Spatial distribution dataset of 30-m resolution cropland in agricultural regions, Qinghai-Tibet Plateau

Abstract

This study developed a 30-m resolution annual cropland dataset spanning 1988–2024 to resolve the unstable data quality and high sample acquisition costs in mapping cropland distributions in two agricultural regions of the Qinghai–Tibet Plateau (QTP): the Hehuang Valley (HV) and middle basin of the Yarlung Zangbo River and its two tributaries (the Lhasa and Nianchu rivers; MBYZR and LNR, respectively). This dataset was generated using Landsat imagery and training samples derived from visual interpretation. An initial classification was conducted using a Random Forest classifier. To ensure the stability of training sample quality across time, a sample cleaning approach was applied annually, based on spectral consistency constraints, allowing for the temporal extension of samples. The dataset demonstrated high classification accuracy, whereas the MBYZR and LNR demonstrated better classification performance, reflecting strong stability and robustness. Both regions showed favorable results regarding precision and recall, validating this approach’s effectiveness in multi-temporal remote sensing classification. Therefore, this dataset provides critical support for cropland monitoring, food security assessment, and agricultural adaptation in QTP studies, offering a practical reference for time-series sample construction and transfer in remote sensing classification.

Background & Summary

Cropland is the most fundamental material basis for agricultural production; changes in its quantity and quality directly affect food production^1–3. Therefore, accurate cropland distribution data can help governments and relevant departments at all levels comprehend the spatial characteristics of cropland resources across different regions, providing scientific support for land management, resource allocation, and conflict resolution related to land use. Long-term time-series cropland datasets further enable spatiotemporal analysis and monitoring of changes in farmland, revealing regional differences and trends, as well as assessing the prospects of cropland balance. Such insights are crucial for food production forecasting, agricultural policy formulation, and ensuring food security⁴.

The Qinghai-Tibet Plateau (QTP), known as the “Third Pole of the Earth,” is the highest plateau in the world⁵. Owing to its harsh climate and complex topography, agricultural activities are primarily concentrated in deeply incised river valleys with steep slopes and rugged terrain. Cropland in this region is typically small in area and fragmented in distribution, making the acquisition of high-quality spatial data particularly challenging. Moreover, the agricultural regions of the QTP are not only ecologically sensitive on a global scale^6,7, but they also play a key role in sustaining regional food self-sufficiency. Spatiotemporal changes in cropland here impact regional agricultural productivity and are vital indicators of socio-economic and ecological shifts under the context of global change^8,9. Therefore, accurately mapping cropland in this highland region is foundational for understanding ecological and socio-economic transformations and is essential for the successful implementation of food security strategies.

Presently, cropland identification and monitoring primarily rely on spatial information technologies, especially remote sensing^10–12. Over the past 30 years, significant advances have been made in remote sensing data acquisition, analytical methods, and product development, greatly enhancing the efficiency and accuracy of land cover classification, including cropland, supporting refined land resource management while providing a robust data foundation for evaluating land use change and ecological impacts. Datasets, such as the 30-m resolution GlobeLand30 (GL) developed by the National Geomatics Center of China¹³, 30-m GLC_FCS30 produced by the Aerospace Information Research Institute of the Chinese Academy of Sciences¹⁴, 10-m ESA WorldCover (ESA WC) developed by the European Space Agency¹⁵, 10-m Sentinel-2 Land Use/Land Cover dataset provided by Esri¹⁶, 30-m China Land Cover Dataset created by Wuhan University¹⁷, and 10-m AIEC dataset developed by the AI Earth team at DAMO Academy¹⁸, all offer cropland distribution data at national or global scales. In addition to these land use/land cover (LULC) products, several cropland-specific datasets have been developed, including the 30-meter China Annual Cropland Dataset (CACD) from Ying Tu et al.¹⁹ and the 30-meter Global Food Security-Support Analysis Data (GFSAD) from NASA and partners²⁰. These cropland products are mostly generated using traditional machine learning or deep learning approaches, with resolutions of 30 or 10 m. They are generally high in accuracy and broad in coverage, with some offering multi-temporal versions, which support long-term monitoring. However, these products also present several limitations that complicate their practical application. Specifically, variations in classification schemes, data preprocessing workflows, training sample sources, and classification algorithms used in their production—combined with inherent uncertainties in Earth observation technologies—result in discrepancies among the cropland distribution outputs derived from different datasets^21–27. Additionally, factors such as atmospheric interference and sensor variability lead to inconsistent image quality across time, which introduces temporal fluctuations in classification accuracy even within the same dataset. These issues pose significant challenges for fine-scale agricultural analysis in heterogeneous environments such as the QTP.

At the regional scale, a number of cropland-related products have also been developed based on the specific characteristics of target study areas. In China, for example, most of these datasets focus on economically developed or agriculturally concentrated regions such as the Northeast Black Soil Region²⁸, the Loess Plateau²⁹, and the Yangtze River Economic Belt³⁰. In contrast, the development of cropland data products for special regions, such as the QTP, still lags behind, which may restrict regional application services. This limitation is manifested in the following aspects:

1. Due to the constraints of terrain and climatic conditions, agricultural practices on the QTP commonly exhibit alternating cropping and fallow cycles, biennial cultivation, and a mix of farming and grazing activities. These patterns result in cropland showing highly similar spectral characteristics to those of grassland and forest in remote sensing imagery²¹. As a result, existing remote sensing classification methods struggle to accurately distinguish cropland from other land cover types, leading to a scarcity of cropland data products in this region.

2. Most studies on cropland distribution in this region are based on secondary extraction from LULC products. However, existing datasets exhibit considerable uncertainty in the QTP region, with reduced mapping accuracy compared to low-altitude plain area like eastern China^31,32. These products often fail to capture the complex agricultural realities of the QTP; thus, their applicability remains highly uncertain.

3. Existing cropland datasets have limited spatial coverage, focusing only on certain agricultural regions, often lacking long-term temporal continuity^33,34, which makes it difficult to comprehensively reflect spatiotemporal patterns. This further limits their utility for monitoring the agricultural dynamics of cropland in the alpine agricultural areas of the QTP.

Additionally, producing long-term cropland maps requires extensive training samples³⁵. Traditional approaches, such as field surveys and visual interpretation, while accurate, are time-consuming and labor-intensive—unsuitable for large-scale, long-term mapping tasks³⁶. Meanwhile, using existing products for sample extraction risks propagating classification errors³⁷. Hence, there is an urgent need for mapping strategies that balance sample efficiency and classification accuracy while tailored to complex alpine environments.

To address these challenges, this study proposes a dynamic sample cleaning scheme based on limited initial samples combined with the Random Forest (RF) classifier. Implemented on the Google Earth Engine (GEE) platform³⁸ and utilizing medium- and high-resolution Landsat imagery, we generated a consistent, long-term cropland distribution dataset for the alpine agricultural areas of the QTP from 1988 to 2024. Through processes including initial sample interpretation, spectral and terrain feature analysis, and dynamic sample updating, we mitigated temporal accuracy degradation and developed the QTP-cropland (QTP-CL) dataset with a 30-m spatial resolution over 37 years.

The QTP-CL is characterized by its long temporal span and continuous spatial coverage. It provides a reliable data foundation for monitoring food security, examining land use change, and analyzing ecological responses on the QTP. This dataset holds significant scientific value and broad application potential.

Methods

Study area

The Hehuang Valley (HV) is located in the northeastern part of the QTP, encompassing a total area of approximately 26,000 km² (Fig. 1a,c). It comprises a section of the Yellow River Valley approximately 200 km in length and several valleys of the Huangshui River system. The region is predominantly characterized by high mountains, alluvial plains, hills, and tableland. Meanwhile, various water systems have eroded out numerous wide river valleys.

Fig. 1.

Overview of the study area. (a) Study area location, (b) the middle basin of the Yarlung Zangbo River and its two tributaries (the Lhasa and Nianchu rivers; MBYZR and LNR, respectively), and (c) the Hehuang Valley (HV).

The middle basin of the Yarlung Zangbo River and its tributaries (the Lhasa and Nianchu rivers; MBYZR and LNR, respectively) is situated in the southern part of the QTP (Fig. 1a,b). Its area is approximately 66,500 km². This region belongs to the mountainous corner wide valley area. Excluding a few canyons, there are a large number of wide and gentle floodplains, alluvial terraces, and alluvial fans on both banks of the river.

Owing to valley terrain, the HV, MBYZR and LNR have relatively low altitudes, satisfying the hydrothermal conditions required for alpine agriculture. These regions have become fertile land with concentrated populations on the QTP, supporting approximately two-thirds of the population in Qinghai and Xizang, China. However, because of rugged valley terrain with steep slopes and deep gullies, the distribution of cropland resources is fragmented and scattered. Therefore, accurate cropland distribution data are crucial for supporting agricultural development and regulation in these two major alpine agricultural areas, ensuring grain self-sufficiency and food security.

Workflow overview

The overall process comprised the following four key steps. (1) Feature Preparation: construction of annual spectral and topographic features from Landsat imagery and NASA Digital Elevation Model (DEM) data. (2) Initial Sample Construction: visual interpretation of multi-source imagery in 2020 to generate cropland and non-cropland samples. (3) Dynamic Sample Cleaning: interannual transfer and cleaning of samples based on statistical thresholds using NDVI and NIR indices. (4) Time-series cropland mapping: after sample cleaning, time-series cropland maps were generated using the selected features and RF classifier (Fig. 2).

Fig. 2.

Cropland classification workflow.

Feature preparation

Considering the requirements for time-series remote sensing imagery and operational periods of various satellites, this study ultimately selected Landsat 5 imagery for the period from 1988–2011, Landsat 7 imagery for 2012, and Landsat 8 imagery from 2013–2024 to develop cropland distribution datasets. Missing pixels in the Landsat 7 imagery were spatially interpolated using the “focal_mean” function in GEE.

To mitigate the effects of clouds and shadows, the Normalized Difference Vegetation Index (NDVI) was selected as the quality band for annual compositing, which proved effective for cloud removal³⁹. Additionally, to reduce interannual variation caused by differences in sensor types and illumination conditions, all cloud-free annual images were standardized using Z-score normalization⁴⁰ at the band level. The normalization was conducted as follows:

$$Z=\frac{X-\mu }{\sigma },$$ 1

where X is the original pixel value, μ is the mean, σ is the standard deviation of the image, and Z is the standardized value.

Elevation is a key factor influencing land cover distribution⁴¹. In this study, we used the 30-m DEM⁴² to derive elevation, slope and aspect layers.

Initial sample construction

Owing to the multifunctional nature of land cover, a single land feature may be classified into different categories based on its varying attributes. Additionally, the mapping standards used to acquire, represent, and generalize land characteristics varied according to research objectives, application requirements, and regional contexts⁴³. Currently, various land cover classification systems were developed across different countries and regions based on diverse data characteristics and classification principles to meet the demands of various application scenarios^44–47. Therefore, selecting a classification system that aligns with the actual conditions of the study area is essential. In this study, the cropland classification was based on the system adopted by GL¹³ to more accurately reflect the actual distribution of cropland in the region (i.e., land used for growing crops includes paddy fields, irrigated dry land, rain-fed dry land, vegetable plots, forage grass planting land, greenhouse land, and land mainly for growing crops with interspersed fruit trees and other economic trees).

The initial samples were categorized into cropland and non-cropland classes. Sampling was conducted based on a 1 km-grid covering the study area, with 10% of the grid centers randomly selected to form the sample set. Visual interpretation was then performed using multi-source remote sensing imagery from 2020, including Landsat 8 Operational Land Imager (OLI), Sentinel-2 Multi Spectral Instrument (MSI), and high-resolution Google Earth imagery. Figure 3 illustrates the spatial distribution of the resulting samples.

Fig. 3.

Sample points. (a) Sample points in the HV and (b) MBYZR and LNR.

In total, 1,639 sample points were collected in the HV, comprising 349 cropland and 1,290 non-cropland samples. In the MBYZR and LNR, 1,551 samples were collected, including 247 cropland and 1,304 non-cropland samples.

Dynamic sample cleaning

Feature selection is instrumental in land cover classification using remote sensing imagery, as it significantly affects classification accuracy and model performance^48,49. To better characterize land cover types and improve classification efficiency and accuracy, this study first integrated spectral features, texture features, and topographic factors within the GEE platform to construct a multidimensional feature set for cropland classification (Tables 1, 2).

Table 1.

Comparison of regional characteristic differences in the HV.

Feature	Mean		Standard deviation
	Cropland	Non-cropland	Cropland	Non-cropland
BLUE	−0.132	−0.015	0.395	1.022
GREEN	−0.115	−0.011	0.437	1.032
RED	−0.217	0.002	0.452	1.027
NIR	0.440	−0.038	1.040	0.957
SWIR1	−0.510	0.096	0.744	0.994
SWIR2	−0.432	0.061	0.653	1.017
NDVI	0.402	−0.016	0.727	1.006
ASM	0.016	0.017	0.025	0.028
Contrast	0.001	0.002	0.002	0.006
Correlation	0.390	0.410	0.099	0.107
IDM	0.108	0.119	0.096	0.100
Entropy	0.956	0.953	0.056	0.060
Elevation	2,684.077	3,180.114	285.012	561.795
Slope	10.983	20.578	7.596	10.810
Aspect	173.970	172.213	110.933	102.690

Table 2.

Comparison of regional characteristic differences in the MBYZR and LNR.

Feature	Mean		Standard deviation
	Cropland	Non-cropland	Cropland	Non-cropland
BLUE	−0.498	0.026	0.195	0.996
GREEN	−0.483	0.031	0.231	0.995
RED	−0.678	0.031	0.275	0.993
NIR	1.106	0.033	0.877	0.972
SWIR1	−1.064	0.079	0.537	0.979
SWIR2	−1.142	0.067	0.589	0.991
NDVI	1.910	−0.011	0.856	0.992
ASM	0.016	0.017	0.021	0.030
Contrast	0.001	0.007	0.003	0.020
Correlation	0.416	0.426	0.101	0.109
IDM	0.137	0.125	0.098	0.115
Entropy	0.955	0.956	0.053	0.062
Elevation	3,897.914	4,827.059	174.501	509.967
Slope	3.415	19.662	3.232	11.780
Aspect	162.889	175.308	100.526	103.967

Subsequently, based on the statistical characteristics of different features and classification accuracy achieved by various feature combinations (Table 3), feature selection was conducted. Texture features that had minimal impact on classification performance were excluded; spectral and terrain features were retained as input features.

Table 3.

Accuracy evaluation results of different feature combinations.

Features	SF	TF1	TF2	SF + TF1	SF + TF2	TF1 + TF2	SF + TF1 + TF2
HV	0.82	0.78	0.81	0.83	0.85	0.84	0.85
MBYZR and LNR	0.95	0.89	0.94	0.95	0.96	0.95	0.96

(SF = Spectral Features; TF1 = Texture Features; TF2 = Topographic Factors).

Additionally, NDVI and Near Infrared (NIR) were selected as key indices for sample filtering. Both indicators have demonstrated significant statistical differences between cropland and non-cropland samples in previous studies^50,51 and our preliminary experiments. These indices are strongly associated with vegetation cover and crop canopy properties⁵². Accordingly, the mean (μ) and standard deviation (σ) of NDVI and NIR were calculated separately. Under the assumption of a normal distribution, the μ ± σ interval covered approximately 68% of the initial samples. Therefore, a dual-threshold filtering criterion was applied, retaining only those samples whose NDVI and NIR values both fell within their respective μ ± σ ranges. This filtering strategy effectively excluded outliers and ensured that the retained training samples exhibited representative and consistent spectral characteristics.

Time-series cropland mapping

RF is a supervised classification method based on ensemble learning, which improves classification performance by constructing multiple decision trees and combining them with a voting mechanism. The core idea is to introduce data and feature randomness into the classification process. The advantages of RF include high classification accuracy, strong noise resistance, the ability to assess feature importance, and ability to manage high-dimensional data and missing values. In remote sensing classification, its ensemble nature and generalization ability make it an efficient classification choice, especially in scenarios involving multi-source feature fusion⁵³.

During cropland mapping, spectral and terrain feature values were first extracted from the samples. Then, based on the set threshold, the samples were cleaned to obtain the sample set for the target classification year. The sample set was then split into training and validation samples in a 7:3 ratio. The RF classifier, along with the “spectral + terrain” feature combination, was used to classify the images and generate interannual cropland distribution maps. Finally, the classification results were evaluated for quality based on the validation samples, ensuring the accuracy and reliability of the classification.

Data Records

The cropland distribution dataset developed in this study covers two major agricultural regions on the QTP: the HV in Qinghai and the MBYZR and LNR in Xizang, China. The dataset spans from 1988 to 2024, comprising 37 time-series data layers. All data are stored in a GeoTIFF format, managed under a unified geographic coordinate system (GCS_WGS_1984), with a spatial resolution of 30 m. Each product is a single-band binary raster, where a pixel value of 1 indicates cropland and 0 indicates non-cropland. The files are named in the format “CL_Region_Year” (e.g., “CL_HV_2020.tif”) to facilitate easy identification and use.

Geographical scope: 37°52′9″–35°0′33″ N, 100°52′3″–103°4′8″ E (HV) and 31°3′29″–28°8′42″ N, 87°4′55″–92°37′1″ E (MBYZR and LNR). Specific areas include the 18 county-level administrative units in Xining, Haidong, Haibei, Huangnan, and Hainan of Qinghai, China, as well as the 24 county-level administrative units in Lhasa, Shannan, and Xigaze of Xizang, China.

The dataset can be accessed from the National Tibetan Plateau/Third Pole Environment Data Center⁵⁴.

Technical Validation

Accuracy assessment

To ensure the reliability of the remote sensing-based cropland identification results, this study conducted systematic accuracy assessments in both the HV and MBYZR and LNR. The evaluation metrics included the Overall Accuracy (OA), Precision, Recall, and F1-score, covering classification results for the full time-series (Fig. 4).

Fig. 4.

Product accuracy assessment results in the HV and MBYZR and LNR. (a) F1-score, (b) Overall Accuracy (OA), (c) Precision, and (d) Recall.

From an overall perspective, both regions demonstrated consistently high classification performance, indicating strong credibility and stability of the results. In terms of the F1-score, the HV showed slightly lower values overall, yet still maintained scores above 0.70. This suggests that despite complex terrain and fragmented cropland distribution, the classification model effectively identified cropland areas. In contrast, the MBYZR and LNR achieved F1-scores of ≥0.80 in most years, reflecting better generalization and region adaptability of the model, which resulted in more stable and reliable classifications. Regarding OA, the HV remained stable between 0.80 and 0.85, whereas that of the MBYZR and LNR consistently exceeded 0.95, further confirming the high accuracy of the identification results. For Precision, the HV remained above 0.70 with no significant fluctuations. The MBYZR and LNR generally maintained Precision above 0.80, with only minor declines in several years, indicating more effective control of false positives. Recall in the HV was relatively weaker compared to other metrics, but still remained between 0.70 and 0.80 for most years, suggesting acceptable capability in detecting actual cropland. Meanwhile, the MBYZR and LNR maintained Recall above 0.80, despite a slight downward trend, demonstrating comprehensive and stable cropland detection performance.

Comparison with other products

Previous studies have demonstrated that ESA WC and GL perform relatively well in the QTP⁵⁵. To further evaluate the reliability of cultivated land identification, we compared the classification accuracy of ESA WC, GL, and QTP-CL in 2020. Table 4 lists the results for the HV, MBYZR and LNR.

Table 4.

Accuracy comparison in 2020 (HV/MBYZR and LNR) among ESA WC, GL, and QTP-CL.

Precision index	ESA WC	GL	QTP-CL
F1	0.679/0.728	0.781/0.797	0.790/0.910
OA	0.818/0.882	0.818/0.848	0.857/0.972
Precision	0.874/0.993	0.794/0.876	0.827/0.910
Recall	0.556/0.576	0.768/0.730	0.755/0.910

In the HV, QTP-CL achieved an F1-score of 0.790, outperforming ESA WC (0.679) and GL (0.781). Its OA reached 0.857, higher than both ESA WC (0.818) and GL (0.818). Particularly in terms of Recall, QTP-CL achieved 0.755, significantly exceeding ESA WC (0.556) and slightly surpassing GL (0.768), indicating a stronger capability for identifying complete cultivated land areas. In the MBYZR and LNR, QTP-CL exhibited even more impressive performance, with an F1-score of 0.910, significantly higher than ESA WC (0.728) and GL (0.797). Its OA reached 0.972, considerably higher than ESA WC (0.882) and GL (0.848). With both Precision and Recall at 0.910, the results indicate that QTP-CL provides high classification accuracy and minimizes omission errors, demonstrating strong robustness in this region.

Overall, the QTP-CL dataset outperforms existing publicly available products in typical alpine agricultural areas, particularly excelling in Recall. These findings highlight its superior ability to capture the full spatial extent of cultivated land and its potential as a reliable base dataset for cultivated land monitoring on the QTP.

These comprehensive assessments underscore the robustness and accuracy of the QTP-CL dataset across both time and space. Its consistent performance demonstrates suitability for long-term cropland monitoring and comparative land use analysis on the QTP, providing a valuable reference for future agricultural and ecological research in alpine environments.

Cropland area dynamics from 1988 to 2024

To validate the temporal reliability of the QTP-CL, we conducted a cropland change analysis focusing on interannual area variations and the temporal stability of cropland extent in the HV, MBYZR and LNR.

Annual cropland area was calculated for each year using the classified raster products. To assess temporal stability, a three-year sliding window approach (year i – 1, i, i + 1) was applied to detect changes in cropland status at the pixel level (Fig. 5).

Fig. 5.

Conceptual diagram for identifying cropland change types using a three-year sliding window. Each pixel is classified into one of four temporal categories based on its cropland status in three consecutive years: stable cropland (S), abandoned cropland (A), fallow cropland (F), and reclaimed cropland (R). Green pixels represent cropland (value = 1) and orange pixels represent non-cropland (value = 0).

Figure 6 displays the annual cropland area and stable cropland ratio in two regions. In the HV, the total cropland area fluctuated considerably, especially from 1995–2005. Although the total area recovered afterward, the overall trend continued to slightly decline. In contrast, the MBYZR and LNR maintained a relatively larger and more stable cropland area. However, a clear data gap was observed from 2001 to 2003 due to missing Landsat imagery in 2002, which affected classification continuity during this period.

Fig. 6.

Annual cropland area and stable cropland ratio from 1989 to 2024. (a) HV and (b) MBYZR and LNR. Note: data for 2001–2003 in the MBYZR and LNR are missing due to a lack of usable imagery in 2002.

An inverse relationship was observed between the total cropland area and proportion of stable cropland in the HV. Specifically, years with increased cropland area often corresponded to lower stability ratios, indicating that the additional cropland was likely composed of temporally unstable or fragmented parcels. This suggests that land use in the region maybe highly sensitive to climatic variability and land management practices. In contrast, the MBYZR and LNR exhibited a consistently higher and more stable proportion of persistent cropland, reflecting greater temporal consistency in both land use and classification results.

This cropland change analysis confirms the spatiotemporal reliability of the QTP-CL across two distinct high-altitude agricultural regions. By incorporating a sliding-window-based stability classification, the dataset captures long-term trends in cropland extent and reveals subtle land use dynamics, such as abandonment, fallow, and reclamation. This enhances the dataset’s value for agricultural monitoring, land use planning, and ecological assessment in complex and variable environments such as the QTP.

Usage Notes

The QTP-CL provides annual cropland distribution data from 1988 to 2024, covering two major agricultural regions on the QTP: the HV and MBYZR and LNR. This dataset distinguishes cropland from non-cropland using a robust classification framework based on RF classifiers, spectral-terrain features, and a dynamic sample cleaning strategy.

This medium-resolution dataset supports a wide range of applications, including:

• Regional agricultural monitoring and sustainable land use assessment.

• Evaluating land use change and policy impacts in alpine agricultural areas.

• Supporting food security research and climate adaptation strategies in fragile ecosystems.

Considering the increasing attention to environmental sustainability and agricultural stability in alpine areas, QTP-CL offers researchers, planners, and decision-makers a reliable resource to analyze long-term trends in cropland expansion, abandonment, and reclamation. The consistent temporal coverage and validated classification results make it valuable for understanding land dynamics in complex terrains.

The dataset is stored in a GeoTIFF format, with spatial reference GCS_WGS_1984 and a resolution of 30 m. It follows a standardized naming convention (“CL_Region_Year,” e.g., “CL_HV_2020.tif”) to facilitate efficient data retrieval and use. The dataset is particularly suited for applications in environmental change studies, regional planning, and precision agriculture in the QTP.

Limitations and prospects

The cropland data extracted in this study exhibited generally high classification accuracy, effectively capturing the spatial distribution of cropland and distinguishing it from artificial surfaces and bare land. However, in edge areas—particularly those adjacent to forest or grassland—the classification results showed some instability, primarily in the form of commission and omission errors (Fig. 7). These issues are likely caused by spectral confusion due to ambiguous land cover boundaries. In sloped farmland or valley areas, where vegetation types are often mixed, the classifier struggled to accurately differentiate between land cover classes, resulting in some forest or grassland areas being misclassified as cropland.

Fig. 7.

Local classification performance and typical misclassification patterns in edge transition areas.

Additionally, although the overall classification accuracy remained high throughout the time-series, a slight downward trend was observed in certain years. This decline may be attributed to variations in image quality, cloud coverage, and reduced availability of reliable training samples, particularly in the earlier years of the dataset. While the trend is relatively weak and does not substantially affect the usability of the dataset in most scenarios, it may pose limitations in applications that require extremely high temporal stability and precision.

To mitigate these issues, future improvements may include the integration of higher-resolution imagery, additional phenological indicators, or multi-temporal sampling to better distinguish complex edge conditions and enhance classification stability in heterogeneous landscapes. Moreover, incorporating long-term stable reference datasets or applying temporal domain adaptation methods may help maintain consistent classification accuracy across the entire time-series.

Author contributions

X.X. and S.L. designed the study; X.X. and S.L. drafted the manuscript; X.X., Q.C. and Y.P. contributed to revising the manuscript; and M.L. and M.Y. participated in data processing.

Code availability

The code is available at 10.5281/zenodo.15591002.

Competing interests

The authors declare no competing interests.