Integrating remote sensing temporal trajectory and survey statistics to update land use/land cover maps

ABSTRACT

Remote sensing and land resource surveys have been used in recent decades for land use/land cover (LULC) mapping; however, keeping the developed LULC up-to-date and consistent with land survey statistics remains challenging. This study developed a practical and effective framework to automatically update existing LULC products and bridge the gap between remote sensing classification results and land survey data. This study employed Landsat imagery time series, change detection algorithms, sample migration, and random forests to develop a framework for updating existing LULC products in China from 1980–2015 to 1980–2022. The updated LULC maps reflect the post-2015 LULC changes well and maintain continuity with the pre-2015 products. Additionally, a statistical space allocation method based on the minimum cross-entropy strategy was proposed to optimize the LULC maps, increasing the correlation coefficient (r) with China’s second and third national land survey statistics from 0.41–0.89 to 0.86–0.99. Thus, the framework and products developed in this study provide valuable tools for sustainable land use and policy planning.

KEYWORDS:

• Land use
• land cover
• change detection
• land survey
• statistical constraints

1. Introduction

Global land use/land cover (LULC) is undergoing extensive and rapid changes due to human activities and climate change, which substantially influence the biodiversity, surface energy balance, and sustainable development of Earth (Anderson 1976; Gong et al. 2013). Accurate and reliable LULC data are typically utilized as a crucial basis for research on earth system modeling, resource management, and ecological and environmental evaluation to measure and respond to LULC changes and their impact (Reyers et al. 2009; Timmermans et al. 2007; J. Yang et al. 2013). In this context, the past decades have witnessed efforts in remote sensing technology and land resource surveys for LULC mapping.

LULC mapping is one of the most essential applications of remote-sensing technology (Hansen et al. 2013; Redo and Millington 2011; Yu et al. 2022). With the accumulation and diversification of remote sensing imagery, advances in artificial intelligence technology, and the increased demand for LULC products, LULC remote sensing mapping has undergone manual image interpretation, semi-automatic interpretation, and automatic mapping based on statistical theory, progressing to the current era of intelligent remote sensing mapping, which has significantly improved in terms of multi-source data, algorithm variety, and mapping accuracy (Gong 2021; Vali, Comai, and Matteucci 2020). Machine learning classifiers, such as random forest, support vector machines, and convolutional neural networks, are typically used for LULC mapping. These techniques have enabled the creation of LULC products on both national and global scales, with examples including the 500-m Moderate Resolution Imaging Spectro-radiometer annual Land Cover product (MCD12Q1) (Friedl et al. 2010; Sulla-Menashe et al. 2019), 300-m European Space Agency Climate Change Initiative Land Cover Dataset (ESA CCI LC) (ESA 2017), 30-m Finer Resolution Observation and Monitoring of Global Land Cover Plus (FROM-GLC Plus) products (Gong et al. 2013; Yu et al. 2022), and 10-m ESA WorldCover products. The spatial resolution and consistency of LULC products have gradually improved in recent years since the public availability of Landsat and Sentinel data; however, challenges remain in keeping LULC products up-to-date, impeding further updates of relevant studies based on these data.

Generally, the process of updating large-scale LULC products is hindered by significant issues, such as the high quantity of remote sensing imagery collection and processing and the massive amount of data computation required. These issues can be addressed by leveraging cloud computing platforms, such as the Google Earth Engine (GEE), which provides a powerful tool for LULC monitoring (Gorelick et al. 2017). Despite this, a lag in LULC products remains evident, characterized by the time gap between data publication and mapping years. For instance, the CLUD-A (Y. Xu et al. 2020) and GLASS-GLC (H. Liu et al. 2020) products released in 2020 reflect a five-year lag for the mapping years 1980–2015 and 1982–2015, respectively. Furthermore, annual LULC product updates are still plagued by discontinuity and unreasonable changes owing to differences in the spatial and temporal distributions of remote sensing imagery and the stochastic nature of machine learning classifiers. Challenges in yearly sample updates are another issue hampering product updates. Faced with these challenges, change detection algorithms such as Continuous Change Detection and Classification (CCDC) and sample generation or migration methods (Huang et al. 2020) demonstrate the potential for automatic LULC product updates (Brown et al. 2020). Nevertheless, a comprehensive and operational framework for updating LULC products nationally and globally is still lacking.

Land resource surveys and remote sensing technology are essential for understanding LULC patterns in regions worldwide (X. Chen et al. 2022; Zhang, Dong, and Ge 2022). Early land resource surveys predominantly relied on manual field surveys. They were mainly conducted in economically developed countries, such as the first national land survey in the United Kingdom in 1931 (Southall, Baily, and Aucott 2007) and the 1934 National Erosion Reconnaissance Survey in the United States (Harlow 1994). With the rapid development of Earth observation and computer technology in recent decades, remote sensing image interpretation has evolved into a valuable tool for land resource surveying (Fuller, Sheail, and Barr 1994; Shu and Du 2022). Although remote sensing technology has provided important insights into LULC patterns, image interpretation for land resource surveys relies mainly on manual methods (Fan et al. 2022; Sannier, McRoberts, and Fichet 2016). Moreover, land resource surveys are typically conducted at the national level, and statistics are generally published in administrative units, such as at the provincial level. While these surveys provide important basic LULC statistics (Nedd et al. 2021; Wang et al. 2022), finer-scale LULC distributions are rarely available through land surveys, resulting in inconsistencies with several published gridded socioeconomic and ecological datasets, such as Gross Domestic Product (GDP) (J. Chen et al. 2022) and carbon emissions (Huo et al. 2022). This disparity impedes the application of land survey statistics to sustainable land use and policy planning. Moreover, remote sensing LULC mapping rarely incorporates land survey data into the mapping process. Differences in mapping or survey tools and discriminatory criteria have led to a gap between remote sensing classification areas and land survey statistics (X. Chen et al. 2022; X. Liu et al. 2018; Yu et al. 2013). Conversely, advanced technologies in agriculture, such as the Spatial Production Allocation Model (SPAM) (You et al. 2014), have successfully integrated satellite-derived LULC maps with tabular statistics for crop distribution and yield estimation, demonstrating the potential for improved mapping accuracy through such a strategy.

Considering the lack of LULC product updates and insufficient access to finer-scale land survey statistics, this study presents an approach for LULC product updates and optimization to provide an efficient and practical updating framework and bridge the gap between remote sensing classification results and land survey statistics. Taking China as an example, the time series of Landsat imagery, along with change detection algorithms, sample migration, and a random forest classifier, were used to construct a framework to update the existing LULC product of CLUD-A, and a statistical space allocation method using the minimum cross-entropy strategy was developed to optimize the classification maps based on the land survey statistics of China.

2. Data

2.1. Satellite imagery

All available Landsat surface reflectance images from 2000 to 2022 were used for land cover updating in this study, including the Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), and Operational Land Imager (OLI). Landsat images were processed by the United States Geological Survey (USGS) using the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) (Masek et al. 2006) and the Land Surface Reflectance Code (LaSRC) (Vermote et al. 2016). The clouds and shadows were masked using the C Function of Mask (CFMask) (Foga et al. 2017).

2.2. Land cover products

CLUDs is a 30-m LULC dataset for the 1980s, 1995, 2000, 2005, 2010, 2015, and 2020 in China. This dataset was developed using Landsat TM/ETM+/OLI imagery and human–computer interactive interpretation method with a classification system of grassland, forest, unused land, cropland, water, and built-up land (J. Liu et al. 2003; J. Liu et al. 2003; J. Liu et al. 2014; J. Liu et al. 2010; Ning et al. 2018). Furthermore, based on the CLUDs dataset, an annual LULC dataset (CLUD-A) was developed from 1980 to 2015 (Y. Xu et al. 2020). This study aimed to update annual land cover maps based on the CLUD-A product.

2.3. Land survey statistics

The Chinese government conducted the first national land survey (FNLS) in 1984 to ensure sustainable land use and management. China launched the second national land survey (SNLS) in 2007 and the third national land survey (TNLS) in 2017 to provide county-level land use status statistics (X. Chen et al. 2022). Currently, the SNLS provincial statistics for 2009–2016 and TNLS provincial statistics for 2019–2021 can be accessed publicly. This study utilized these provincial statistics for space allocation in the corresponding year’s LULC map optimization. The land use systems of the SNLS and TNLS were converted into the CLUDs classification system. Furthermore, with the support of the Ministry of Natural Resources (China), the 2019 TNLS county-level statistics of Inner Mongolia, Heilongjiang, Sichuan, Yunnan, and Shaanxi were obtained and used to evaluate the map optimization method.

3. Methods

This study aimed to develop an approach for LULC product updates and optimization to provide an efficient and practical updating framework and bridge the gap between remote sensing classification results and land survey statistics. To achieve this, a framework for updating and optimizing the LULC product (CLUD-A) was developed by leveraging a combination of continuous change monitoring and national land survey statistics (Figure 1). The proposed method integrates the change detection results from CCDC and LandTrendr, which are the foundation for identifying potential change areas. Subsequently, the migrated samples and a random forest classifier were employed to update the LULC product of CLUD-A from 1980–2015 to 1980–2022. Furthermore, the updated product was optimized using a minimum cross-entropy formulation based on SNLS and TNLS statistics.

Figure 1. Workflow of the LULC product update using LandTrendr and CCDC, and optimization using the statistics space allocation method.

3.1. LULC map update

3.1.1. Map update using CCDC and LandTrendr

The updated LULC map in this study was developed using all available Landsat images in China from 2000 to 2022. The CCDC algorithm (Zhu and Woodcock 2014) was first used for time-series modeling and change detection. For every pixel, the Normalized Difference Vegetation Index (NDVI) (Rouse et al. 2007), Modified Normalized Difference Water Index (MNDWI) (H. Xu 2006), and Normalized Difference Buildings Index (NDBI) (Zha, Gao, and Ni 2003) were obtained for the spectral time-series fitting of CCDC. Specifically, a harmonic model was built based on the spectral or index time series of historical remote sensing images with variable coefficients to fit and predict the spectral or index values on a specific date (Equation (1)). (1) $\hat{p}(i,t)=c_{oi}+c_{1i}t+{\sum }_{n=1}^3\left(a_{ni}\cos {\displaystyle \frac{2\pi nt}{T}+b_{ni}\sin {\displaystyle \frac{2\pi nt}{T}}}\right)$(1) where the predicted value of the $i$th Landsat band at Julian date $t$ is represented by $\hat{p}(i,t)$, $T$ is the average number of days per year ($T=365.2425$), $a_{ni}$ and $b_{ni}$ are the $n$th order harmonic coefficients for the intra-annual change in the $i$th band, $c_{oi}$ and $c_{1i}$ correspond to the intercept and slope coefficients of the $i$th band, respectively. The fit coefficients of $c_{oi}$, $c_{1i}$, $a_{ni}$ and $b_{ni}$ were estimated using the least absolute shrinkage and selection operator (LASSO) regression model (Tibshirani 1996). By dynamically updating the time-series model with the surface reflectance of a new remote sensing image, pixels were recognized as changed when the difference between the reflectance and predicted reflectance values exceeded a certain threshold generated based on three times the Root Mean Square Error (RMSE) value obtained from the time-series model estimation of the historical period.

However, abrupt land surface changes detected by the CCDC do not indicate ground-truth LULC changes (Zhu et al. 2020). Certain ephemeral land disturbances, such as floods and grass fires, may lead to false-positive changes in the LULC. Moreover, interannual differences in the effective Landsat imagery distribution may lead to the misinterpretation of change detection by the CCDC. Therefore, considering the continuous monitoring capability of the LandTrendr algorithm for long-term vegetation changes, it was included in the change-detection procedure to capture vegetation changes in this study. The NDVI was selected as the input feature based on its wide application in forest monitoring (Guo and Gong 2018), and the mean value of NDVI in the growing seasons (i.e. June to August) during 2000–2022 was used in this study. The LandTrendr parameters were set following a previous study (Kennedy, Yang, and Cohen 2010), which was considered reasonable for vegetation change monitoring (Du et al. 2022; Y. Yang et al. 2018). Based on the temporal fitting results of LandTrendr, the maximum magnitude of the fitted observational values between 2015 and 2022 was calculated as $Mag$. Pixels with $Mag>0.1$ or $Mag<-0.1$ were determined as potential vegetation gain or loss, respectively. The potential land cover change regions comprised the change detection results of the CCDC and LandTrendr. The change detection results from the CCDC provided a break in 2017; however, the Sentinel 2 images illustrated that the detected area did not experience LULC changes in 2017 (Figure 2(a)). The abrupt change detected by the CCDC may have been induced by changes in the cultivated crops. In this case, the LandTrendr algorithm helped filter out the misjudgments. Additionally, LandTrendr has an advantage over CCDC in monitoring gradual land surface changes, which complement slow LULC changes, such as the gradual recovery of the forest shown in Figure 2(b). Although the LandTrendr algorithm is a valuable tool for detecting changes in vegetation over time, its ability to detect annual changes in non-vegetated areas is limited. Therefore, relying solely on LandTrendr change detection results may be insufficient. To complement this, the monitoring results of the CCDC (Figure 2(c)) necessary to accurately identify and track changes in LULC over time. Therefore, combining the strengths of both algorithms is crucial to achieve a comprehensive and accurate mapping of LULC changes.

Figure 2. Change detection examples for a single pixel (the red dot on the true color combined Sentinel 2 images or Google Earth images on the blow). The combination of CCDC and LandTrendr can (a) filter out abrupt vegetation LULC changes detected by CCDC in Baoding, Hebei Province; (b) complement slow vegetation restoration undetected by CCDC in Yuanping, Shanxi Province; and (c) complement new buildings undetected by LandTrendr in Tangshan, Hebei Province.

Random forest was used as the supervised classifier to classify the potential change regions. Annual samples from 2016 to 2022 were generated based on the first all-season sample set (FAST) collected in 2015 (Li et al. 2017) using an automated sample migration method (Huang et al. 2020; Yu et al. 2022) and converted to the CLUD classification system. Finally, the annual LULC maps for 2016–2022 were updated with the 2015 map of the CLUD and the classification results of potential change regions by year. Additionally, vegetation changes were carefully scrutinized by crosschecking the outputs of the LandTrendr algorithm. Specifically, vegetation gain, such as the transition from grassland/shrubland to forest or grassland to shrubland, was masked at $Mag>0.1$. Conversely, vegetation loss, such as the transition from forest to grassland/shrubland or from shrubland to grassland, was identified by masking with $Mag<-0.1$.

3.1.2. Evaluation

The updated LULC maps were assessed based on LULC classification accuracy and LULC area-change performance. The 30-m China Land Cover Dataset (CLCD) and 300-m ESA CCI LC were utilized to facilitate this evaluation. To ensure comparability, both datasets were converted to the CLUD classification system before being used for comparison.

First, annual LULC evaluation samples from 2016 to 2022 were generated using the sample migration method, and the FAST dataset was used for annual training sample generation. To assess the overall accuracy (OA) and uncertainty of the three datasets, an error matrix-based calibration estimator was used with a 95% confidence interval (Olofsson et al. 2014), covering the entire region of China. Second, the LULC area change performance of the LULC datasets was conducted to evaluate whether the area changes since 2010 were consistent with CLDA and ESA CCI LC and reasonable, based on visual comparison. Specifically, because this study provided updated LULC maps for 2016–2022, the LULC maps for 2010–2015 for the area change comparison were from CLUD-A, providing a basis for verifying the continuity of the updated maps.

3.2. LULC map optimization

3.2.1. Map optimization using statistics space allocation

The aim of statistics space allocation is to seek the optimized area share of each LULC class within 1-km grids, and a better fit between the optimized maps and statistics is expected. Inspired by SPAM (You et al. 2014), a map optimization method was developed using the inverse distance weighted (IDW) algorithm and the minimum cross-entropy formulation (Figure 3). The map optimization method was implemented for each Chinese province because the land survey statistics were at the provincial level. However, the method developed here can be applied to statistics at other scales, such as the country level. Taking Shaanxi Province as an example, Figure 3 illustrates the optimization procedure. First, the 1-km grided LULC area percentage of each grid $i$ and each land class $j$ was calculated as $o_{ij}$ based on the 30-m LULC maps. The provincial land use area statistics were then assigned to each grid as $p_{ij}$ based on $o_{ij}$. We assume that the statistical and map areas of land class $j$ are $LandAre{a}_j$ and $MapAre{a}_j$, respectively. Notably, if $LandAre{a}_j>MapAre{a}_j$, there is a possibility that $p_{ij}>1$, which makes it difficult for the subsequent minimum cross-entropy formulation to determine the optimized area share. Therefore, the IDW algorithm was applied to the potential distribution estimation for pixels with $p_{ij}>1$. Specifically, a fraction > 1 (i.e. $p_{ij}-1$) is redistributed to the grids around grid $i$ according to the weights provided by the IDW. Notably, these assigned surrounding grids must satisfy (1) $p_{ij}<1$, and (2) the sum of the mapping area shares of land class $j$ and classes $j^{\mathrm{\prime }}$ often confused with class $j$ (such as grass and unused land), is > 0 (i.e. $o_{ij}+o_{i{j}^{\mathrm{\prime }}}>0$). The estimated distribution potential of grid $i$ and land class $j$ was then developed as ${\pi }_{ij}$.

Figure 3. The workflow of land use/land cover (LULC) map optimization using the statistics space allocation method with Shaanxi Province as an example. (IDW: inverse distance weighted)

The minimum cross-entropy formulation was then developed for each grid $i$ to seek the optimized area share $s_{ij}$ based on ${\pi }_{ij}$, such that. (2) $MINCE(s_{ij},{\pi }_{ij})=\sum _j{s}_{ij}ln{s}_{ij}-\sum _j{s}_{ij}ln{\pi }_{ij}$(2) Subject to: (3) $\sum _j{s}_{ij}=1\mathrm{\forall }i,$(3) (4) $1\ge s_{ij}\ge 0\mathrm{\forall }i\mathrm{\forall }j$(4) where $MIN$ is minimize, $i$ is the grid identifier, $j$ is the LULC identifier, and $GridAre{a}_i$ is the area of grid $i$.

Generally, the objective function (Equation (2)) minimizes the cross-entropy between the estimated area share $s_{ij}$ and potential area share ${\pi }_{ij}$. The adding-up constraint in Equation (3) was used to ensure that the sum of the area shares was 100%. Equation (4) provides a range constraint for each LULC class area share: This cross-entropy formulation involves assessing the optimized LULC distribution based on statistics and the principle of minimum cross-entropy.

3.2.2. Evaluation

The 2019 TNLS county-level statistics of five provinces (Inner Mongolia, Heilongjiang, Sichuan, Yunnan, and Shaanxi) were used as evaluation datasets for the optimized LULC area share to prove the effectiveness of the proposed map optimization method. The five provinces selected for validation purposes are situated on both sides of the Hu line (Hu 1935), a demarcation line in China that reflects spatial differences in the natural environment and socioeconomic development in China. These provinces boast rich, topographic variations and diverse LULC types, indicating that the validation results more accurately reflect the performance of the optimization method across different ecological and environmental conditions.

Based on the mapping and optimized results of the 2019 LULC map, the land area for each LULC class was obtained for each county in the five provinces. Linear regression was then applied to evaluate the optimization effects for each province. The Pearson Correlation Coefficient ($r$) was used to quantify the correlation and agreement between the statistics and mapping/optimized LULC distribution. In addition, inspired by the Spectral Angle Distance (SAD) index in the Change Vector Analysis (CVA), the Area Angle Distance (AAD) index was introduced to the validation process to quantify the similarity between the classified area and the statistics before and after optimization. (5) $\theta ={\cos }^{-1}{\displaystyle \frac{{\sum }_{i=1}^{N}Are{a}_{i(c)}Are{a}_{i(s)}}{\sqrt{{\sum }_{i=1}^N{(Are{a}_{i(c)})}^2{\sum }_{i=1}^N{(Are{a}_{i(s)})}^2}}}$(5) (6) $AAD=\cos (\theta )$(6) where $N$ is the number of LULC classes, $Are{a}_{i(c)}$ is the mapping/optimized area of class $i$, and $Are{a}_{i(s)}$ is the statistical area of class $i$. If $r$ and AAD were higher after optimization, the optimized map obtained using remote sensing mapping results and provincial statistics remained reasonable at the county level, indicating that our optimization method effectively redistributes LULC.

4. Results

4.1. Annual LULC changes in China during 2016–2022

The LULC maps from 2016 to 2022 were produced using the proposed update methodology that involved utilizing Landsat images coupled with the advanced change detection methods of CCDC and LandTrendr and the random forest classification algorithm. These maps were updated based on the original CLUD-A dataset that provided the foundation for this study. The latest map for 2022, which shows the current LULC patterns, is presented in Figure 4, and the maps for 2016–2021 are illustrated in Figure S1. Moreover, the original CLUD-A dataset (1980–2015) was merged with the updated maps to create the updated CLUD-A (1980–2022).

Figure 4. Updated annual land use/land cover (LULC) maps of 2022 in China (for 2016–2021 maps, please see Figure S1.)

To confirm the consistency of the updated maps with the original CLUD-A dataset and the reasonableness of the LULC changes, the annual area changes for various LULC types between 2010 and 2022 were analyzed, and the results were compared to those of the CLCD and ESA CCI LC (Figure 5). Notably, the area changes in Figure 5 may differ from the reported results owing to differences in the classification systems. Because the updated CLUD-A was completed between 2016 and 2022, the red dashed vertical lines in Figure 5 were used to compare the changes that occurred before and after 2015. The updated CLUD-A displayed smooth changes in different LULC types compared to the original dataset, with no unreasonably abrupt changes observed. Moreover, the annual area change trend of the updated CLUD-A was similar to that of the CLCD and ESA CCI LC, demonstrating the potential of the LULC updating method.

Figure 5. Annual area changes of CLUD-A, CLCD, and ESA CCI LC by different LULC types from 2010 to 2022.

Furthermore, the analysis revealed that cropland and grassland areas in China have continuously decreased since 2015, with decline rates of 1.94% (3.31 million ha) and 1.08% (3.22 million ha), respectively. Conversely, forest and urban areas have continuously increased since 2015, with growth rates of 0.35% (0.77 million ha) and 3.18% (1.01 million ha), respectively. The rate of urban expansion has decreased since 2015 (from 1.03 million ha yr^-1–0.13 million ha yr^-1). A decline in area was observed between 2010–2015 for unused land. However, the total area of unused land has increased from 2016 to 2022.

4.2. LULC map update performance

To obtain the annual accuracy of the updated LULC maps, annual validation samples were generated based on the migration results from the FAST dataset. The annual sample size is shown in Figure 6. The FAST dataset was collected primarily in 2014 and 2015, resulting in a larger sample size for these two years. Because the LULC maps for 2016–2022 were obtained based on the 2015 CLUDs product, the classification results for 2010–2015 of CLUD-A were included when calculating the annual accuracy to reflect the continuity of the classification results. Figure 6 shows that the overall accuracy of the updated CLUD-A maps remains stable after 2015, indicating a high level of consistency and continuity with the original product. This accuracy stability highlights the robustness of the methodology, ensuring that the results are reliable and consistent for practical temporal analysis and monitoring of LULC changes.

Figure 6. Annual sample size and overall accuracy of updated CLUD-A (2010–2022).

The reasonableness of the LULC changes was further verified by visual comparison. Based on Figure 5, China’s urbanization has decelerated since 2015, with substantial LULC changes generated primarily by national land development projects and the construction of water conservancy infrastructure and large buildings. Examples of the LULC dynamic changes include (1) construction of agricultural facilities in Shannan, Tibet (Figure 7(a)); (2) water expansion caused by the construction of the Chushandian Reservoir in Xinyang, Henan, located in the main stream of the Huai River (Figure 7(b)); (3) newly reclaimed cropland caused by land development projects in Hangzhou, Zhejiang (Figure 7(c)); and (4) urban expansion caused by the construction of the Beijing Daxing International Airport (Figure 7(d)). According to the enlarged examples shown in Figure 7, China’s economic and social growth during the ‘13th Five-Year Plan’ (i.e. from 2016 to 2020) continues to occur at the expense of cropland loss; however, the government has attempted to implement cropland protection and food security, such as land development projects, including the balance of cultivated land occupation and compensation.

Figure 7. Examples of LULC changes during 2015–2022. (a) Construction of agricultural facilities in Shannan, Tibet; (b) water expansion caused by the construction of water conservancy infrastructure in Xinyang, Henan; (c) newly reclaimed cropland in Hangzhou, Zhejiang; (d) construction of Beijing Daxing International Airport.

4.3. LULC map optimization performance

Based on provincial statistics from the SNLS and TNLS for 2009–2016 and 2019–2021, respectively, optimized 1-km LULC maps were developed for the respective periods using the proposed statistical space allocation method. Each pixel in the optimized maps comprised an area share of different LULC categories, including forests, grasslands, croplands, water, urban land, and unused land. The category with the largest area share for that pixel was determined, and optimized LULC maps were developed by analyzing the area share of each LULC category in each pixel (Figure 8). The two maps illustrate the dynamics of LULC patterns in China over the past decade. One of the most striking features of these maps is the expansion of urban clusters, which reflects the rapid urbanization and economic growth experienced by China over the past decade. However, it also highlights the challenges of sustainable land use management, the need for effective urban planning, and the necessity of balancing economic development with environmental conservation.

Figure 8. Optimized land use/land cover (LULC) maps of China during 2009–2016 and 2019–2021.

To evaluate the provincial statistic-based optimization results, the consistency of the classified and statistical areas before and after the optimization of the five provinces (Inner Mongolia, Heilongjiang, Sichuan, Shaanxi, and Yunnan) by different LULC types was calculated using the 2019 TNLS county-level statistics (Figure 9). Based on the $r$ values (Pearson Correlation Coefficient) shown in Figure 9(a–e), $r$ calculated by linear regression of classified results and county-level statistics was higher after optimization (from $0.41\le r\le 0.89$ to $0.86\le r\le 0.99$). Additionally, the optimization algorithm was more effective for provinces where the consistency between the original classification results and the statistical results was unsatisfactory. For example, when the $r$ value improved from 0.44–0.94 in Shaanxi. The AAD for the counties of the five provinces (Figure 9(f)) before and after the optimization process was obtained to quantify the change in similarity using statistics. As shown in Figure 9(g), the results demonstrate an improvement in the consistency of the classified and statistical areas at the county level, further highlighting the efficacy of our approach. Specifically, certain counties, such as Batang County in Sichuan, have improved from 0.32–0.86. Moreover, the apparent optimization effect of Huanglong County is enlarged in Figure 9(h), and a comparison with the Google Earth image further illustrates the effectiveness of our optimization method. High consistency implies that the optimized map obtained using the remote sensing mapping results and provincial statistics is still reasonable at the county level, suggesting the reliability of the optimized maps.

Figure 9. Comparison of the consistency of classified and statistical areas before and after optimization using the comparison of (a)–(e) area and (g) Area Angle Distance (AAD) of (f) five provinces, and (h) spatial distribution of forest percentage in Huanglong County, Yan’an, Shaanxi Province.

5. Discussion

This study aimed to develop LULC map updating and optimization methods using Landsat imagery and land survey statistics. The results show that the proposed method provides an efficient and reliable framework for updating the maps of existing LULC products. Additionally, considering the generalizability of the CCDC, LandTrendr, and random forest algorithms in this method and the high global coverage of Landsat imagery, our method can potentially be extended to other 30-m LULC products. For products with coarse spatial resolution, such as the 500-m MCD12Q1 dataset, matching 500-m MODIS imagery should be used in the updating framework. However, for 10-m products, such as ESA WorldCover, the change detection results should be given more attention, considering the time-series fitting difficulties due to the shortage of Sentinel imagery before 2017.

Based on the updated LULC maps, a map optimization method was developed using provincial land survey statistics and a minimum cross-entropy formulation. According to the evaluation results with county-level statistics, our method efficiently improved the consistency between the remote sensing classification results and statistics, thus providing a solution for downscaling land survey statistics. The optimized results showed that the gaps between the classification results and statistics varied spatially and between LULC types, and the differences were further quantified based on statistics and classification maps for 2021 using Equations (7)–(9). (7) $Dif{f}_{i,j}={\displaystyle \frac{ClassifiedAre{a}_{i,j}-StatisticAre{a}_{i,j}}{StatisticAre{a}_{i,j}}}$(7) (8) $Dif{f}_i=\sqrt{\sum _{j=1}^{n}Dif{f}_{i,j}^2}$(8) (9) $Dif{f}_j=\sqrt{\sum _{i=1}^{m}Dif{f}_{i,j}^2}$(9) where $i$ and $j$ represent the province and LULC type identifier, respectively, $n$ and $m$ are the numbers of LULC classes and provinces, respectively, $ClassifiedArea$ is the area obtained from remote sensing classification maps, and $StatisticArea$ is the area obtained from land survey statistics.

The analysis revealed that Xinjiang, Gansu, and Chongqing have the greatest differences in LULC mapping, as shown in Figure 10(a). However, the classification maps showed better agreement with the statistical data for the eastern part of the Hu line (Hu 1935). In contrast, the northwestern provinces exhibited greater differences between the classification maps and the statistics. Our analysis also revealed that these provinces tended to have greater discrepancies in the extent of unused land and grasslands. This finding is consistent with the observed differences in the LULC types. Specifically, by applying Equation (9) to different LULC types, the most significant differences were observed in unused land and grassland (with differences of 115.29 and 60.49, respectively), which was also evident from the linear fit results in Figure 10(b). Confusion between unused land and grassland in LULC mapping tasks is frequent and could be the reason for the observed discrepancies. Moreover, croplands and grasslands were classified more extensively than indicated in the statistics, whereas the opposite was true for other land types except for unused land. However, there was a greater consistency for LULC types with larger areas, which may be attributed to the greater ease of mapping and classification of larger land masses.

Figure 10. The differences between remote sensing classification maps and provincial land survey statistics for each LULC type in Chinese provinces. (a) Differences for each province; (b) linear fit results for each LULC type.

Although helpful and practical methods for map updating and optimization were developed in this study, there were certain limitations. First, because the update approach proposed in this study is based on current LULC products, it is feasible to transfer the classification error of the original map to the updated map. Furthermore, the pixel-level change detection and classification approach introduced ‘salt and pepper’ issues into the updated findings, which are frequent in pixel-level mapping assignments. Spatial filtering is a typical suppression strategy; however, it reduces the range of detected changed areas. Furthermore, the potential of the CCDC algorithm remains to be fully explored. Although the CCDC is generally employed for yearly LULC change detection, it is capable of finer time-scale change detection (such as monthly or even daily), which is still underutilized. However, although the optimization results obtained in this study were consistent with the land survey statistics, there is potential for improvement. For example, the corrosion and expansion algorithm can be used to search for adjacent grids to improve patch smoothness in the optimization results.

6. Conclusions

This study presents a comprehensive framework for updating and optimizing LULC products using a combination of continuous LULC monitoring and national land survey statistics. The proposed method involves integrating change detection results from CCDC and LandTrendr, which are used as the foundation for potential change areas. Subsequently, a sample migration method and a random forest classifier were used to update the LULC product of CLUD-A from 1980–2015 to 1980–2022. The updated maps show smooth changes compared to the original dataset, indicating the effectiveness of the proposed approach. Additionally, this study introduced a statistics space allocation-based map optimization method using the IDW algorithm and minimum cross-entropy formulation based on SNLS and TNLS statistics. This method significantly improved the consistency between the LULC classification maps and land survey statistics, resulting in an improved correlation coefficient ($r$) in Shaanxi from 0.44 to 0.94. The proposed framework provides efficient and practical methods for updating current LULC products, bridging the gap between classification maps and land survey statistics, and has vital implications for continuous LULC change monitoring and sustainable management of land resources.

Disclosure statement

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

Data availability statement

The data supporting the findings of this study are available from the corresponding author upon reasonable request.

Additional information

Funding

This work was supported by Fundamental National Key R&D Program of China (grant number 2019YFA0606601); Tsinghua University Initiative Scientific Research Program (grantnumber 20223080017); National Natural Science Foundation of China (grant number 42201367); Fundamental ResearchFunds for the Central Universities (grant number DUT23RC(3)064.