean were exploited to construct two sets of multi-temporal imagery: (1) imagery from March to mid-July for identifying wheat cultivation areas, and (2) imagery from April to early November for identifying maize, rice, and soybean cultivation areas. The parameter calibration process of the Random Forest (RF) model was conducted as follows. First, based on the number of feature variables and the sample size, the parameters Features per Node, Min Leaf, and Max Nodes were empirically set to 4, 1, and 140, respectively. Subsequently, two key parameters—Number of Trees and Bag Fraction—were further tuned, with Number of Trees ranging from 0 to 500 at an interval of 5, and Bag Fraction ranging from 0.01 to 0.8 at an interval of 0.01. For each parameter combination, model performance was evaluated through cross-validation using the ground-truth samples. Finally, by comparing classification accuracies across different parameter settings, the optimal RF model parameters and feature variable combinations for accurate crop identification were determined (Table 5).

Table 5.

RF model parameters and selected features for the classification of wheat, maize, rice, and soybean.

Category	Number of Tree	Bag Fraction	Features per Node	Min Leaf	Max Nodes
Wheat	320	0.53	4	1	140
Maize, Rice, and Soybean	360	0.58	4	1	140
Features	B02, B03, B04, B05, B06, B08, B8A, B11 NDVI, MNDWI, EVI, REI

To verify the effectiveness of the hierarchical classification framework in crop classification, the RF model was also applied directly to crop classification without the cropland mask, and the results were compared. Furthermore, to assess performance differences between the RF and SVM_RBF in crop classification, the above classification procedure was repeated using the SVM_RBF, with its optimal parameters and feature variables summarized in Table 6.

Table 6.

SVM_RBF model parameters and selected features for the classification of wheat, maize, rice, and soybean.

Category	G	C
Parameters	20	22
Features	B02, B03, B04, B05, B06, B08, B8A, B11 NDVI, MNDWI, EVI, REI

Classification results and accuracy assessment

Cropland

Notable differences in the spatial delineation of cropland were observed among the reference map, the RF classifier, and the SVM_RBF classifier (Fig. 7). The reference map exhibited well-demarcated parcels, reduced cropland fragmentation, and clear separation between cropland and adjacent land-cover classes (e.g., roads, residential zones, field paths, and parks) (Fig. 7A and D). The RF classifier showed misclassification in spectrally mixed zones (e.g., cropland–residential interfaces) and limited ability to resolve fine-scale features: field paths were frequently labeled as cropland, and portions of urban green space and forest were occasionally misclassified as cropland (Fig. 7C and F). By contrast, the SVM_RBF classifier produced more clearly delineated cropland boundaries and improved discrimination in mixed pixels, resulting in clearer separations between cropland and residential areas. Although very narrow field paths remained challenging to detect, SVM_RBF accurately identified roads and most field paths and substantially reduced misclassification of urban green space and forest patches relative to RF (Fig. 7B and E).

Fig. 7.

Cropland distribution in the Qingtongxia Irrigation District in 2023. (A) and (D) reference map; (B) and (E) cropland map from the SVM_RBF model; and (C) and (F) the cropland map from the RF model. Panels (D), (E), and (F) are enlarged views of the areas outlined by black rectangles in (A), (B), and (C), respectively. Maps were created using QGIS version 3.34 (https://qgis.org/).

To further quantify the accuracy of the reference and model-derived cropland maps, accuracy assessments were conducted for all three maps using validation samples (Table 7). The reference map achieved the highest accuracy scores, with PA of 96.81%, UA of 96.46%, and OA of 96.64%. SVM_RBF achieved PA of 95.39% and OA of 94.61%, whereas RF attained PA of 91.30% and OA of 89.77%. Overall, among the automated classification approaches, the SVM_RBF classifier outperformed the RF classifier.

Table 7.

Accuracy assessment for cropland maps in the Qingtongxia Irrigation District in 2023.

Classification Method	PA (%)	UA (%)	OA (%)	Kappa
Reference map	96.81	96.46	96.64	0.92
RF	91.30	88.24	89.77	0.78
SVMRBF	95.39	93.83	94.61	0.88

Note: PA is Producer’s Accuracy, UA is User’s Accuracy, and OA is Overall Accuracy.

Comparison of cropland classification results with three high-resolution land use/land cover products

The cropland maps derived from three high-resolution, publicly available land-use/land-cover products (ESRI-GLC10, ChinaCropland10m, and SinoLC-1), along with the SVM_RBF derived cropland map, were evaluated using complementary metrics. Spatial consistency and area-based agreement were quantified by comparison with the reference map (Fig. 8), whereas classification accuracy was assessed independently based on validation samples.

Fig. 8.

Spatial consistency assessment of multi-source cropland maps in the Qingtongxia Irrigation District. (A, E) Comparison between the SVM_RBF map and the reference map; (B, F) comparison between ESRI-GLC10 and the reference map; (C, G) comparison between ChinaCropland10m and the reference map; (D, H) comparison between SinoLC-1 and the reference map. Panels E-H show enlarged views of the areas outlined by black rectangles in A-D, respectively. Maps were created using QGIS version 3.34 (https://qgis.org/).

Although all products show reasonable agreement in large and homogeneous cropland areas (Fig. 8), differences become evident in heterogeneous environments such as urban peripheral areas and retired farmland near rivers. These discrepancies are likely related to the fact that existing public products are primarily designed for large-scale or global applications, where fine-scale heterogeneity is not the main focus. In urban areas, ESRI-GLC10 and ChinaCropland10m misclassified portions of parks and green spaces as cropland; SinoLC-1 mislabeled extensive build-up areas enclosed by irrigation as cropland; the SVM_RBF map mainly misclassified small, fragmented green patches. Overall, SVM_RBF performed comparatively better for built-up area identification. In the retired-cropland zones of the Yellow River corridor, the SVM_RBF map and ESRI-GLC10 showed some misclassification, ChinaCropland10m tended to label the area broadly as cropland, while SinoLC-1 accurately identified it as non-cropland. Overall spatial consistency of the SVM_RBF map, ESRI-GLC10, and ChinaCropland10m exceeded that of SinoLC-1. Despite SinoLC-1’s 1-m resolution, misclassification in built-up areas reduced its overall spatial consistency.

The accuracy of cropland maps from the three public products was evaluated using validation samples (Table 8). ChinaCropland10m performed best among the three products (OA = 90.32%, Kappa = 0.80), followed by ESRI-GLC10 (OA = 85.36%, Kappa = 0.70). SinoLC-1 showed lower performance (OA = 80.25%, Kappa = 0.61). The reference map achieved the highest scores across all metrics (OA = 96.64%, Kappa = 0.92). Although slightly below the reference map, the SVM_RBF map shows improved performance in this study area, particularly in handling fragmented cropland patterns and complex land-cover mixtures typical of irrigation districts. (OA = 94.61%, Kappa = 0.88).

Table 8.

Accuracy assessment of multi-source cropland maps in the Qingtongxia Irrigation District.

Datasets	UA (%)	PA (%)	OA (%)	Kappa
ESRI-GLC10	84.10	86.10	85.36	0.70
ChinaCropland10m	91.20	89.44	90.32	0.80
SinoLC-1	81.36	79.14	80.25	0.61
Reference map	96.46	96.81	96.64	0.92
SVMRBF map	93.83	95.39	94.61	0.88

Note: PA is Producer’s Accuracy, UA is User’s Accuracy, and OA is Overall Accuracy.

Integrating the area statistics with the accuracy metrics and using the reference map’s area as the benchmark indicates systematic overestimation of cropland extent by all other products. The reference area is 341,000 ha; the SVM_RBF map (374,000 ha) and SinoLC-1 (359,000 ha) exhibit relatively small positive biases—9.5% and 5.3%, respectively—consistent with higher areal fidelity. In contrast, ESRI-GLC10 (401,000 ha) and ChinaCropland10m (406,000 ha) overestimate by roughly 17–19%, plausibly because deep-learning classifiers mislabel portions of roads, forest, and urban green spaces as cropland. Although SinoLC-1 exhibits a small net area error, this conceals poor thematic accuracy because limitations in the source imagery lead to widespread misclassification of urban built-up areas as cropland and omission of numerous true cropland patches, with the opposing biases offsetting in area estimates but lowering the product’s accuracy compared with other 10-m products. More broadly, area totals are an inadequate proxy for accuracy because aggregated statistics fail to reveal pixel-level errors or local misclassification patterns.

These findings underscore the advantage of the reference map derived from large-scale manual visual interpretation for cropland classification, while also highlighting the challenges that existing regional and global land-use products may face when applied to complex irrigation-district environments, given their different design objectives and spatial scales. Overall, the SVM_RBF offers high accuracy and stability for automated cropland mapping and serves as a reliable substitute when high-precision reference maps are unavailable.

Crop classification

Two comparative experiments were conducted for crop classification. (i) Under a hierarchical framework, a cropland mask was first generated with SVM_RBF; crop classification within the mask then used either RF or SVM_RBF, yielding two combinations (SVM_RBF - RF; SVM_RBF - SVM_RBF; Fig. 9). (ii) Without the hierarchical framework (i.e., no mask), RF and SVM_RBF were applied directly (RF; SVM_RBF). Performance for the four dominant crops (wheat, maize, rice, soybean) in the Qingtongxia Irrigation District in 2023 was evaluated using validation samples (Table 9), supplemented by a confusion matrix for the SVM_RBF - RF (Table 10) to localize classification errors.

Fig. 9.

Spatial distribution of major crops in the Qingtongxia Irrigation District, 2023, derived under a hierarchical classification framework. Panels A and D show the reference crop map; B and E show the SVM_RBF- RF crop map; C and F show the SVM_RBF - SVM_RBF crop map. Panels D-F show enlarged views of the areas outlined by black rectangles in A-C, respectively. Maps were created using QGIS version 3.34 (https://qgis.org/).

Table 9.

Accuracy assessment of crop maps derived from different classification methods in the Qingtongxia Irrigation District.

Method	Evaluation Metric	Wheat	Maize	Rice	Soybean	Others	Overall
Reference map	PA(%)	92.42	96.34	84.12	88.63	76.12	OA:91.24
	UA(%)	95.83	93.28	89.47	91.24	74.32
	Kappa	0.92	0.94	0.83	0.88	0.75	Kappa:0.89
	F1-score	0.94	0.95	0.87	0.90	0.75	F1:0.89
SVMRBF-RF	PA(%)	83.87	88.46	91.49	68.42	34.78	OA:80.79
	UA(%)	91.23	92.01	95.56	54.17	28.57
	Kappa	0.81	0.86	0.92	0.78	0.74	Kappa:0.75
	F1-score	0.87	0.90	0.93	0.60	0.31	F1:0.81
SVMRBF-SVMRBF	PA(%)	80.43	86.17	89.34	79.41	75.31	OA:78.17
	UA(%)	86.17	82.54	92.43	74.12	68.42
	kappa	0.80	0.84	0.74	0.83	0.72	Kappa:0.71
	F1-score	0.83	0.84	0.91	0.77	0.72	F1:0.80
RF	PA(%)	87.12	78.43	90.12	61.34	73.18	OA:76.42
	UA(%)	82.54	83.17	86.34	68.27	69.12
	kappa	0.80	0.76	0.86	0.59	0.71	Kappa:0.70
	F1-score	0.85	0.81	0.88	0.65	0.71	F1:0.76
SVMRBF	PA(%)	84.32	77.11	88.43	72.14	68.17	OA:74.24
	UA(%)	79.22	82.34	84.17	66.42	73.14
	kappa	0.77	0.75	0.83	0.65	0.71	Kappa:0.65
	F1-score	0.82	0.80	0.86	0.69	0.71	F1:0.75

Note: PA is Producer’s Accuracy, UA is User’s Accuracy, and OA is Overall Accuracy.

Table 10.

Confusion matrix of the SVM_RBF-RF model for crop classification.

Actual Predicted	Wheat	Maize	Rice	Soybean	Others	UA(%)
Wheat	52	1	0	1	3	91.23
Maize	1	69	1	2	2	92.01
Rice	0	1	43	0	1	95.56
Soybean	1	1	0	13	9	54.17
Others	8	6	3	3	8	28.57
PA(%)	83.87	88.46	91.49	68.42	34.78	80.79

The reference map showed consistently high accuracy across most crops, with overall accuracy (OA) and Kappa of 91.24% and 0.89, respectively, and serves as a reliable benchmark. For wheat and maize, the hierarchical SVM_RBF - RF achieved producer’s accuracy (PA)/user’s accuracy (UA) of 83.87%/91.23% and 88.46%/92.01%, respectively, outperforming SVM_RBF - SVM_RBF (80.43%/86.17%; 86.17%/82.54%) and the flat single-models (RF: 87.12%/82.54%, 78.43%/83.17%; SVM_RBF: 84.32%/79.22%, 77.11%/82.34%). The F1-scores for wheat and maize under SVM_RBF - RF (0.87 and 0.90) also exceeded those of other methods, indicating superior stability and robustness for these two crops.

For rice, SVM_RBF - RF attained the highest PA/UA/F1 (91.49%/95.56%/0.93), showing slightly higher agreement with the validation samples than the reference map. However, this apparent improvement does not necessarily indicate higher actual accuracy and may instead reflect limitations in the validation framework. Specifically, the relatively small number of rice samples may increase the sensitivity of accuracy metrics to sample distribution, while the partial dependence between validation samples and the reference data—both derived from similar visual interpretation—may introduce bias in the assessment. In addition, SVM_RBF - SVM_RBF yielded a high PA (89.34%) but a comparatively low Kappa (0.74), also pointing to residual class confusion or imbalanced labels.

Soybean is the most challenging class, with all methods performing markedly worse than for the other crops. The confusion matrix (Table 10) reveals that the SVM_RBF - RF model, despite achieving the highest overall OA (80.79%) among model methods, showed substantial misclassification for soybean: only 13 out of 19 validation samples were correctly identified (UA = 54.17%), with the rest being confused mainly with “Others” (9 samples) and wheat/maize (2 samples). Within the hierarchical framework, SVM_RBF - SVM_RBF performed best for soybean (PA/UA/F1 = 79.41%/74.12%/0.77), exceeding SVM_RBF - RF (68.42%/54.17%/0.60) and the flat single-models (RF: 61.34%/68.27%/0.65; SVM_RBF: 72.14%/66.42%/0.69). This pattern likely reflects higher spectral similarity between soybean and other land covers, as well as class imbalance in the validation samples (only 19 soybean samples; Table 10).

In summary, hierarchical classification generally outperforms direct single-model classification, but the optimal combination is crop-dependent: SVM_RBF - RF is preferable for wheat, maize, and rice, whereas soybean benefits from SVM_RBF - SVM_RBF; overall, SVM_RBF - RF surpasses SVM_RBF - SVM_RBF on balance. OA and Kappa usually move in tandem, yet instances such as rice (high OA but lower Kappa) and soybean (low UA despite moderate PA) underscore the need to consider PA, UA, F1-score, and confusion matrices to avoid inflated accuracy assessment under class imbalance. Pre-delineating cropland with SVM_RBF reduces non-target confusion and, in most cases, substantially improves the accuracy and consistency of crop classification.

Feature importance and ablation analysis

To quantitatively evaluate the contribution of crop-specific phenological features to crop classification, an ablation experiment was conducted by comparing classification performance under different feature combinations. Specifically, two key features identified in this study—MNDWI for rice discrimination during the flooding stage and the near-infrared band (B08) for separating maize and soybean—were selectively removed or included to assess their individual effects. Three experimental settings were designed: (1) a full feature set (All); (2) removal of MNDWI; and (3) removal of B08. All experiments were conducted using the same SVM_RBF-RF classifier and parameter settings to ensure comparability.

The results (Table 11) show that removing MNDWI leads to a substantial decline in rice classification accuracy, with producer’s accuracy (PA) dropping from 91.49% to 74.32% (a decrease of 17.17%) and user’s accuracy (UA) from 95.56% to 89.74% (a decrease of 5.82%). The Kappa coefficient for rice also decreases from 0.92 to 0.72, while the F1-score falls from 0.93 to 0.81. This pronounced degradation highlights the critical role of MNDWI in capturing flooding-related spectral signals for accurate rice discrimination. In contrast, removing MNDWI has only minor effects on other crops, with overall accuracy (OA) decreasing from 80.79% to 76.43% and Kappa from 0.78 to 0.69.

Table 11.

Ablation experiment results of the SVM_RBF-RF model for crop classification.

Experiment Scheme	Evaluation Metric	Wheat	Maize	Rice	Soybean	Others	Overall
Full features (All)	PA(%)	83.87	88.46	91.49	68.42	34.78	OA:80.79%
	UA(%)	91.23	92.01	95.56	54.17	28.57
	Kappa	0.81	0.86	0.92	0.78	0.74	Kappa:0.78
	F1-score	0.87	0.90	0.93	0.60	0.31	F1:0.82
Remove MNDWI	PA(%)	83.12	87.64	74.32	67.31	34.56	OA:76.43
	UA(%)	91.11	91.98	89.74	54.02	28.32
	Kappa	0.79	0.85	0.72	0.76	0.73	Kappa:0.69
	F1-score	0.87	0.90	0.81	0.60	0.31	F1:0.75
Remove B08	PA(%)	82.46	76.92	90.88	52.63	31.12	OA:73.54
	UA(%)	90.87	86.44	95.47	46.42	25.74
	Kappa	0.78	0.74	0.91	0.65	0.69	Kappa:0.64
	F1-score	0.86	0.81	0.93	0.49	0.28	F1:0.72

Removing B08 substantially impairs the classification of maize and soybean. For maize, PA declines from 88.46% to 76.92% (a decrease of 11.54%), UA from 92.01% to 86.44% (a decrease of 5.57%), and F1-score from 0.90 to 0.81. For soybean, the impact is even more severe: PA drops from 68.42% to 52.63% (a decrease of 15.79%), UA from 54.17% to 46.42% (a decrease of 7.75%), and F1-score from 0.60 to 0.49. The Kappa coefficient for maize and soybean decreases from 0.86 to 0.78 to 0.74 and 0.65, respectively. These results confirm that the near-infrared band (B08) is essential for separating maize and soybean, which exhibit similar NDVI temporal profiles during key growth stages. The overall classification performance also deteriorates notably, with OA decreasing to 73.54% and Kappa to 0.64.

Discussion

Factors influencing classification performance

The classification accuracy is primarily influenced by the reliability of the training samples, the design of input features, and the issue of class imbalance^42,43.

With regard to training data, ESRI-GLC10 employs dense markup (i.e., vector-based delineation of land-cover classes) rather than single-pixel labels. This strategy incorporates richer spatial context and spectral information, although it may introduce a degree of subjectivity in the absence of extensive ground-based verification⁷. SinoLC-1 constructs “stable” supervision by integrating multiple 10 m land-cover products (FROM-GLC10, ESRI-GLC10, and ESA-GLC10) through pixel-wise consensus and uses these as training labels within the L2HNet framework¹⁰. Such approaches are designed to ensure large-scale consistency and generalization; however, uncertainties inherited from the source products may propagate into the final outputs, particularly in complex landscapes^6–8. In addition, differences in spatial resolution and spectral information between the source datasets (Sentinel-2 multispectral imagery) and the target high-resolution imagery (Google Earth RGB data) may introduce cross-scale and cross-sensor inconsistencies, which can affect the delineation of fine-scale cropland boundaries. Although the L2HNet framework incorporates strategies such as CAS (Confident Area Selection) and DVA (Dynamic Vague-area Adaptation) to mitigate spatial inconsistencies, residual discrepancies may remain in spectrally complex environments. Overall, these characteristics reflect the design priorities of existing large-scale land-cover products. In contrast, this study relies on field surveys and manual interpretation of high-resolution imagery at the irrigation-district scale, enabling better representation of local cropping patterns and thereby improving classification performance in this specific application context.

For input features, ESRI-GLC10 utilizes six spectral bands from Sentinel-2 (red, green, blue, NIR, SWIR1/2)⁷, enabling deep-learning models to capture rich spatial–spectral information. In contrast, SinoLC-1 is based on RGB imagery from Google Earth and does not include near-infrared or broader spectral information¹⁰. This difference in feature design may influence the ability to distinguish between cropland and other vegetation types, particularly in spectrally complex environments where phenological signals are important. As shown in Fig. 10D and F, both products exhibit some confusion between forest, parkland, and cropland. In irrigation districts, spectral differences between certain land-cover types can be subtle. For example, during the flooding period, rice fields and wetlands may exhibit similar spectral responses, while slight differences in near-infrared reflectance related to crop structure can aid discrimination. The absence of near-infrared information may therefore affect the separability of such classes. Similarly, in terraced agricultural areas, shadow effects may introduce additional uncertainty in RGB-based classification due to reduced contrast between vegetation and non-vegetation features. ChinaCropland10m integrates multiple feature types, including vegetation indices (EVI2), soil indices (DBSI), pigment indices (Cire, BRI), and SAR backscatter (VV/VH), to capture crop growth dynamics¹¹. While this approach enhances large-scale mapping capability, some confusion between cropland and other land-cover types, as well as blurred boundaries in heterogeneous areas, can still be observed (Fig. 10E). In comparison, the SVM_RBF-based approach in this study combines spectral bands and crop-specific phenological signals to better represent local conditions in the irrigation district, resulting in clearer delineation of cropland boundaries and improved identification of forests, park areas, and narrow roads, although some scattered cropland remains difficult to capture (Fig. 10C). The resulting cropland map achieved an overall accuracy of 94.61%, indicating improved performance in this specific irrigation-district application.

Fig. 10.

Detailed comparison of croplands from different products. Panel A shows a high-resolution image from Google Earth Pro (the green box highlights areas of forest and parkland), while Panels B-F were visualized and arranged using QGIS version 3.34 (https://qgis.org/).

Class imbalance, defined as unequal class representation, substantially impacts accuracy⁴². The class-frequency distribution of land cover is inherently skewed and minority classes are often spatially sparse/clustered, so models struggle to learn stable decision boundaries for under-represented classes, leading to degraded performance. In the study area, soybean occupies the smallest spatial proportion in the irrigation district, with relatively few samples. In addition to the spectral similarity between corn, soybean, rice and other crops, class imbalance directly contributes to lower classification accuracy of soybean, with the combined effect of both factors amplifying errors. To address this, class weighting strategies could be considered for improvement⁷.

ChinaCropland10m adopts a hierarchical framework to differentiate fallow and active cropland within a cropland mask. At the first level, cropland masks are derived from ESA-GLC10 and FROM-GLC10; at the second level, a knowledge-driven fusion of Sentinel-1/2 time series using the VSPS (Vegetation–Soil–Pigment–SAR) approach is applied to detect cropland dynamics without relying on training samples¹¹. While this design supports large-scale and sample-independent mapping, uncertainties in the initial cropland masks may influence subsequent classification results, particularly in heterogeneous agricultural environments. In this study, a cropland mask with relatively high accuracy was first generated using the SVM_RBF approach, which helps improve the representation of cropland boundaries and provides a more reliable basis for subsequent crop classification within the irrigation-district context.

Cropland classification accuracy at the irrigation district scale depends on the reliability of training samples, the design of input features, and class balance. Although knowledge-driven approaches and spectral indices can improve classification, misclassification of non-cropland areas remain, underscoring the need for high-precision cropland classification as a first step in crop classification. Integrating high-quality field samples, carefully selected features, and robust classifiers such as SVM_RBF can effectively enhance accuracy and reduce these errors.

Limitations and future work

Despite the promising performance of the proposed framework, several limitations should be acknowledged. First, the model was developed and evaluated using data from a single irrigation district and a single year, and its spatial and temporal transferability has not yet been fully validated. Future work should assess the robustness of the framework across different regions and multiple growing seasons. Second, the framework relies primarily on optical time-series features derived from Sentinel-2 imagery. Although indices such as NDVI, MNDWI, and red-edge bands capture key phenological characteristics, their effectiveness may be affected by cloud contamination or incomplete temporal coverage. The integration of multi-source data, such as SAR backscatter or texture features, could further improve classification robustness under challenging observation conditions. Third, variations in agricultural management practices and environmental conditions, such as irrigation scheduling, climate variability, water stress, or nutrient deficiency, may lead to intra-class variability in spectral and phenological signatures. While the use of ensemble classifiers (e.g., RF) improves model robustness to some extent, these factors may still introduce uncertainty in classification results. Finally, shifts in agricultural calendars due to climate change may alter the timing and shape of crop phenological trajectories. Incorporating adaptive temporal feature extraction or dynamic phenology modeling represents an important direction for improving the generalization ability of the framework.

Conclusions

A hierarchical SVM_RBF-RF framework implemented on Google Earth Engine (GEE) achieved accurate mapping of cropland extent and crop planting structures in the Qingtongxia Irrigation District. By integrating Sentinel-2 imagery time-series, ground truth data, and multi-source features, the approach reduced spectral confusion between cropland and non-cropland and enhanced crop-type discrimination in a complex irrigation district. The principal findings are as follows:

The hierarchical strategy—first delineating cropland extent using SVM_RBF and then identifying crop types with RF—proved effective in minimizing interference from non-target land covers. The inclusion of shortwave-infrared responses during the paddy flooding stage, as well as the distinct deep-green spectral signature of maize, further improved crop separability. Moreover, the construction of long NDVI time series on the GEE platform and the integration of multi-source features reduced reliance on high-resolution imagery and dense in-situ sampling, thereby enhancing computational efficiency.

The use of SVM_RBF for cropland-extent mapping achieved an overall accuracy of 94.61%, surpassing comparative products such as ESRI-GLC10, ChinaCropland10m, and SinoLC-1 in boundary delineation and area estimation. Within the hierarchical SVM_RBF-RF framework, the dominant crops (wheat, maize, rice, and soybean) were classified with consistently high accuracy. Notably, the OA for maize and rice exceeded 86%, representing an improvement of 5–10% over the direct (non-hierarchical) classification approach.

In the Qingtongxia Irrigation District, benchmarking against cropland maps extracted from publicly available LULC products using independent validation samples and standard accuracy metrics shows that ChinaCropland10m attains the highest accuracy (OA = 90.32%, Kappa = 0.80). ESRI-GLC10 ranks second (OA = 85.36%, Kappa = 0.70), while SinoLC-1 exhibits lower accuracy (OA = 80.25%, Kappa = 0.61). The comparatively weaker performance of ESRI-GLC10 and SinoLC-1 likely reflects the lack of ground-truth labels and reliance on inputs limited to raw spectral bands (i.e., without phenological time series), which exacerbates confusion with spectrally similar non-cropland classes (e.g., forest and grassland).

Overall, the results suggest that the hierarchical framework enables efficient and accurate mapping of cropland and crop types in the study area, demonstrating potential for application in similar complex agro-ecosystems pending further validation across multiple regions and years.

Author contributions

Wen Wang conceived and designed the study and supervised the study. Xiaoqian Xu performed the formal data analysis, curated the dataset, and drafted the original manuscript. Houjun He contributed to ground truth data collection and reference data preparation. Xiaoqiang Zhang contributed to data analysis and methodological validation. All authors discussed the results, revised the manuscript, and approved the final version.

Funding

This study was supported by the National Key Research and Development Program of China under Grant No. 2023YFC3209201.

Data availability

The cropland and crop classification datasets for the Qingtongxia Irrigation District (Ningxia, China; 2023) are publicly available at Zenodo (https://doi.org/10.5281/zenodo.17104444; Xu⁴⁴).The release comprises (i) point-based ground-truth crop samples in ESRI Shapefile format, with attributes for geometry/location, acquisition source (field survey or visual interpretation), and class labels; (ii) three 10-m cropland-extent maps derived from manual interpretation (reference cropland map), Random Forest (RF), and support vector machine with an RBF kernel (SVM_RBF); and (iii) three 10-m crop-type maps produced by manual interpretation (reference crop map) and hierarchical classification frameworks (SVM_RBF -RF and SVM_RBF -SVM_RBF). All maps are projected to the WGS 1984/UTM Zone 48 N coordinate system.

Declarations

Competing interests

The authors declare no competing interests.