A multi-source data fusion method for land cover production: a case study of the East European Plain

ABSTRACT

Large-area high-precision land cover mapping faces challenges such as a lack of uniform classification systems and the inability to compare different products. The current use of deep learning methods in land cover data product generation provides opportunities to address these issues. However, this requires the creation of many manually labeled samples, and this involves high time and labor costs. Therefore, research is being conducted to examine methods for producing land cover products by integrating multiple data sources. This study focuses on the East European Plain and is based on land cover types that include water, forest, grass, wetland, crop, shrub, built area, bare area, ice, and tundra. Label images were fused using data from Dynamic World, ESA World Cover, ESRI Global LULC, GlobeLand30, and Open Land Map. Using a modified Dynamic World model, predictions for the East European Plain for 2022 were made, ultimately resulting in a land cover product at 10 m resolution. Compared to Dynamic World data, the classification system of this dataset aligns with the land cover conditions of the study area. The dataset possessed higher accuracy. This method integrates the advantages of existing data products, automates the generation of training labels, and effectively reduces manual costs.

KEYWORDS:

• Land cover
• East European Plain
• multi-source data fusion
• deep learning

1. Introduction

Land cover refers to the natural or artificial elements that cover the surface of the Earth, including water bodies, vegetation, buildings, sandy areas, and bare soil. Different regions (latitude, longitude, country, altitude, and climate type) exhibit variations in surface elements, and the generation of land cover products can reflect surface cover conditions and differences in these areas. Land cover data can be used for research examining changes in water resources (Aghsaei et al. 2020; Nie et al. 2011), geology and landslides (Ghaderpour et al. 2024; Quevedo et al. 2023), land degradation and desertification evolution (Jafari and Hasheminasab 2017; Ying et al. 2014), environmental monitoring (Ghaderpour et al. 2023), the calculation of ecosystem service functions (Ji et al. 2023; Yirsaw et al. 2017), and ecological security (Ji et al. 2023; Pan et al. 2022). Therefore, land cover data, particularly large-scale coverage data, are among the most important public and fundamental data for surface systems science research.

Remote sensing image-processing technology is the main method for obtaining large-scale, wide-ranging, and multi-temporal land cover classification products. Examples of large-scale land cover products include Coordination of Information on the Environment (CORINE) by the European Environment Agency (EEA) (Büttner et al. 2012; Büttner, Soukup, and Kosztra 2014), World Cover by the European Space Agency (ESA) (Zanaga et al. 2021), GlobalLand30 by the National Geomatics Center of China (Chen et al. 2015), Copernicus Global Land Cover by the European Commission (GLC100, GLC200) (Buchhorn et al. 2020), Global Land Cover Characterization by the U.S. Geological Survey (GLCC) (Loveland et al. 2000), and Dynamic World developed by Google (Brown et al. 2022). Although numerous publicly available land cover products exist, they face challenges such as inconsistent resolution, difficulty in comprehensive utilization owing to large data volumes, inconsistent classification systems, and difficulties in mutual comparison between different products.

Land cover classification methods that use remote sensing technology include manual annotation, unsupervised classification, and supervised classification. Manual annotation or visual interpretation involves classifying land cover based on the visual features of objects in images. This offers high accuracy but is time-consuming and challenging to conduct on a large scale. The 3rd National Land Resource is an example of manual annotation, involving 21.9 million surveyors in China over three years that collected 295 million survey patches and provided a comprehensive understanding of China's land use (Chen et al. 2022). Wang et al. (2022) used manual annotation combined with an object-oriented approach to accurately obtain land cover data for Mongolia in 1990, 2000, 2010, and 2020, with overall classification accuracies of 82.26, 82.77, 92.34, and 81.84%, respectively.

Unsupervised classification involves the classification of images without manual sample labeling by iteratively self-classifying different feature channels (bands) in the feature space. Common remote sensing methods include K-means (Kavzoglu and Tonbul 2018) and ISODATA (Spruce et al. 2018). However, the effectiveness of this classification method varies markedly under different scenarios and poses challenges for large-scale applications.

The difference between supervised and unsupervised classification lies in the use of prior sample knowledge to train the classifier in supervised classification while utilizing statistical methods to establish the relationship between land cover features and samples. The effectiveness of supervised classification is often determined by the number of samples and the ability of the classifier to distinguish the target objects. Supervised classification methods are divided into traditional machine learning methods (such as the minimum distance method (Sun and Ongsomwang 2020), the maximum likelihood method (Hu and Zhang 2013), etc.), and deep learning methods (Li, Wang, and Yao 2021). Traditional machine-learning methods can effectively use expert knowledge. Gong et al. (2019) constructed a random forest classifier to train land cover labels with 30 m resolution in 2015 and transferred it to Sentinel-2 images in 2017, thus obtaining the first 10 m resolution global land cover product. Chen et al. (2015) selected multi-source remote sensing image data, collected land cover data and multimodal regional geographic information, fused them with images, and produced global land cover data products for 2000 and 2010 with an overall accuracy of 83.50%. With the development of deep learning, various network models have experienced considerable growth. Convolutional Neural Networks (CNNs) play an important role in extracting image features and are widely applied in image processing, with remote sensing land cover extraction techniques mostly relying on the CNN architecture. Google's Dynamic World model can produce 10 m-resolution land cover products using 24,000 Sentinel-2 land cover labeled patches obtained through Google crowdsourcing and training an improved Fully Convolutional Neural Network (FCNN) (Brown et al. 2022). The Alibaba DAMO Academy developed the Dual-Net model based on the Dynamic World, ultimately obtaining a 10 m-resolution land cover product for all of China (Liu et al. 2023).

The traditional method of manual interpretation relies on visual judgment and provides high accuracy; however, it is challenging to deploy it rapidly and on a large scale. The production of existing land cover data products based on deep learning methods requires a large amount of manual label sample creation that incurs high time and labor costs. Therefore, this study aims to address these issues and investigate a method that integrates multiple data sources for land cover production. This study proposes a method that use fused land cover to train deep learning models and that leverages the advantages of existing data products, automates the generation of training labels, and effectively reduces manual costs. Moreover, it allows users to adapt to local conditions, develop a classification system tailored to specific characteristics of the research area, and conduct training and predictions accordingly. The structure of the paper is organized as follows. An introduction of the data source, the strategy of data fusion, and the construction of the deep learning model is described in section 3. We indicate the details of model training and verification in section 4, discuss the distribution of land cover in Eastern Europe in 2022 in section 5, and then highlight the applicability of the multi-source data fusion method for land cover production.

2. Study area

The East European Plain, which is located in Eastern Europe, is the second-largest plain in the world, with boundaries extending from the Arctic Ocean in the north to the Black Sea, Caspian Sea, and Greater Caucasus Mountains in the south. Moreover, it stretches from the Ural Mountains to the Baltic Sea and covers an area of approximately 5 million square km. The average elevation is approximately 170 m, and the terrain is generally low and flat. The research area covered Ukraine, Belarus, Lithuania, Latvia, Estonia, Moldova, and parts of western Russia, and it shares borders with Kazakhstan, Georgia, Azerbaijan, Romania, Hungary, Slovakia, Poland, and Finland. The climate is predominantly temperate continental. The major land cover types in the region include crops, forests, grass, water, built areas, tundra, bare areas, wetlands, shrubs, and ice Figure 1.

Figure 1. Topographic map of the East European Plain.

3. Data and methods

3.1. Data

The data primarily consisted of image and related data such as land cover classification (Table 1). The image data used to construct the feature channels included bands 2, 3, 4, 5, 6, 7, 8, 11, and 12 from the Sentinel-2 imagery, longitude and latitude grids, and a digital elevation model with a total of 12 channels. The training label data were obtained through the fusion of 2020 data from Dynamic World, ESRI Land Cover, ESA World Cover, GlobeLand30, and Open Land Map Potential Distribution of Biomes.

Table 1. List of data sources.

Data	Google Earth Engine ID	Channels	Purpose
Sentinel-2	COPERNICUS/S2_SR_HARMONIZED	2, 3, 4, 5, 6, 7, 8, 11, 12	constructing feature channels
Grids of longitude and latitude	ee.Image.pixelLonLat()	longitude, latitude
DEM	COPERNICUS/DEM/GLO30	DEM
Dynamic World	GOOGLE/DYNAMICWORLD/V1	label	label data fusion
ESRI Land Cover	projects/sat-io/open-datasets/landcover/ESRI_Global-LULC_10m_TS	/
ESA World Cover	ESA/World Cover/v100	/
GlobeLand30	projects/ee-carylee/assets/DO01 projects/ee-carylee/assets/DO02 projects/ee-carylee/assets/DO03 projects/ee-carylee/assets/DO04	/
Open Land Map	OpenLandMap/PNV/PNV_BIOME-TYPE_BIOME00K_C/v01	/

Figure 2 illustrates the transit situations of the Sentinel-2 satellite over the East European Plain during the growing seasons of 2020 and 2022, from June to August, when the satellite achieved full image coverage during the growing season with a coverage rate of 100%. This guaranteed the availability of imagery for training and prediction purposes.

Figure 2. Image coverage during the growing seasons of 2020 (a) and 2022 (b) in the study area.

3.2. Methods

The primary approach involved the fusion of multiple land cover products from the same year (Figure 3). Consistent regions across various land cover products were used to construct label samples that were then used to train the deep learning model. Finally, the trained model was applied to map the land cover products. The main tools that were utilized included the Google Earth Engine (GEE) (for data preprocessing, primarily serving as the data source), Python (for the construction and training of the deep learning model and basic mapping tasks), and ArcGIS (to construct the fishnet vector, perform image mosaicking, and other related tasks).

Figure 3. Overall technical route.

3.2.1. Label data preprocessing and data fusion

When analyzing the current land cover status of the East European Plain, 10 main land cover types were identified, including water, forest, grass, wetlands, crops, shrubs, built areas, bare areas, ice, and tundra. Due to variations in the labels of different land cover products, it is essential to map the data into a unified index system. The mapping is represented in Table 2, where the values from various land cover products are mapped in the range of 1–10. The water, forest, grass, wetland, crop, built area, bare area, and ice categories were chosen by fusing the Dynamic World, World Cover, and ESRI Land Cover. Shrubs were selected by fusing Dynamic World, World Cover, and GlobeLand30. Tundra was obtained by fusing GlobeLand30 and Open Land Map. Figure 4(a)–(c) illustrate the land cover classifications of ESA, ESRI, and Google in 2020. Misclassifications in the northern tundra exist in all three data products. Additionally, there are inconsistencies in the classification of grassland areas in the southeastern region. Figure 4(d) indicates the distribution of tundra and shrublands in the GlobeLand30 product, and Figure 4(e) shows the distribution of tundra in the Open Land Map product for 2020. The overall distribution of tundra in GlobeLand30 and Open Land Map is consistent and is mainly concentrated in the northern part of the East European Plain.

Figure 4. Multi-source data products: (a) land cover of ESA World Cover in 2020; (b) land cover of ESRI Land Cover in 2020; (c) land cover during the growing season of Google Dynamic World in 2020; (d) distribution of tundra and shrublands in GlobeLand30; (e) distribution of tundra in Open Land Map.

Table 2. Multiple source data label mapping.

Value	Google Dynamic World		ESA World Cover		ESRI Land Cover		GlobeLand30		Open Land Map
1	0	Water	80	Permanent water bodies	1	Water	60	Water	1–27	Others
2	1	Trees	10	Tree cover	2	Trees	20	Forest
3	2	Grass	30	Grassland	11	Rangeland	30	Grass
4	3	Flooded vegetation	90	Herbaceous wetland	4	Flooded Vegetation	50	Wetland
5	4	Crops	40	Cropland	5	Crops	10	Crop
6	5	Shrub & scrub	20	Shrubland	/	/	40	Shrub
7	6	Built	50	Built-up	7	Built Area	80	Built
8	7	Bare	60	Bare / sparse vegetation	8	Bare Ground	90	Bare
9	8	Snow & ice	70	Snow and ice	9	Snow/Ice	100	Ice
10	/	/	/	/	/	/	71–74	Tundra	≥28	Tundra

The underlying assumption of this study is that land cover types in regions with consistent representation across multiple data products are considered true surface covers. The labels for water, forest, grass, wetland, crop, built area, bare area, and ice were determined in areas where Dynamic World, World Cover, and ESRI Land Cover categories were aligned. The three types of land cover data products exhibited strong consistency in the categories of water, forest, crop, built area, and ice, with consistency (the ratio of the intersection to the union of the area of each category among the products) exceeding 80%. The consistency of grass, wetland, and bare area categories was relatively low. Shrubs were identified in regions where the Dynamic World, World Cover, and GlobeLand30 categories were consistent with a consistency of 3%. The tundra in the northern part was determined in areas where the GlobeLand30 and Open Land Map categories were aligned with a consistency reaching 49%. After statistics, the total area of the Eastern European plain was approximately 4.93 million square km, with the area of the consistent region after fusion being approximately 3.4 million square km and accounting for 68% of the total area. Any pixels with labeling discrepancies were assigned a value of zero and were not involved in the model training process.

3.2.2. Model construction

The model was inspired by Google's Dynamic World model (https://github.com/google/dynamicworld) and was designed for global land cover mapping using Sentinel-2 data. The selected input channels were based on the original Sentinel-2 channels (2, 3, 4, 5, 6, 8, 11, and 12), with the addition of longitude and latitude bands and a digital elevation model. The model possessed 12 input channels and 10 output channels corresponding to different land cover categories that included water, forest, grass, wetland, built area, bare area, ice, and tundra. During the training process, the input image size was 180 × 180 × 12, and the label size was 180 × 180 × 11. The 11 label channels included nonconflicting regions (labeled 0–10) from multiple source products and were fused into 10 classes. Regions with pixel values of ‘0’ represent inconsistent areas in the data product fusion process, serve as masks, and are not involved in model training. The loss was not calculated for these regions during the loss function computation. The model primarily consists of convolution, depth-wise separable convolution, add, and concatenate operations. The activation function that was used throughout was the ReLU function, with the Softmax activation function used in the final layer (Figure 5). The parameters denoted by ‘i’ and ‘C_i’ in the grey-dash box represent the iteration number and the size of convolution kernel for each iteration. The total number of parameters in the model was 255,866.

Figure 5. Model structure (i is the iteration parameter, C_i is the size of convolution kernel in each iteration).

4. Results

4.1. Loss weight allocation

During the data training process, differences in the number of sample labels for each category can lead the model to learn the category features with larger sample sizes more easily while neglecting those with smaller sample sizes. Weighting the loss function based on the number of training samples can effectively mitigate the impact of sample bias.

The obtained land cover types consisted of 10 categories, each with varying sample quantities. Directly applying these to model training could result in the model being biased towards learning category features with larger sample sizes and overlooking those with fewer samples. When the disparity in sample quantities reaches a certain magnitude, even without learning from smaller labeled samples, the training accuracy of the model can still be high. Therefore, a solution to this issue is adopting a loss weight allocation. By iterating through the obtained dataset of over 3,000 labels, the number of samples (C_i) for each category was calculated. The loss weights for each label type are determined using Equations (1(1) $P_i={\displaystyle \frac{\prod _{i=1}^{10}{C}_i}{C_i}}$(1) ) and (2(2) $W_i={\displaystyle \frac{P_i}{\sum _{i=1}^{10}{P}_i}}$(2) ). These weights were assigned to the cross-entropy losses for each category by employing the weighted cross-entropy loss function as the training loss function for the model. (1) $P_i={\displaystyle \frac{\prod _{i=1}^{10}{C}_i}{C_i}}$(1) (2) $W_i={\displaystyle \frac{P_i}{\sum _{i=1}^{10}{P}_i}}$(2) where C_i and W_i represent the pixel count and weight of loss for each category (i = 1, 2, … , 10), respectively.

The weights for each category in the loss function were calculated using Equations (1) and (2), and the weight of loss allocations are listed in Table 3. Categories with fewer labeled samples, such as ice, shrubs, and bare areas, had higher allocated weights of loss. Conversely, categories with a larger number of labeled samples allocated lower weights of loss. This approach helps to mitigate the impact of differences in sample quantities.

Table 3. Weight allocation.

CLASS	Pixel	Weight
Water	26307308	9.68231E-06
Forest	503643	0.000505746
Grass	52665383	4.83649E-06
Wetland	2433405	0.000104674
Crop	34465	0.007390552
Shrub	20348818	1.25175E-05
Built area	499	0.510451665
Bare area	385432	0.000660857
Ice	10109	0.025196892
Tundra	559	0.455662577
SUM	102689621	1

4.2. Model training

The model was trained on 3,196 samples with a size of 180 × 180 for 1,000 iterations until it reached an accuracy of 99.9%. The training curve is represented in Figure 6. Due to the insufficient GPU memory on the personal computer, the maximum batch size was set to 12, and this caused oscillations in the training curve. However, overall, the loss of model exhibited a decreasing trend, whereas accuracy exhibited an increasing trend. The model selected here was from the 997th iteration and achieved a training accuracy of 0.999 and a loss of 0.001.

Figure 6. Model training iteration curve.

Using the ArcMap software, a vector grid file with dimensions of 5 km × 5 km was constructed. This grid was uploaded to the GEE, and after cropping features from 10 m resolution images, 200 thousand tiles of 500 pixels in size with 12 feature channels were obtained. The total size of these tiles was 2.2 TB, stored locally. The model was applied for prediction and stitching, resulting in an initial land cover dataset for the study area (Figure 7). Due to the marked classification differences in the southeastern region between Google's land cover product and ESA and ESRI, the labels of this region could not participate in training, thus causing misclassification issues.

Figure 7. Comparison of misjudged areas in preliminary results.

4.3. Land cover mapping

Upon visual examination using Google Earth Pro, it was observed that the misclassification issue in the southeastern portion of the study area mostly manifested as grassland being classified as cropland and an overestimation of built areas. To address this misclassification, rough delineation was performed using rectangular boxes. Misclassified crop land was replaced with fusion-classified labels from ESA World Cover and ESRI Land Cover. This process was repeated for model training and prediction, ultimately resulting in land cover data for the study area during the 2022 growing season. Figures 8 and 9 indicate the land cover data for the entire study area and a detailed view, respectively. The East European Plain can be broadly divided into three regions based on latitude, including tundra and snow in the north (north of 66° N), forest regions in the central region (between 55° and 66° N), and crop land and grassland in the south (south of 55° N). Various land cover types, including grass, crops, tundra, and forests, are widespread and contiguous. The built areas were more scattered, possessed a radial distribution pattern, and were denser in the southern part of the East European Plain. The central forested region includes major cities such as Moscow and St. Petersburg. Snow is mainly distributed in high-latitude northern and southern regions, including the Caucasus Mountains.

Figure 8. Land cover data for the growing season of the East European Plain in 2022.

Figure 9. Land cover details of the East European Plain in 2022.

4.4. Accuracy validation

Based on historical land cover data products, 1,527 verification points were selected according to a manual examination of high-resolution images from Google for the East European Plain in 2022. The numbers of points for each category, including water, forest, grass, wetland, crop, shrub, built area, bare area, ice, and tundra, were 131, 379, 213, 35, 377, 45, 49, 83, 76, and 139, respectively. Tables 4 and 5 describe the confusion matrices for the land cover data produced by fusing multiple data products and Google Dynamic World for the 2022 growing season in the study area. The method employed here yielded superior land cover data. Both the producer and user accuracies for each category (Figure 10) were higher than the results from Google Dynamic World land cover. Categories such as water, ice, tundra, and forest demonstrated high classification accuracy, whereas the accuracy for bare areas and shrubs requires improvement. The calculations exhibited an overall classification accuracy of 89.3% and a kappa coefficient of 0.873 for the land cover data during the 2022 growing season.

Table 4. Confusion matrix of 2022 East European Plain land cover data by integrating multiple source data product methods.

		Our method
		Water	Forest	Grass	Wetland	Crop	Shrub	Built area	Bare area	Ice	Tundra
Ground Truth	Water	130							1
	Forest	2	338	27	3	4		1			4
	Grass	1	4	175	4	23	5				1
	Wetland			1	27		7
	Crop		1	23		339			14
	Shrub				4		37		4
	Built area			1		1		32	15
	Bare area						7		76
	Ice									74	2
	Tundra			3							136

Table 5. Confusion matrix of 2022 East European Plain land cover data from Google Dynamic World.

		Google Dynamic World 2022
		Water	Forest	Grass	Wetland	Crop	Shrub	Built area	Bare area	Ice	Tundra
Ground Truth	Water	126	9	1		2			1		3
	Forest	7	284	74	7	23		6	1		88
	Grass		15	63	3	49	13				14
	Wetland		3	10	27	11	12
	Crop		19	53	1	261					7
	Shrub		7	21		7	20		13		15
	Built area		2	1		3	2	27	16		2
	Bare area			6		11	9		74		8
	Ice		4	1					5	74	6
	Tundra	/	/	/	/	/	/	/	/	/	/

Figure 10. Comparison of user accuracy and mapping accuracy.

5. Discussion

5.1. Land cover patterns and distribution

In the land cover data analysis for the 2022 growing season in the East European Plain, which covers a total area of 4.93 million square km, it was observed that forests, crops, grass, tundra, and water bodies are widespread (Figure 11(a)). Among these, forest cover had the largest area of nearly 2.4 million square km (49% of the total area), while crop land occupied an area of 1.22 million square km (26%). Figure 11(b) describes the country-specific breakdown of the land cover categories. The study area encompasses the Russian region as well as seven countries and territories, including Ukraine, Belarus, Moldova, Lithuania, Latvia, and Estonia. The Russian region consistently exhibited the highest proportion of each land cover type, and this was followed by Ukraine. Ukraine, often referred to as the ‘breadbasket of Europe,’ possessed a wide crop distribution, constituting 30% of the total crop land area in the East European Plain and surpassing other European countries.

Figure 11. Overview of land cover data in the East European Plain in 2022.

Next, the land cover proportions for each country were analyzed (Figure 12). Russia region, Latvia, Estonia, and Belarus possess the largest forested areas of all countries and regions. Ukraine and Moldova primarily feature crop land as the dominant land cover type. Lithuania possesses a relatively balanced proportion of forest and crop land. Built areas in the study area account for less than 2% of the entire region. Compared to other countries/regions, Lithuania had the highest ratio of built areas to its land area (7%), and this is followed by Ukraine (nearly 4%).

Figure 12. Proportion of land cover by country/region.

5.2. Applicability of the method

The land cover production method that integrates multiple data products, combines data from the same year, constructs consistent labels for uniform regions, and is effective in leveraging the strengths of different data products. This fusion strategy contributes to improved accuracy and comprehensiveness of land cover classification. Deep learning, particularly by referencing Google's Dynamic World model, enhances the ability of the model to capture complex land cover features and adapt to changes in terrain and cover types. This method applies to large-scale, high-resolution data research and provides a feasible and customized solution for regional land cover mapping. By fusing different data sources, users can choose and combine classification standards applicable to specific regions, making land cover data more representative of actual surface conditions. This personalized classification system better meets user needs, offering more targeted land cover information for local studies and planning.

However, this study assumed the accuracy and error-free nature of consistent regions across multiple data products. Real-world scenarios may involve misclassification or omission of regions in multiple data products, thus posing a challenge for future research focused on achieving higher-precision data fusion and quality improvement. Additionally, the high parameter count of the model makes it difficult to achieve interface-level calls on the GEE platform (Li et al. 2022), thus requiring local deployment.

In summary, this research method is not only applicable to large-scale, high-resolution data studies and provides a viable solution for customized land cover data production in regional contexts. This method demonstrates broad prospects in the field of land cover classification through the fusion of multiple data sources and the application of deep learning models.

6. Conclusion

To address the challenge of effectively assessing land cover in the East European Plain using existing products, this study researched on a global 10 m resolution land cover product. This research involved multiple steps, including the fusion of data products, model training, land cover product mapping, and post-processing. To address the issue of an incomplete classification system, this study incorporated data fusion from Google Dynamic World, the European Space Agency's World Cover, ESRI Land Cover, Globe Land30, and the Open Land Map Potential Distribution of Biomes. The data were uniformly processed, and further optimization was performed using the land cover classification system for the East European Plain.

Sentinel-2 imagery data for 2020 and 2022, along with latitude and longitude grids and digital elevation models, were obtained from the GEE platform for training and prediction data. The model that was optimized based on Google’s Dynamic World underwent iterative training, ultimately resulting in a high-precision land cover prediction model for the East European Plain. The model was applied to over 200 thousand 10 m-resolution 5 × 5 km feature images, and manual discrimination was employed to supplement the land cover data. Confusion matrix calculations revealed superior accuracy in the 2022 growing season land cover data for the East European Plain, ultimately outperforming Google's Dynamic World products across various categories.

This study provides a novel approach for mapping large-scale global land cover products with potential implications for applications in other regions. Future studies could expand this research area, enhance the accuracy of data fusion labels, and improve the global applicability of products.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Data will be made available on request.

Additional information

Funding

This research was funded by the Science & Technology Fundamental Resources Investigation Program of China (grant number 2022FY101902), the Special Exchange Program of Chinese Academy of Sciences (grant number E2X20060Y2), and the Construction Project of the China Knowledge Center for Engineering Sciences and Technology (grant number CKCEST-2023-1-5).