1-km Grid Dataset Development of Land Surface Human Activity Intensity of China (1980–2023)

CHAI Yi¹  ZHANG Rui^1,2  SHI Shengxiang^3,4  ZHOU Qiang¹  ZHAO Lei^1,2*  MIAO Peipei^1,2*

1.  Faculty of Geography, Yunnan Normal University, Kunming 650500, China;

2. Engineering Research Center of GIS Technology for West-China Resources and Environment, Ministry of Education, Yunnan Normal University, Kunming 650500, China;

3. State Key Laboratory of Remote Sensing and Digital Earth, Beijing Normal University, Beijing 100875, China;

4. Faculty of Geographical Science/Beijing Engineering Research Center for Global Land Remote Sensing Products, Beijing Normal University, Beijing 100875, China

Abstract: Harmonious coexistence between humans and nature is fundamental to the sustainable development of human civilization. However, continuous observations of land-surface human activity intensity at the national scale and over extended periods remain limited. In response to the major data requirements of coupled human-environment systems and global change research, this study used multiperiod land-use remote sensing monitoring data from China, comprising 12 epochs covering 1980–2023 and released by the Resource and Environmental Science Data Platform, as the baseline data. Relying on the Google Earth Engine (GEE) cloud platform, we integrated multisource feature variables, including long-term remote sensing imagery, meteoro­logical, and topographic features. By employing the random forest algorithm and spatiotemporal consistency constraint mechanisms, we reconstructed a long-term sequence of land use type data products for China from 1980 to 2023. On this basis, using the construction land equivalent coefficient method, we converted different land use types into a unified metric of human contri­bution, generating a 1-km spatial resolution dataset of the land-surface human activity intensity (HAI) index. The contents of this dataset include (1) 40 epochs of land-surface HAI data for 1980 and continuously from 1985 to 2023; (2) land-use transition matrix data at 5-year intervals from 1980 to 2023, along with code explanation files; (3) historical statistical data on China’s equivalent construction land area and HAI; and (4) the intermediate process data used for calculations. The dataset is archived in .tif, .xlsx, clr, and .txt formats, consisting of 93 data files with a total size of 1.93 GB (compressed into a single 307 MB file). This dataset supported the first author’s Master of Science thesis in Cartography and Geographic Information System.

Keywords: human activity intensity; construction land equivalent coefficient; sustainable development; China; Master of Science thesis

DOI: https://doi.org/10.3974/geodp.2026.05.02

Dataset Availability Statement:

The dataset supporting this paper was published and is accessible through the Digital Journal of Global Change Data Repository at: https://doi.org/10.3974/geodb.2026.06.01.V1.

1 Introduction

The profound impacts of anthropogenic activities on the Earth system mark the advent of the “Anthropocene”^[1,2]. As interactions within coupled human-environment systems intensify, accurately characterizing and measuring the spatiotemporal impacts of anthropogenic activities on the land surface has become a central focus of global environmental change and sustainability research^[3–5]. Although point-source or latent disturbances—such as industrial emissions, groundwater overextraction, and deep mineral exploitation—exert significant influences on ecosystems, land use and land cover change (LUCC) remains the most direct and spatially extensive core indicator of human occupation and modification of the land surface^[6,7]. Processes such as construction land expansion^[7], agricultural intensification^[8], and forest degradation and restoration^[9] intuitively reflect the intensity of human impacts on land-surface material cycles and energy flows.

As a crucial representation of human disturbance to natural systems, the land-surface human activity intensity (HAI) index provides an integrated, standardized metric for measuring the extent to which construction land expansion, land use structural transfor­mation, and spatial pattern evolution impose graduated pressures on ecological processes and environmental outcomes^[10,11]. Consequently, HAI holds fundamental value for terri­torial spatial planning, ecological civilization assessment, ecological redline effecti­veness evaluation, and monitoring progress toward China’s “dual carbon” (carbon peaking and carbon neutrality) goals.

Existing studies and products mapping human activity intensity largely follow 3 technical pathways. The first pathway involves the construction of composite indices from remotely sensed variables such as nighttime lights, population density, and impervious surfaces^[12,13]. While advantageous in terms of update frequency and spatial coverage, this approach is only indirectly linked to categorical land use planning. The second pathway overlays and weights socioeconomic factors such as population density, GDP, and road density^[13], which offers strong explanatory power but is highly susceptible to temporal inconsistencies in statistical definitions. The third approach employs equivalence weighting factors based on land use/cover datasets^[10]. This method aligns well with territorial manage­ment classification systems and associated policy implications, yet it is traditionally con­strained by the spatiotemporal resolution, classification consistency, and temporal continuity of the underlying data.

At the national scale in China, earlier products commonly suffer from discrete obser­vation periods, spurious changes induced by shifting classification schemes over time, a lack of annual continuity, and difficulties in achieving high-resolution, nationally consistent spa­tial representation. These limitations hinder the accurate characterization of rapid urbani­zation, industrial relocation, and large-scale ecological projects (e.g., the Grain for Green Program) since the 1980s and obscure the pronounced geographic gradients across the HU Huanyong Line. Recently, the accumulation of long-term remote sensing archives, the widespread adoption of cloud computing platforms, and the maturation of machine learning methods have created unprecedented opportunities for constructing long-term, cross-com­parable land-surface products^[14,15].

To address these scientific and practical demands, this study developed a spatiotemporally continuous HAI dataset for China covering the period from 1980 to 2023 by undertaking the following work:

(1) Utilizing the multiperiod land use remote sensing monitoring data of China as a baseline^[16], we leveraged the Google Earth Engine (GEE) cloud platform and a random forest classifier to construct a temporally consistent national land use product and applied cross-period consistency constraints to suppress classification noise and spurious changes. (2) Based on the built-up land equivalence coefficient method, we generated an annual, 1-km gridded HAI dataset (1980, 1985–2023), which was aligned with the territorial spatial management classification system under a unified metric. Compared with previous products that rely on single-source proxies or statistical overlays, this unified-metric dataset achieves a superior balance among indicator interpretability, policy relevance, and cross-period consistency.

2 Metadata of the Dataset

The metadata of the 1-km grid dataset of land surface human activity intensity of China (1980–2023)^[17]—including the dataset title, authors, geographic coverage, year of the data, temporal resolution, spatial resolution, data files, publication and sharing platforms, and data-sharing policy—are summarized at the Table 1.

Table 1  Metadata summary of the 1-km grid dataset of land surface human activity intensity of China (1980–2023)

3 Methods

3.1 Technical Workflow

The overall workflow for dataset construction is illustrated in Figure 1. First, we utilized 12 epochs of land use remote sensing monitoring data from China (1980, 1990, 1995, 2000, 2005, 2008, 2010, 2013, 2015, 2018, 2020, and 2023) released by the Resource and Environmental Science Data Platform, Chinese Academy of Sciences, as baseline datasets^[19]. To overcome the lack of dense, year-by-year historical ground truth samples, we adopted a temporally stable sample extraction strategy. By overlaying adjacent baseline land use maps separated by more than 5 years (e.g., 1990–1995) at the pixel level, we identified pixels with unchanged land use classes and treated them as high-confidence, temporally stable samples. Through stratified random sampling, we acquired approximately 1.95 million sam­ple points covering 26 land use classes. Second, leveraging the GEE platform, we syste­matically acquired surface reflectance imagery from the Landsat TM, ETM+, and OLI/ OLI-2 sensors for the period from 1980–2023 period (30-m spatial resolution). Utilizing quality assessment (QA) bands for cloud masking, we applied a median compo­siting algorithm over the vegetation growing season (May–October) to generate high-quality, cloud-free annual remote sensing mosaics.

For feature space construction, in addition to the six original Landsat multispectral bands, we calculated 4 core spectral indices representing vegetation growth, surface moisture, and built-up characteristics: the NDVI, NDWI, NDBI, and EVI. For the terrain variables, we incorporated the 30-m SRTM DEM from NASA to derive elevation, slope, and aspect. We further integrated core bioclimatic covariates from the approximately 1-km WorldClim global dataset (BIO1, BIO12, BIO4, and BIO15). Together, these variables formed a 17-dimensional base feature space for the pre-2000 data. After 2000, we integrated annual MODIS vegetation indices (MOD13A2 V6.1), expanding the feature space to 19 dimensions. Third, by employing the random forest (RF) ensemble learning algorithm, we partitioned the samples into training (70%) and validation (30%) sets, optimizing the hyperparameters via cross-validation. To guarantee temporal continuity and classification consistency, we intro­duced Savitzky-Golay temporal smoothing, cross-period classification standardization, and trend interpolation as spatiotemporal constraints, ultimately producing 40 epochs of annual land use maps.

Finally, following the construction land equivalent coefficient framework proposed by Xu, et al.^[10], we assigned equivalent coefficients to 26 land-use classes according to their relative levels of human disturbance, treating construction land as the reference (equivalent coefficient=1). We then computed the HAI index for each 1-km grid cell using an area-weighted averaging approach, thereby generating a spatially continuous HAI dataset for all of China (Figure 1).

Figure 1  Flowchart of the dataset development

3.2 Algorithm

3.2.1 Construction of Multisource Feature Space

To balance robustness for long-term computation with the richness of land-surface infor­mation, we constructed a dynamically dimensioned multisource feature space (Table 2). The base RF feature space comprises 17 variables: 6 original Landsat multispectral bands (blue, green, red, near-infrared, shortwave infrared 1, and shortwave infrared 2), 4 derived spectral indices, 3 SRTM-based topographic variables, and 4 WorldClim-based clima­tic variables. For 2000 and later, we incorporated annual composite vegetation index products (MOD13A2 V6.1) from the MODIS sensor, extending the feature space to 19 dimensions. Owing to its high temporal resolution, MODIS can effectively compensate for the limited observation frequency of Landsat 30-m imagery in cloud-prone regions, providing complementary multiscale information.

To obtain continuous, high-quality annual mosaics, we departed from conventional single- step compositing and adopted a multilevel image prioritization and cooperative gap-filling strategy. This method first prioritizes high-quality imagery within each year and applies median compositing at the pixel level. For local areas where effective observations were missing because of persistent cloud masking, we dynamically retrieved Landsat L2 surface reflectance products, rigorously masked clouds via QA bands, and performed median compositing for valid cloud-free pixels during the growing season (May–October) for gap filling. For a very small fraction of areas with extreme observation scarcity, we applied a constant zero-value masking. This approach ensured complete spatial coverage and consis­tent temporal series based on the foundational imagery.

Table 2  Variables included in the feature space for random forest classification

3.2.2 Random Forest Parameterization

Land use/cover classification is essentially a complex, nonlinear mapping problem in a high-dimensional feature space^[20]. In this study, we adopted the RF ensemble learning algorithm as the core classifier. As a decision-tree-based ensemble model, the RF introduces bootstrap aggregation of samples and random selection of split features at each node, effectively reducing overfitting. When dealing with LUCC, RF not only exhibits strong tolerance and robustness to highly collinear, multisource, high-dimensional data (spectral, terrain, climate, etc.) but also efficiently captures nonlinear decision boundaries among land-use types (e.g., paddy field vs. dry cropland and grassland with different cover levels) in complex spectral and phenological spaces^[21].

To balance classification accuracy with computational efficiency at the national scale and for long-term remote sensing data, we systematically tuned and strictly constrained the global RF hyperparameters: the number of decision trees was set to 500 to ensure model convergence and generalization; the maximum number of features for node splits was set to the square root of the total feature count; the minimum number of samples per leaf node was set to 1; the bootstrap sampling rate was 0.7; and the maximum node complexity was constrained to 128. These settings promote the stability and reproducibility of long-term classification predictions.

3.2.3 Construction Land Equivalent Coefficients and HAI Calculation

Following the theoretical framework of land surface-human activity intensity proposed by Xu, et al.^[10], we treat construction land—where the natural landform is completely trans­formed and human influence is dominant—as the reference (equivalent coefficient=1). The other land-use types are assigned coefficients according to their relative levels of anthro­pogenic disturbance (Table 3). The HAI for each 1-km grid cell is computed as an area-weighted average:

                                                   (1)

where  is the area of land-use category i within the grid cell,  is its equivalent coefficient, and  is the total area of the grid cell. The index ranges from 0 to 1, with higher values indicating greater human activity intensity.

Table 3  Construction land equivalence coefficients for land use/land cover classes

4 Data Results and Validation

4.1 Dataset Composition

The “1-km grid dataset of land surface human activity intensity of China (1980–2023)” includes the following components: (1) 40 epochs of continuous HAI data for 1980 and annually from 1985 to 2023; (2) land-use transition matrices at 5-year intervals from 1980 to 2023 and associated code description files; (3) annual equivalent construction land area and national-level HAI statistics; and (4) intermediate-process data used during HAI computation.

4.2 Data Validation and Uncertainty Attribution

Using the reserved 30% independent validation samples, we constructed confusion matrices and evaluated classification performance using overall accuracy (OA) and the kappa coefficient. The results show an OA of 86.92% and a kappa coefficient of 0.86. Spatial patterns of accuracy exhibit clear geographical regularities: the classification of lakes (95.03%) and forestland (91.98%) is highly reliable. The accuracy loss is concentrated mainly in rural residential land within the construction land category (F1 score of approximately 32.27%). This limitation is rooted in the “mixed-pixel” effect under specific geomorphological conditions: in the extensive mountainous areas of Southwest China and the fragmented rural settlements of the Huang-Huai-Hai Plain, rural homesteads are highly scattered and intermixed; at a 30 m spectral resolution, they easily blend with surrounding croplands or forest canopies. In addition, in the desert-grassland ecotone in western China, slope shadows and other terrain effects further blur the boundaries between bare land and low-coverage grassland (OA<80%).

Despite these confusions in certain fine-scale land-use subclasses, aggregation to the 1-km grid and area-weighted conversion substantially reduce their impact on HAI estimates. Overall, the resulting HAI patterns are consistent with known geographical gradients and socioeconomic development trajectories, and are therefore considered robust and reasonable for macroscale applications.

4.3 Data Result Analysis

4.3.1 Spatiotemporal Characteristics of Land use Transitions in China

The spatiotemporal patterns of land use transitions over the past four decades (1980–2023) was illustrated at the Figure 2, it covers multiple subperiods with a cumulative transition area of approximately 2,312,772 km², or approximately 24.4% of the national land area. Four dominant transition flows characterize the core features of land use change: cropland to construction land (approximately 289,649 km², 12.5% of total transitions), grassland to forestland (approximately 214,495 km², 9.3%), cropland to forestland (approximately 222,552 km², 9.6%), and forestland to cropland (approximately 242,127 km², 10.5%). Net construction land area increased by 2.7 percentage points, from 1.5% of the national land area in 1980 to 4.2% in 2023, whereas cropland decreased by 1.8 percentage points; grassland and unused land declined slightly, by 0.7 and 0.6 percentage points, respectively.

From a temporal perspective, the overall transition rate was relatively low (approximately 15%) from 1980 to 2000, peaked from 2000 to 2015 (approximately 20%), and remained high from 2020 to 2023, while construction land expansion continued to be substantial (approximately 19%). Overall, land use underwent a transformation from agricultural and natural dominance toward an urban-ecological codominance pattern, with pronounced spatial heterogeneity. In eastern China, construction land expansion was dominant, accoun­ting for approximately 45% of all transitions, especially in major urban agglomerations such as the Yangtze River Delta and Pearl River Delta. In western China, mutual transitions between grassland and unused land were prominent (approximately 12%), driven by the vulnerability of dry climates and complex terrain, where grassland tends to degrade into unused land. Moreover, reforms and urbanization led to path dependence, supporting the conversion of cropland to construction land.

The socio-political mechanisms exhibit two-way coupling: ecological restoration programs such as Grain for Green provide negative feedback that promotes forest recovery (forest gains accounting for approximately 15% of transitions from 2000 to 2010), whereas urbanization exerts positive feedback that challenges the protection of cropland “red lines”. The overall pattern is scale-dependent: at the national scale, transition shares among land- use categories change only slightly, and the system appears relatively stable; however, at regional scales, changes can be dramatic, with ecological degradation-dominated trajectories in parts of western China and rapid expansion of construction land in the east, exemplifying the principle of human dominance in land-surface processes in the Anthro­pocene.

Figure 2  Maps of the land use/land-cover transitions patterns in China during 9 periods (1980–2023)

4.3.2 Spatiotemporal Change of Land Surface Anthropogenic Activity Intensity in China

The significant spatiotemporal changes in China’s HAI between 1980 and 2023 was illustrated at the Figure 3. The national mean HAI increased from 0.18 in 1980 to 0.37 in 2023, a growth of 105.6%. Spatially, high-value HAI zones are concentrated in the eastern coastal regions (e.g., the Yangtze River Delta and Pearl River Delta) and key northern urban clusters (e.g., Beijing-Tianjin-Hebei), reflecting the imprint of urbanization. The western plateau and northern grassland regions are characterized by low HAI, but localized hotspots are emerging around mining areas and major infrastructure corridors.

In terms of temporal evolution, 1990–2010 was a period of rapid growth, with an average annual HAI increase of approximately 2.5%, corresponding to the peak of industrialization and urbanization under the reform and opening-up policy. After 2010, growth slowed to approximately 1.2% per year, largely attributable to the strengthened implementation of ecological protection policies.

The evolution of HAI follows a “point agglomeration-corridor expansion-polarization intensification” pattern. From 1980 to 2000, the HAI increased mainly through broad-scale expansion, with low-intensity areas shrinking by approximately 60,097 km², primarily in the eastern monsoon region plains. From 2000 to 2020, rapid urbanization drove medium-high HAI expansion, with high-intensity areas increasing by more than 4,500 km² per year and

Figure 3  Distribution maps of HAI class composition in China

extending into central hilly regions and major river valleys. From 2020 to 2023, patterns underwent rapid restructuring, with high-intensity areas surging by approximately 192,897 km² and an average annual growth rate of 6.43%, forming a “string of pearls” like network of urban clusters.

This pattern is strongly shaped by geographic fundamentals. East of the HU Huanyong Line—occupying approximately 45% of China’s land area but hosting approximately 94% of its population—HAI is highest, while western high-mountain and desert regions exhibit the lowest intensity, with a clear gradient of attenuation from coast to inland. The observed evolution reflects the interplay between environmental determinism and spatial polarization theory: biophysical conditions in eastern plains (including fertile soils and dense river networks) underpin the transition from low to medium HAI (approximately 32% of transi­tions), whereas human geographic factors, such as population concentration in urban clusters, such as Beijing-Tianjin-Hebei, drive the increase from medium to high HAI (approximately 40%). These patterns highlight geographic trade-offs in socioecological processes.

5 Discussion and Conclusion

The dataset developed in this study fills an important gap in long-term HAI products for China, providing annual continuity and explicit alignment with land-use policy frameworks. Nonetheless, uncertainty remains, mainly arising from limitations in remote sensing spatial resolution and the subjective nature of construction land equivalent coefficients. Future work can integrate higher-resolution imagery, more advanced machine-learning methods, and richer multisource socioeconomic data to further improve classification accuracy and broaden the application scope. The dataset can serve as a foundational input for assessing progress toward the Sustainable Development Goals (SDGs), evaluating the effectiveness of ecological red-line policies, and modeling coupled human–environment systems.

5.1 Uncertainties in the Classification Framework and Scale Effects

Large-scale Earth observations inevitably entail methodological uncertainties. Owing to the limited effective coverage of Landsat TM imagery before 2000—particularly in some border regions of western China—this study relies on machine-learning models that integrate climate and terrain variables supplemented by temporal logic filters to predict land-use classes and fill gaps. This approach may reduce sensitivity to some extent to localized, abrupt land-cover disturbances. In addition, while the 1 km spatial resolution is well suited to national-scale ecological assessments, it inevitably introduces homogenization errors when the extremely fragmented human-environment landscapes of Southwest China’s karst mountains are represented through upscaling.

5.2 Limitations of the Physical Representation of HAI

The HAI derived from LUCC-based equivalent conversion essentially measures the intensity of human-induced alteration of land-surface material and energy fluxes. This theoretical framework cannot yet capture multidimensional “hidden” human footprints such as industrial chemical pollution, atmospheric aerosol forcing, deep groundwater depletion, or the disruption of ecological connectivity due to habitat fragmentation. Looking ahead, frontier research in Earth system science may combine this two-dimensional LUCC-derived indicator with nighttime light radiance, carbon emission inventories, mobile phone signaling, and other big-data proxies of socioeconomic processes to construct a more comprehensive representation of human activity footprints.

This dataset supported the first author’s Master of Science thesis in Cartography and Geographic Information System.

Author Contributions

Chai, Y. designed the overall framework and led the development of the dataset. Chai, Y. and Zhang, R. jointly wrote the data paper. Shi, S. X. and Zhou, Q. collected and processed the remote sensing data. Zhao, L. and Miao, P. P. reviewed and revised the dataset and manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

References

[1]        Crutzen, P. J. Geology of mankind: the anthropocene [J]. Écologie & Politique, 2007, 34(1): 141–148.

[2]        Zalasiewicz, J., Waters, C. N., Williams, M., et al. When did the anthropocene begin? a mid-twentieth century boundary level is stratigraphically optimal [J]. Quaternary International, 2015, 383: 196–203.

[3]        Mottl, O. E., Flantua, S. G. A., Bhatta, K. P., et al. Global acceleration in rates of vegetation change over the past 18,000 years [J]. Science, 2021, 372(6544): 860–864.

[4]        Zuo, L. J., Zhang, Z. X., Carlson, K. M., et al. Progress towards sustainable intensification in China challenged by land-use change [J]. Nature Sustainability, 2018, 1(6): 304–313.

[5]        Yang, Y. S. The impacts of global environmental changes on typical ecosystems: status, challenges and trends [J]. Acta Ecologica Sinica, 2017, 37(1): 1–11.

[6]        Fu, B. J., Zhang, L. W. Land-use change and ecosystem services: concepts, methods and progress [J]. Progress in Geography, 2014, 33(4): 441–446.

[7]        Kuang, W. H., Liu, J. Y., Dong, J. W., et al. The rapid and massive urban and industrial land expansions in China between 1990 and 2010: a CLUD-based analysis of their trajectories, patterns, and drivers [J]. Landscape and Urban Planning, 2016, 145: 21–33.

[8]        Yuan, J. T., Chen, W. X., Zeng, J. Spatio-temporal differentiation of cropland use change and its impact on cropland NPP in China [J]. Journal of Natural Resources, 2023, 38(12): 3135–3149.

[9]        Williams, B. A., Beyer, H. L., Fagan, M. E., et al. Global potential for natural regeneration in deforested tropical regions [J]. Nature, 2024.

[10]     Xu, Y., Sun, X. Y., Tang, Q. Human activity intensity of land surface: concept, method and application in China [J]. Acta Geographica Sinica, 2015, 70(7): 1068–1079.

[11]     Duan, Q. T., Luo, L. H. Summary and prospect of spatialization method of human activity intensity: taking the Qinghai-Tibet Plateau as an example [J]. Journal of Glaciology and Geocryology, 2021, 43(5): 1582–1593.

[12]     Wu, B., Song, Z. C., Wu, Q. S., et al. A vegetation nighttime condition index derived from the triangular feature space between nighttime light intensity and vegetation index [J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 1–15.

[13]     Xing, X. Y., Huang, Z., Cheng, X. M., et al. Mapping human activity volumes through remote sensing imagery [J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13: 5652–5668.

[14]     Fu, D. J., Xiao, H., Su, F. Z., et al. Remote sensing cloud computing platform development and Earth science application [J]. National Remote Sensing Bulletin, 2021, 25(1): 220–230.

[15]     Liang, J. T., Chen, C., Sun, W. W., et al. Spatio-temporal land use/cover change dynamics in Hangzhou Bay, China, using long-term Landsat time series and GEE platform [J]. National Remote Sensing Bulletin, 2023, 27(6): 1480–1495.

[16]     Liu, J. Y., Ning, J., Kuang, W. H., et al. Spatio-temporal patterns and characteristics of land-use change in China during 2010–2015 [J]. Acta Geographica Sinica, 2018, 73(5): 789–802.

[17]     Chai, Y., Zhang, R., Shi, S. X., et al. 1-km grid dataset of land surface human activity intensity of China (1980–2023) [J/DB/OL]. Digital Journal of Global Change Data Repository, 2026. https://doi.org/10.3974/ geodb.2026.06.01.V1.

[18]     GCdataPR Editorial Office. GCdataPR data sharing policy [OL]. https://doi.org/10.3974/dp.policy.2014.05 (Updated 2017).

[19]     Xu, X. L., Liu, J. Y., Zhang, S. W., et al. Dataset of remote sensing monitoring of land use in China over multiple periods (CNLUCC) [DS]. Resource and Environmental Science Data Platform, 2018. DOI: 10.12078/2018070201.http://www.resdc.cn/.

[20]     Feng, Q. L., Niu, B. W., Zhu, D. H., et al. Review for deep learning in land use and land cover remote sensing classification [J]. Transactions of the Chinese Society for Agricultural Machinery, 2022, 53(3): 1–17.

[21]     Ma, H. J., Gao, X. H., Gu, X. T. Random forest classification of Landsat 8 imagery for the complex terrain area based on the combination of spectral, topographic and texture information [J]. Journal of Geo-Information Science, 2019, 21(3): 359–371.