Annual 30 m land cover dataset on the Tibetan Plateau from 1990 to 2023

Abstract

Accurate land cover data was fundamental for formulating sound land planning and sustainable development strategies. This study focused on the Tibetan Plateau (TP), a globally sensitive ecological area, and developed a locally tailored annual 30 m resolution land cover dataset from 1990 to 2023 (TPLCD). Leveraging the Google Earth Engine (GEE) platform for Landsat data processing, LandTrendr was employed to generate robust, high-precision training samples. Subsequently, random forest classification and spatiotemporal smoothing strategies were applied to precisely map the land cover dynamics of the TP. Rigorous validation through visual interpretation, authoritative third-party datasets (Geo-Wiki and GLCVSS), and thematic dataset cross-comparisons, revealed an overall accuracy of 84.8%, and a Kappa coefficient of 0.78, fully affirming the dataset’s high reliability. This dataset provided invaluable empirical evidence for understanding the vulnerability and adaptability of the TP’s ecosystem.

Background & Summary

Land cover, as a pivotal indicator for climate regulation and biodiversity conservation^1,2, highlighted the intricate and persistent interplay between global ecosystem changes and human activities³. Since the 1950s, global agricultural expansion⁴, urban sprawl^5,6 and other activities related to Land Use/Cover Change (LUCC) have increasingly influenced Earth system processes, including the surface energy balance^7,8, carbon cycle^9,10, hydrological dynamics^11,12, and biodiversity patterns^13–15. These changes have not only emerged as a primary driver of global ecological issues, but have also engendered significant uncertainties for regional ecosystems^16–18. Obtaining continuous and accurate land cover products at fine spatial scales is essential for formulating effective ecological restoration strategies¹⁹ and supporting the implementation of climate change adaptation measures^20–22. Remote sensing plays a central role in this field, enabling precise mapping of land cover dynamics^23,24.

Over the past several decades, a wide range of remote sensing sensors, including MODIS, SPOT, Landsat, and Sentinel-2, have been extensively applied. Among these, Landsat has been particularly prominent in the detection and analysis of LUCC²⁵ due to its long-term, freely accessible and continuous data record²⁶, providing critical evidence for revealing historical trends in Earth’s surface dynamics^27,28. Notable examples include FROM-GLC²⁹, GlobeLand30³⁰, and GLC_FCS30³¹. Concurrently, remote sensing data processing technology has undergone a profound paradigm shift from traditional workstations to cloud-based platforms. Google Earth Engine (GEE) epitomizes this trend³², offering users instant access to petabytes of satellite imagery and a diverse range of geospatial datasets, thereby enabling the generation of accurate annual land cover products on a global scale³³.

Traditional national or regional land cover maps were typically updated annually, consuming considerable time and resources while posing challenges for the high-frequency collection of large-scale remote sensing imagery³⁴. In light of these, some research has proposed classification expansion or generalization techniques to enable land cover classification without the need for annually updated training samples^35,36. For instance, Gong et al. utilized metrics such as spectral distance to measure the similarity of sample data between different years, allowing for the selection of unchanged samples in inter-annual transfer tasks, thus demonstrating the feasibility of stable classification with limited samples for global land cover mapping^37,38. Furthermore, continuous change detection schemes, such as Continuous Change Detection and Classification (CCDC)³⁹, Time-Series Classification approach based on Change Detection (TSCCD)⁴⁰, and Landsat-based detection of Trends in Disturbance and Recovery (LandTrendr)⁴¹ have been proposed. While CCDC algorithm can be executed on GEE, it was not well-suited for detecting more gradual land cover transitions unrelated to distinct disturbance events, such as vegetation regeneration^42,43. Similarly, the TSCCD method, which integrates Prophet and Dynamic Time Warping, demanded significantly higher computational resources⁴⁰. In contrast, LandTrendr, which can also run on the GEE platform, is more easily scaled to broad-area analyses⁴⁴, and effectively applied to time-series change detection in medium-resolution satellite imagery, primarily Landsat. Moreover, LandTrendr demonstrates the ability to efficiently and accurately track LUCC at large scales, adeptly handling extensive time-series datasets while detecting and identifying changes in land cover types. It maintains high classification accuracy and exhibits strong resistance to disturbances^41,45,46, showing robust performance in the presence of noise. Considering factors such as data storage requirements, computational efficiency, and scalability, LandTrendr offers a practical and reliable solution for mapping dynamic land cover changes with limited training samples.

The Tibetan Plateau (TP), renowned as the “Roof of the World” and the “Third Pole on Earth”⁴⁷, serves not only as a pivotal ecological security barrier for Asia but also as one of the highest and most environmentally challenging regions globally, with human activities dating back 40,000 to 160,000 years^48,49. In recent years, the dual impacts of climate change and human activities have triggered severe ecological issues, including grassland degradation and habitats fragmentation^50,51, particularly affecting ecosystems dominated by alpine meadows and alpine steppes^52,53. Nevertheless, existing datasets have not sufficiently delineated grassland systems in accordance with the unique ecological characteristics of the TP. Most studies have been confined to localized or typical areas^54,55, lacking a systematic understanding of LUCC across the entire plateau, thereby highlighting limitations in research perspectives.

The Second Comprehensive Scientific Expedition to the Tibetan Plateau has defined a more precise geographic boundary based on plateau elevation and mountain integrity, encompassing not only southwestern China but also extending into neighboring countries, ensuring geographic continuity of the plateau’s margin⁵⁶. Although the annual China Land Cover Dataset (CLCD)⁵⁷ provided long-term, annually updated land cover data, its coverage does not fully align with the entire geographical boundary of the TP. While global land cover datasets, such as GlobeLand30³⁰, offer basic land cover information for the region, their limitations in temporal resolution hinder in-depth analyses of long-term land cover dynamics. Furthermore, these datasets fail to adequately characterise the grassland-dominated ecosystem, impeding a deeper understanding of grassland ecological functions, degradation status, and restoration potential. Given the distinctive ecological characteristics of the TP and their vital role in global biodiversity conservation, it is imperative to develop a land cover classification system that distinguishes between different grassland types and provides precise, long-term monitoring of land cover dynamics across this integral geographical unit.

The study addressed the gap in high-resolution, long-term land cover monitoring of the TP. By leveraging the LandTrendr algorithm, a representative training dataset was developed. Utilizing GEE to process Landsat data, a series of spectral feature indices were derived, and a Random Forest (RF) classification model was employed to generate the Tibetan Plateau Land Cover Dataset (TPLCD), which provides annua land cover data for the TP from 1990 to 2023. The study pursued two main objectives: (1) to accurately monitor land cover dynamics on the TP over a long time series from 1990 to 2023 by applying a continuous change detection algorithm to construct representative training samples; and (2) to precisely assess the accuracy of the derived TPLCD using multi-source validation datasets. The findings provided highly credible data to analyse the current status of land cover on the TP, aiding policymakers to make informed, science-based decisions.

Methods

Study area

The TP, located between 25°59′-40°1′N and 67°40′-104°40′E, spans across nine countries, including China, India, and Nepal, with a total area of 3.08 million km². Approximately 2.58 million km² (accounting for 83.7% of the total area) lies within Chinese borders⁵⁶. The TP has an average elevation of around 4,320 m and exhibits a diversity of climates, with temperature and precipitation generally decreasing from the southeast to the northwest. The southeastern region enjoys a warm and humid climate, while the northwest is colder and drier, with an annual average temperature ranging from −6 °C to 20 °C and an annual average precipitation varying from 20 mm to 4,500 mm. In recent decades, the TP has faced a series of ecological crises such as grassland degradation and habitat loss⁵⁸, it is particularly urgent to conduct continuous and precise land cover monitoring, which is essential for capturing subtle ecological changes, providing timely warnings of potential risks (Fig. 1).

Fig. 1.

The study area of the research. (a) the elevation of the Tibetan Plateau, (b) geographical distribution of Landsat scenes used on the Tibetan Plateau, (c) number of Landsat scenes used on the Tibetan Plateau for each year from 1990 to 2023.

Overview

The workflow included training sample generation, feature variable input, RF classification, accuracy assessment, and dataset comparison (Fig. 2). This process primarily relied on the GEE platform. Initially, annual available Landsat images were collected, and key spectral indices were calculated. The LandTrendr algorithm was used to identify spectrally stable pixels, thereby gathering accurate and robust training samples. Subsequently, the RF algorithm was employed for annual classification, with the classification results smoothed and corrected using a spatiotemporal optimization strategy. Finally, the classification accuracy was evaluated by validation sets to produce a continuous and high-precision land cover maps for the TP from 1990 to 2023.

Fig. 2.

Research workflow for generating land cover products on the TP.

Landsat data processing

Landsat satellite imagery provided the longest time series of surface observations, with atmospheric corrections applied and cloud and shadow interference effectively removed using the CFMask technique⁵⁹. This study utilized Surface Reflectance data from Landsat 5 (L5), Landsat 7 (L7), and Landsat 8 (L8) spanning from 1990 to 2023, totaling 93,520 scenes, specifically 27,694 scenes from L5, 43,394 scenes from L7, and 22,432 scenes from L8 (Fig. 1). To construct annual composite images, a 12-month temporal window was employed, implementing a pixel-level median compositing strategy. This approach effectively eliminated extreme values caused by clouds and shadows, while also mitigating edge effects at the junctions of different Landsat scenes, ensuring high fidelity and accuracy in the annual composite images⁶⁰.

Classification system

The TP is predominantly covered by alpine grasslands, with alpine meadows and alpine steppes serving as typical components of this ecosystem⁶¹. Considering the unique regional attributes of the TP, the study developed a classification system comprising ten key land cover types: cropland, forest, shrubland, alpine steppe, alpine meadow, water bodies, bare land, impervious surfaces, wetlands, and snow/ice. This system not only subdivided the alpine grasslands ecosystem but also ensured compatibility and seamless conversion with international standardized systems, such as International Geosphere-Biosphere Programme (IGBP) (Table S1).

Feature variables input

In land cover mapping practices, integrating auxiliary environmental parameters and spectral indices can enhance classification accuracy^60,62. This study calculated eight spectral indices to optimize classification performance, including the Modified Normalized Difference Water Index (MNDWI), Enhanced Vegetation Index (EVI), Bare Soil Index (BSI), Soil Adjusted Vegetation Index (SAVI), and the Normalized Difference Built-up and Bare Index (NDBBI), Normalized Difference Water Index (NDWI), 95th percentiles of NDVI (maxNDVI) and 5th percentiles of NDVI (minNDVI). Furthermore, Shuttle Radar Topography Mission Version 3 elevation data provided by NASA-JPL was incorporated to accurately represent topographic features. By combining six basic spectral bands, eight spectral indices, and three topographic variables (DEM, slope, and topographic diversity), a total of 17 feature variables were constructed (Table S2). This comprehensive strategy not only maintained the integrity of the time series, but also effectively reduced data dimensionality and significantly minimized the interference of external factors.

Training samples generation

In large-scale land cover mapping tasks, the study adopted an innovative strategy to collect high-quality training samples by leveraging the advanced change detection technique—the LandTrendr algorithm—to construct training samples with both spatiotemporally robust and reliable. LandTrendr, a trajectory and spectral time series segmentation algorithm, has demonstrated strong effectiveness in detecting and analysing subtle land cover changes over time, making it highly efficient for distinguishing stable areas from changing ones^41,46. We applied LandTrendr to analyse three spectral indices sensitive to vegetation, water bodies, and non-vegetated hard surfaces: EVI, NDWI, and NDBBI. Specifically, the NDBBI was a crucial indicator for analysis of urban changes, excelling in highlighting the distribution of built-up areas and was highly sensitive for monitoring urban sprawl and assessing LUCC⁶³. Parameter configurations for the LandTrendr algorithm are detailed in Table S3.

By applying linear regression to the spectral indices of each pixel, deviations potentially caused by land cover changes were identified (Fig. 3), enabling the selection of regions that have remained stable over the long term, which formed the ideal source of our training samples. This method was efficiently implemented on the GEE platform, providing robust computational support for large-scale analysis. Based on these, LandTrendr was employed to precisely locate pixels with high temporal and spatial consistency, forming the foundation of the classification model. Within these preselected stable regions, polygon samples were uniformly extracted to balance the accuracy and representativeness of the training samples while maintaining computational efficiency. Pixel values derived from these samples were then used for classification training (Table S4).

Fig. 3.

Schematic representation of four typical spectral changes detected using the LandTrendr algorithm. (a) no abrupt change, (b) single change, and (c) two changes.

Classification and spatial–temporal consistency

Although spatiotemporally stable areas significantly enhanced the reliability of samples, they were not entirely error-free. Previous studies have confirmed the RF models exhibit high robustness to training samples^31,64, effectively supporting the use of samples derived from spatiotemporally stable regions for land cover classification. Consequently, we adopted the RF algorithm for land cover classification.

RF, proposed by Breiman, is a machine learning method that improves overall classification performance by constructing multiple decision trees and aggregating their outputs. Each tree is built using a random subset of the training data to reduce the risk of overfitting⁶⁵. Considering the memory limitations of the GEE platform, after several iterations of experiments, we determined that setting the number of trees to 200 satisfied the classification accuracy requirements while avoiding resource overload.

To enhance the robustness of the land use classification, the study adopted an annual repeated classification strategy, conducting 10 classifications per year. These results were subsequently mode-synthesised at the raster scale to produce the final map, thereby reducing the variability inherent in single classifications. Additionally, to improve the temporal consistency of the classification series, a sliding window technique was employed to process multi-year data, which updated the land use types of the central year based on time series patterns within the sliding window. By comparing the confusion matrices of the smoothed results with window lengths of 3 and 5 years, we chose the 5-year window that more effectively reduced noise and improved classification accuracy (Fig. S1).

Accuracy assessment

In the field of land cover classification, accuracy assessment is crucial for ensuring the reliability of classification results. This study employed two validation strategies to assess the reliability of the classification results: (1) Visual Interpretation of Google Earth Images: Using the same strategy as for obtaining training samples, we applied spatially uniform sampling to randomly select 7,000 samples for visual interpretation, aiming to evaluate the data quality from 1990 to 2023 (Fig. 4). (2) Use of Authoritative Third-Party Datasets: Geo-Wiki⁶⁶ and Global Land Cover Validation Sample Set (GLCVSS)⁶⁷, for additional accuracy verification. Geo-Wiki is a crowdsourced validation resource providing samples covering ten major global land cover types with imagery spanning from 2000 to 2012. Considering data reliability, we selected 1,667 samples located in the TP and marked as “sure”. GLCVSS was designed based on random sampling principles to ensure global spatial representativeness and uniformity, making it suitable for use across many years and large spatial scales. From this dataset, we selected 1,005 samples. This study utilized confusion matrices based on Geo-Wiki and GLCVSS to evaluate the classification accuracy of TPLCD for the periods 2000–2012 and post-2013, quantifying key metrics such as Producer’s Accuracy (PA), User’s Accuracy (UA), Overall Accuracy (OA), Kappa coefficient, and F1 score. Besides, using these three datasets, we conducted cross-comparison analyses of existing land cover products including GLC_FCS30D⁶⁸, CLCD⁵⁷, MODIS Land Cover product (MCD12Q1)⁶⁹, European Space Agency’s Global Land Cover product (ESA’s GLC)^70,71, and ESRI’ Global Land Cover data (ESRI’s GLC)⁷². To ensure the comparability of different datasets, all were mapped to a common classification system following the IGBP classification scheme (Table S5).

Fig. 4.

Spatial distribution of the validation samples in TP. (a) Samples from visual interpretation referring to Google earth images; (b) Samples from Geo-wiki and GLVSS, and the number of samples for land use type on the two platforms was indicated in the photos.

Datasets inter-comparison

In addition to existing annual land cover products, this study enhances the comprehensiveness of quality assessment by conducting detailed comparisons with multiple specialized datasets. Specifically, TPLCD was meticulously compared with the Global Surface Water dataset (GSW)⁷³, the Global Forest Change dataset (GFC)^74,75, and global impervious surface datasets (GISA 2.0⁷⁶ and GAIA⁷⁷). The GSW dataset (JRC Yearly Water Classification History, v1.4), covering the period from 1984 to 2021, features low error rates (omission error < 5%, commission error < 1%), ensuring high accuracy in water body identification. The GFC datasets included Hansen Global Forest Change v1.11 (2000–2023), focusing on areas with cover greater than 10%, and EC JRC global map of forest cover 2020, v1. These two datasets were used to validate the forest cover in TPLCD for the years 2000 and 2020, respectively. GISA 2.0 offered improved accuracy and stability compared to GISA 1.0; GAIA, for the first time, provided a 30 m resolution representation of dynamic artificial impervious surfaces from 1985 to 2018, with an average overall accuracy exceeding 90%. Based on these datasets, we integrated data on a spatial grid of 0.05° × 0.05° to create dot distribution maps showing the area covered by different datasets, and assessed the consistency and accuracy of TPLCD using linear regression and the coefficient of determination (R²)⁷⁸.

Data Records

TPLCD provides an annually updated land cover dataset for the TP from 1990 to 2023, with a spatial resolution of 30 m, using the WGS-84 as the spatial reference, and publicly available in GeoTIFF format on Zenodo⁷⁹. The pixel values range from 1 to 10, corresponding to the following land use types: 1 for cropland, 2 for forest, 3 for shrubland, 4 for alpine steppe, 5 for alpine meadow, 6 for water body, 7 for bare land, 8 for impervious surfaces, 9 for wetlands, and 10 for snow/ice.

Technical Validation

Accuracy assessment of TPLCD

Based on all available Landsat data on GEE, we generated the annually updated TPLCD from 1990 to 2023. To validate the accuracy of TPLCD, we first compared it with samples visually interpreted based on Google Earth imagery. The dataset demonstrated stable reliability, with OA ranging from 72.04% to 90.11% and an average OA of 79.07%. Additionally, the Kappa coefficient varied between 0.68 and 0.88, with an average value of 0.76. Forests exhibited the highest average F1 score (87.76%), followed by water bodies (86.54%), impervious areas (85.59%), and wetlands (83.24%) (Table 1).

Table 1.

Comparison of mapping accuracy of TPLCD with GLC_FCS30D, CLCD, MCD12Q1, ESRI’s GLC, and ESA’s GLC within the TP based on the visual interpretation samples.

Visual Interpretation of Google Earth Images
		Cropland	Forest	Shrub	Alpine steppe	Alpine meadow	Water	Barren	Impervious	Wetland	Snow/Ice	OA (%)	Kappa
TPLCD	PA (%)	80.67	90.14	70.71	79.94	72.85	93.48	66.28	84.32	84.49	80.46	79.07	0.76
	UA (%)	62.43	85.59	66.53	79.50	78.49	81.00	83.36	87.00	82.22	81.08
	F1 (%)	70.35	87.76	68.50	79.71	75.52	86.54	73.73	85.59	0.00	80.74
GLC_FCS30D		Cropland	Forest	Shrub	Grassland		Water	Barren	Impervious	Wetland	Snow/Ice	OA (%)	Kappa
	PA (%)	33.95	90.13	0.25	82.10		92.94	46.92	65.78	0.00	32.85	63.72	0.54
	UA (%)	41.21	64.91	2.59	55.68		80.52	63.26	95.36	0.00	91.50
	F1 (%)	39.57	76.21	0.00	66.91		92.12	55.41	82.76	0.00	54.48
CLCD	PA (%)	37.11	86.73	1.64	89.48		84.63	72.69	8.11	4.91	39.62	62.79	0.51
	UA (%)	56.82	80.85	27.72	53.75		96.13	60.35	95.18	99.03	91.44
	F1 (%)	44.84	83.65	3.10	67.15		89.95	65.93	14.71	9.21	55.23
MCD12Q1	PA (%)	31.32	64.93	0.25	77.31		61.00	81.75	49.54	0.35	13.44	56.07	0.43
	UA (%)	58.35	92.64	7.58	52.03		99.79	41.19	93.06	15.37	89.85
	F1 (%)	35.52	97.96	0.84	46.90		99.11	75.77	65.01	12.69	94.37
ESRI’s GLC	PA (%)	35.09	66.86	0.60	75.68		92.98	74.71	93.07	3.36	44.35	63.13	0.53
	UA (%)	76.40	88.16	42.71	54.75		92.62	45.73	86.59	52.15	81.39
	F1 (%)	48.07	76.04	1.17	63.53		92.79	56.71	89.70	6.27	57.04
ESA’s GLC	PA (%)	37.55	78.76	4.04	63.90		96.60	90.56	66.02	30.87	34.18	62.56	0.54
	UA (%)	80.65	81.65	23.88	60.50		92.84	38.98	98.26	92.62	85.78
	F1 (%)	51.24	80.17	6.91	62.15		94.68	54.50	78.97	46.31	48.88

To further validate the accuracy of TPLCD, we utilized Geo-Wiki to conduct a detailed comparison between TPLCD, GLC_FCS30D, CLCD, and MCD12Q1 from 2000 to 2012. The statistics indicated that TPLCD consistently exhibited higher classification accuracy, with OA ranging from 84.95% to 89.62% and Kappa coefficients between 0.74 and 0.82. The average OA and Kappa values were 86.31% and 0.77, respectively, significantly outperforming MCD12Q1 (average OA = 71.98%, Kappa = 0.49), GLC_FCS30D (average OA = 76.34%, Kappa = 0.57), and CLCD (average OA = 81.98%, Kappa = 0.66) (Table 2). Among the various land cover types, grasslands had the highest F1 score at 91.61%, followed by forests (90.07%), water bodies (81.79%), barren lands (75.96%), and impervious surfaces (75.38%). Furthermore, using GLCVSS, we evaluated the performance of TPLCD relative to GLC_FCS30D, CLCD, ESA’s GLC, ESRI’s GLC, and MCD12Q1 after 2015. During this period, TPLCD maintained the highest accuracy, with average OA and Kappa coefficients of 88.44% and 0.81, respectively (Table 3). Among the other datasets, ESA’s GLC performed second best with an average OA of 87.9% and a Kappa of 0.81, followed by GLC_FCS30D with an average OA of 86.66% and a Kappa of 0.74.

Table 2.

The comparison of the mapping accuracy of TPLCD with GLC_FCS30D, CLCD, and MCD12Q1 within the TP based on the Geo-wiki.

Geo-Wiki
		Cropland	Forest	Shrub	Grassland	Water	Barren	Impervious	Wetland	Snow/Ice	OA (%)	Kappa
TPLCD	PA (%)	73.93	94.19	61.04	93.44	95.14	65.54	61.54	66.67	73.60	86.31	0.77
	UA (%)	62.71	86.35	51.14	89.87	71.19	89.95	100.0	60.26	71.59
	F1 (%)	67.82	90.07	55.59	91.61	81.79	75.96	75.38	63.00	72.55
GLC_FCS30D	PA (%)	32.97	93.24	1.92	85.72	76.92	66.39	38.46	0.00	23.16	76.34	0.57
	UA (%)	30.91	70.02	7.14	83.99	63.38	55.23	30.00	0.00	88.19
	F1 (%)	31.89	79.98	3.03	84.84	69.40	60.30	0.00	0.00	36.67
CLCD	PA (%)	32.54	90.53	1.10	90.77	81.12	71.40	0.00	7.43	38.11	81.98	0.66
	UA (%)	50.00	81.01	5.68	85.93	70.66	69.71	0.00	92.62	89.77
	F1 (%)	39.34	85.49	4.79	88.29	75.51	70.53	0.00	94.56	52.75
MCD12Q1	PA (%)	33.77	62.57	0.00	81.22	34.21	83.55	50.00	0.00	8.98	71.98	0.49
	UA (%)	67.33	97.07	0.00	82.60	100.0	48.30	100.0	0.00	94.98
	F1 (%)	44.80	76.07	0.00	81.90	50.92	61.20	0.00	0.00	16.35

Table 3.

The comparison of the mapping accuracy of TPLCD with GLC_FCS30D, CLCD, MCD12Q1, ESRI’s GLC, and ESA’s GLC within the TP based on the GLCVSS.

GLCVSS
		Cropland	Forest	Shrub	Grassland	Water	Barren	Impervious	Wetland	Snow/Ice	OA (%)	Kappa
TPLCD	PA (%)	0.00	100.0	0.00	99.52	84.95	0.00	0.00	85.54	80.46	88.44	0.81
	UA (%)	0.00	68.22	0.00	85.77	69.95	83.82	0.00	0.00	99.23
	F1 (%)	0.00	81.07	0.00	92.13	72.21	84.36	0.00	0.00	91.88
GLC_FCS30D	PA (%)	0.00	100.00	0.00	99.19	85.71	90.91	0.00	0.00	82.51	86.66	0.78
	UA (%)	0.00	38.80	0.00	79.98	48.05	79.58	0.00	0.00	100.00
	F1 (%)	0.00	55.91	0.00	88.37	61.23	84.86	0.00	0.00	90.42
CLCD	PA (%)	0.00	0.00	2.52	99.73	66.67	84.02	0.00	0.00	75.73	83.63	0.74
	UA (%)	0.00	0.00	100.0	79.98	50.00	63.88	0.00	0.00	100.0
	F1 (%)	0.00	0.00	6.11	88.77	57.14	72.55	0.00	0.00	86.18
MCD12Q1	PA (%)	0.00	100.0	0.00	99.76	25.00	98.73	0.00	0.00	64.37	78.00	0.67
	UA (%)	0.00	81.36	0.00	83.44	100.0	34.95	0.00	0.00	100.0
	F1 (%)	0.00	89.72	0.00	90.88	40.00	51.62	0.00	0.00	78.32
ESRI’s GLC	PA (%)	0.00	100.0	0.00	96.98	75.00	87.76	0.00	50.00	84.76	86.65	0.79
	UA (%)	0.00	74.27	0.00	87.38	79.17	47.99	0.00	100.0	99.85
	F1 (%)	0.00	67.09	0.00	91.92	76.79	61.89	0.00	0.00	91.65
ESA’s GLC	PA (%)	0.00	100.0	0.00	95.38	75.00	97.40	0.00	0.00	86.58	87.90	0.81
	UA (%)	0.00	69.23	0.00	92.37	75.00	48.70	0.00	0.00	100.0
	F1 (%)	0.00	81.82	0.00	93.85	75.00	64.94	0.00	0.00	92.81

In summary, through visual interpretation verification and comparisons with third-party datasets, TPLCD has demonstrated superior OA compared to MCD12Q1, ESA’s GLC, CLCD, GLC_FCS30D, and ESRI’s GLC. TPLCD achieved an average OA of 84.8% and an average Kappa coefficient of 0.78. Notably, TPLCD offers a continuous time series spanning 34 years (1990–2023), providing more extensive temporal coverage compared to ESRI’s GLC (2017 to 2022), ESA’s GLC (2020 to 2021), and MCD12Q1 (2001 to 2022).

Comparison with other products

Through thematic comparison with GFC datasets, TPLCD-forest demonstrated a high degree of consistency with both Hansen-forest and JRC-forest. Specifically, the R² between TPLCD-forest and Hansen-forest was 0.88, while the R² with JRC-forest was as high as 0.96. These high values strongly validate the reliability and accuracy of TPLCD in forest classification (Fig. 5). The spatial distribution of TPLCD-forest also showed significant similarity to that of the two GFC datasets, visually confirming the quality of the TPLCD-forest and further highlighting its potential and application value in global-scale forest monitoring (Fig. 6).

Fig. 5.

Scatterplots of forest areas between the TPLCD and GFC in the TP in 2000 and 2020.

Fig. 6.

Forest map examples for TPLCD and GFC in 2000 and 2020.

This study also analysed the R² between TPLCD and GSW dataset concerning water body area at the grid scale from 1990 to 2021 to validate the accuracy of TPLCD. The results showed the R² between the two datasets fluctuated between 0.61 and 0.9 during 1990–1999, with the R² of 0.73 in 1990. Since 2000, the R² has shown a steady increase, reaching 0.98 in 2021 (Fig. 7), indicating a significant improvement in the consistency between the datasets. To further assess the accuracy of water bodies in TPLCD, we selected three representative lakes—Qinghai Lake, Namtso, and Zhari Namco—and analyzed their distribution maps for the four periods of 1990, 2000, 2010, and 2023. The findings indicated an expansion trend in the lakes on the TP (Figs. S2–4), which aligns with numerous existing studies^80–82. Moreover, further exploration in the spatial dimension confirmed the high consistency between TPLCD and GSW in depicting water distribution patterns, while also demonstrating the finesse of the GSW dataset in capturing the dynamic changes of seasonal and permanent water surfaces (Fig. 8).

Fig. 7.

Scatterplots of water body areas between the TPLCD and GSW during 2000 and 2020.

Fig. 8.

Examples of water body morphology for TPLCD and GSW in 1990 and 2021.

In recent years, with the acceleration of urbanization, the impervious surface area (ISA) on the TP has shown a significant growth trend. The study conducted a comparative analysis with existing 30 m resolution ISA products (GAIA and GISA 2.0) and used a linear regression model to evaluate the consistency between these datasets. The results indicated strong consistency between TPLCD-ISA and GAIA, with R² values ranging from 0.55 to 0.76 (Figs. S5, S6), while TPLCD-ISA showed a high spatial fit with both GAIA and GISA 2.0 in the urban areas of Xining and Srinagar, but GISA 2.0 failed to capture the dynamics of the city in Lhasa (Fig. 9). Notably, the average slope of TPLCD-ISA was higher compared to GAIA (0.57) and GISA 2.0 (0.3), suggesting that existing ISA data might be underestimating ISA changes. TPLCD-ISA provided a more accurate and detailed depiction of ISA changes, particularly in identifying the expansion of rural settlements in the complex terrain of the TP, showcasing its unique advantages. This is attributed to the limitations of GAIA and GISA 2.0 at a global scale, which do not fully capture the unique ISA dynamics of the region. Although the TP is not a mainstream region for global urbanization, its urbanization strategies hold a unique position due to their special mission of safeguarding national security and maintaining ecological security.

Fig. 9.

Comparative examples of development dynamics from the cities of Xining, Lhasa and Srinagar in TPLCD, GAIA and GISA 2.0 respectively.

Limitations and prospects

The study developed the TPLCD from 1990 to 2023 using Landsat data on the GEE platform, achieving an OA of 84.8%. Despite this, the dataset still had some limitations, primarily due to issues with the data sources. The L5 and L7 data suffered from uneven spatial and temporal coverage, early data acquisition limitations, and technical faults that affected data integrity, significantly impacting the accuracy and precision of the TPLCD. The research found that the misclassification rates of land cover types and the number of masked pixels before 2000 were significantly higher than those after 2000. Classification accuracy improved significantly after 2000, and the stability of the dataset strengthened since 2012. L5 had lost relay capability since 1992, making its data transmission dependent on receiving station^28,83, while L7, although technologically upgraded, suffered from a Scan Line Corrector failure in 2003, resulting in data stripe loss⁶⁰, severely impacting data acquisition and continuity analysis of the TP. Besides, the high altitude and variable weather conditions of the TP often resulted in cloud cover⁸⁴, and the wide field-of-view design of L5 and L7 further complicated data acquisition, affecting the coherence of the time series.

Indeed, while the LandTrendr algorithm has shown unique advantages in environmental monitoring and tracking land use changes, it also faces certain limitations. The 16-day revisit cycle of Landsat satellites results in a discontinuous observational data sequence⁸⁵, making LandTrendr potentially blind when detecting short-term events such as sudden forest disturbances. Cloud cover is another factor impacting the performance of LandTrendr, frequent cloud coverage can lead to gaps in the time series or anomalies in phenological characteristics due to a lack of valid data⁴⁵. Without access to complete and continuous data, LandTrendr may struggle to accurately capture trends within a region. Additionally, while LandTrendr effectively identifies the year in which a change occurs, its capability to detect changes at finer temporal scales, such as intra-seasonal variations, remains limited, which restricts its effectiveness in specific application scenarios.

Despite these inherent limitations of the data, TPLCD incorporated a locally tailored land cover classification system, enabling precise evaluation of grassland degradation and restoration dynamics. It successfully filled the gap in long-term, high-precision land cover data for the TP, a comprehensive natural geographic unit. Future research aims to address the limitations of long-term land cover change monitoring on the TP by optimizing monitoring efficiency and data quality from multiple dimensions. Multi-source satellite data, such as Sentinel imagery, will be integrated to compensate for the spatial resolution and specific period data gaps of the Landsat series. Moreover, advanced cloud detection algorithms and sub-cloud surface feature restoration techniques will be explored using deep learning technology to accurately process cloud-obscured areas and enhance data validity and purity. Furthermore, integrating ground truth data, meteorological variables, and socio-economic indicators will strengthen the accuracy of tracking and forecasting land cover dynamics.

Usage Notes

This study developed an annual land cover dataset, TPLCD, covering yearly updated land cover products from 1990 to 2023 at a 30 m resolution, including 10 land use types. Through multi-dimensional validation, TPLCD demonstrated high classification accuracy, with an OA of 84.8% and a Kappa coefficient of 0.78, and an average F1 score surpassing those of similar products. As the first land cover monitoring product for the TP covering a 34-year period and tailored to local characteristics, TPLCD deepened the understanding of the states and functional differences within grassland ecosystems through detailed differentiation of alpine grasslands, and successfully filled the gap in long-term, high-precision land cover data for this unique natural geographic unit. Not only does TPLCD provide invaluable resources for researchers studying ecological security issues on the TP and globally, but it also lays a solid foundation for formulating effective conservation strategies.

Author contributions

Siya Li: Methodology, Software, Data curation, Visualization, Writing – original draft. Quansheng Ge: Conceptualization, Project administration, Supervision. Fubao Sun: Conceptualization, Supervision Methodology, Project administration. Qiulei Ji: Methodology, Software, Writing – review & editing. Wenbin Liu: Supervision, Writing – review & editing. Ronggao Liu: Software, Visualization. Duanyang Xu: Visualization, Writing – review & editing. Zexing Tao: Supervision, Methodology.

Code availability

The TPLCD mapping algorithms were implemented in GEE. The map processing codes and produced TPLCD were available on zenodo⁷⁹.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41597-025-04759-6.