Long-term changes in dry-season water bodies and lake-littoral land-cover structure of typical lakes in the middle-lower Yangtze and Huai River basins, China (1999–2019)

Long-term monitoring of dry-season water bodies and lake–littoral land-cover structure is important for lake conservation, shoreline management, and wetland protection. Based on the Google Earth Engine platform, this study used Landsat time-series imagery from 1999 to 2019 and combined Continuous Change Detection and Classification (CCDC) with a Random Forest classifier to generate annual land-cover products for Taihu Lake, Hongze Lake, Poyang Lake, and Dongting Lake. A fixed representative date within the January–February dry-season window was used to extract dry-season land-cover maps, and all area and transition statistics were calculated within fixed lake–littoral analysis boundaries. Accuracy assessment was conducted for three representative years, namely, 1999, 2010, and 2019, corresponding to the beginning, middle, and end of the study period. The overall accuracy values were 90.10%, 91.30%, and 90.40%, respectively, with Kappa coefficients of 0.872, 0.886, and 0.875. The results revealed clear inter-lake differences in dry-season water-area change. From 1999 to 2019, Poyang Lake showed the largest decline, with a net loss of 136.95 km 2 (−14.67%), whereas Taihu Lake, Hongze Lake, and Dongting Lake showed relatively small net increases of 21.07 km 2 (+0.85%), 36.29 km 2 (+3.39%), and 13.00 km 2 (+2.75%), respectively. Land-cover transition analysis showed that conversions among water, mudflat, and field were the dominant processes of lake–littoral restructuring, including 141.40 km 2 of water converted to mudflat and 89.67 km 2 of mudflat converted to field. These results suggest that dry-season lake–littoral dynamics may be associated with hydrological variability, shoreline land use, and local management or restoration activities. However, these relationships should be interpreted as plausible associations rather than direct causal attribution. Overall, the CCDC–RF framework, combined with fixed dry-season extraction and transition-matrix analysis, provides a reproducible approach for monitoring long-term dry-season water-body dynamics and lake–littoral structural change.

The middle and lower reaches of the Yangtze River and the adjacent Jianghuai region are characterized by a dense distribution of lakes, and their lake–wetland systems play important roles in flood regulation, water supply, biodiversity conservation, and regional ecological security. Wetland ecosystems support key ecosystem services through hydrological regulation, nutrient cycling, and habitat provision, and their structure and functions are highly sensitive to hydrological fluctuations and land-use change (Mitsch and Gosselink, 2015; Junk et al., 1989). Over the past 2 decades, under the combined influence of climate variability and human activities, lake water extent, littoral transition-zone morphology, and surrounding land-use patterns have undergone varying degrees of adjustment, thereby affecting wetland habitat stability, water-environment risks, and shoreline management effectiveness. To meet the need for refined lake protection and management, it is necessary to continuously and quantitatively characterize the long-term evolution of lake land–water patterns and to identify the major change pathways and inter-lake differences.

Previous studies have shown that remote sensing can provide consistent observations over large spatial extents and long time spans, making it an important means for monitoring changes in lakes and wetlands (Wulder et al., 2012; Zhu and Woodcock, 2014). Long-term surface-water monitoring has increasingly benefited from dense optical satellite archives and cloud-based computing platforms. Global and large-scale studies have demonstrated that Landsat time series can effectively reveal long-term surface-water occurrence, seasonality, and change at regional to global scales. In China, recent studies have also shown substantial surface-water changes in rapidly urbanizing and highly managed regions, such as the Yangtze River Delta, highlighting the importance of long-term monitoring for understanding regional water-body dynamics. These studies provide an important background for using multi-decadal Landsat observations to examine lake-area change and lake–shore transition processes (Pekel et al., 2016; Donchyts et al., 2016; Sogno et al., 2022; Chen et al., 2023).

Existing remote sensing studies of lake systems can generally be divided into three categories. The first focuses on waterbody/shoreline delineation and area-change statistics, relying on long-term image series to reveal lake expansion–shrinkage processes and abrupt change events, while also providing global evidence of long-term surface-water dynamics (Pekel et al., 2016). Meanwhile, the development of basic geographic databases for lakes, reservoirs, ponds, and wetlands has provided important baseline support for regional comparative studies and cross-scale assessments (Lehner and Döll, 2004). The second category emphasizes changes in littoral zones and wetland landscape patterns, with particular attention to the reorganization of mosaic structures such as “water–mudflat/wetland–vegetation/field–built-up land”. Related studies have noted that remote-sensing identification and change interpretation in transition zones pose common challenges, including the presence of mixed objects and confusion among land-cover types (Ozesmi and Bauer, 2002). The third category addresses long-term change trends and management implications in a more integrated manner, showing that the long-term decline in global wetland area and the risk of functional degradation remain prominent, while water security and river biodiversity are under multiple pressures; watershed development and altered hydrological connectivity further reshape the hydrological processes and ecological responses of lake–river systems (Davidson, 2014; Vörösmarty et al., 2010; Grill et al., 2019). However, lake systems in the middle and lower reaches of the Yangtze River are highly complex. First, the lake area is subject to strong seasonal fluctuations and interannual variability, and “snapshot-based” comparisons based on only a few years or sparse observations may easily misinterpret short-term water-level anomalies as long-term trends. Second, littoral transition zones commonly exhibit spectral mixing and frequent land–water conversion. When temporary exposure/inundation coexists with persistent land-cover conversion, the conventional two-date post-classification comparison is prone to confusion about change types and to insufficient interpretation of change pathways. Therefore, constructing annual land-cover products and conducting land-cover transition analysis are key steps in advancing research from simple descriptions of area increase or decrease to the identification of structural reorganization and change pathways, and are also more consistent with the practical needs of lake management for shoreline regulation and wetland protection.

Long-term remote-sensing time-series analysis provides a clear direction for addressing the above issues. By using continuous, multi-year, multi-temporal observations, temporal consistency constraints can be strengthened, thereby reducing the influence of single-date noise and anomalous hydrological conditions on results and supporting the identification of both abrupt and gradual changes (Verbesselt et al., 2010). Among the various time-series frameworks, Continuous Change Detection and Classification (CCDC) performs segmented modeling and change diagnosis on long Landsat time series, making it possible to extract stable temporal features under the interference of clouds, cloud shadows, missing observations, and seasonal fluctuations, thus providing more consistent feature representations for change-event identification and subsequent classification (Zhu et al., 2019). Compared with other temporal segmentation approaches, such as LandTrendr, CCDC places greater emphasis on continuous-observation modeling and change diagnosis, making it suitable for consistent annual mapping and change characterization in long-term time-series contexts (Kennedy et al., 2010). In terms of classification, random forests have been widely used in remote-sensing land-cover mapping because of their adaptability to high-dimensional features, robustness to noise, and strong generalization ability, and have been systematically reviewed in previous studies (Breiman, 2001; Belgiu and Drăguţ, 2016; Pal, 2005). In parallel, machine-learning algorithms have been widely used for land-use and land-cover mapping and temporal change detection. Random Forest and related classifiers have shown robust performance in LULC modeling because of their ability to handle nonlinear relationships and multi-source predictor variables. Comparative studies using Landsat and Sentinel imagery have also demonstrated the potential of optical satellite time series for producing LULC maps. More recent machine-learning-based temporal change-detection studies using high-resolution imagery further indicate that multi-temporal classification can effectively identify persistent land-cover transitions in complex landscapes. These studies provide methodological support for combining CCDC-derived temporal features with Random Forest classification in lake–littoral environments (Aryal et al., 2023; Acar and Zengin, 2023; Acar et al., 2025). At the same time, cloud-computing platforms such as Google Earth Engine (GEE) have greatly lowered the threshold for acquiring, preprocessing, and batch-mapping long-term Landsat data, making it possible to construct annual products and conduct consistency analysis at regional to basin scales (Gorelick et al., 2017). Google Earth Engine has further improved the feasibility of long-term lake and wetland monitoring by enabling efficient processing of multi-temporal Landsat and Sentinel imagery. Previous studies have estimated lake water extent at long-term and bimonthly scales using Landsat data and GEE, demonstrating the value of dense time-series analysis for capturing intra-annual and interannual water dynamics. Other studies have used multi-temporal Landsat data for long-term water-surface monitoring and applied multi-index classification approaches for wetland mapping in different geographic environments. These studies show that cloud-based platforms and multi-temporal optical imagery are increasingly important for monitoring dynamic inland water and wetland systems (Zhang et al., 2022; Purnama et al., 2023; De Mello et al., 2024). Therefore, the framework of “CCDC temporal features + random forest annual classification + land-cover transition statistics” has the potential to more robustly characterize the evolution of land–water patterns under seasonal lake-level fluctuations and to provide reproducible quantitative evidence for inter-lake comparison.

Despite these advances, several gaps remain in long-term lake–littoral monitoring. First, many studies focus primarily on open-water extent, while transitions among water, mudflat, field, built-up land, and tree-covered areas within lake–shore transition zones are less frequently quantified. Second, highly seasonal lakes require temporally comparable mapping strategies, because annual composites may mix wet-season and dry-season land-cover states. Third, few studies combine segment-based time-series change detection, annual dry-season land-cover extraction, and transition-matrix analysis across multiple large lakes with different hydrological and human-use settings. To address these gaps, this study developed a CCDC–RF framework to map dry-season land-cover dynamics and lake–littoral transitions in Taihu Lake, Hongze Lake, Poyang Lake, and Dongting Lake from 1999 to 2019.

Study area

The study area is located in the middle and lower Yangtze River Plain and the adjacent Jianghuai region. Taihu Lake, Hongze Lake, Poyang Lake, and Dongting Lake were selected as representative lake case areas. This region is characterized by a dense distribution of lakes and close hydrological connections within river–lake systems. Some lakes also maintain significant hydraulic connectivity with the main stem of the Yangtze River, making this area a typical lake–wetland composite region in China (Wang et al., 2022).

The study area is strongly influenced by the East Asian monsoon, and precipitation is unevenly distributed throughout the year, with the highest concentration in spring and summer and relatively limited in autumn and winter. As a result, lake water levels and water extent exhibit pronounced seasonal variation (Hao et al., 2015). For river-connected lakes, this seasonal process is particularly evident during the dry season, when declining water levels are usually accompanied by lake-surface contraction and the extensive exposure of shoals and littoral zones (Yao et al., 2016; Guo et al., 2022; Yuan et al., 2015).

Overall, all four lakes experience a seasonal transition from the wet season to the dry season, with the dry season mainly occurring from winter to early spring of the following year. Among them, Poyang Lake and Dongting Lake are more strongly affected by both river–lake interactions and basin inflow processes and therefore show more pronounced seasonal water-level fluctuations, whereas Taihu Lake and Hongze Lake exhibit different ranges of water-level variation under the combined effects of regional precipitation changes and human regulation (Yao et al., 2016; Guo et al., 2022; Yuan et al., 2015; Zhang et al., 2023). Against this background, the four lakes were selected for comparative analysis in this study, and a unified observation window was adopted to characterize the structural features of dry-season water bodies and littoral zones in typical lakes.

Definition of the fixed lake–littoral analysis zone

In this study, the term “littoral zone” refers to an operational lake–shore transition analysis zone defined for remote-sensing monitoring, rather than a bathymetric littoral belt determined by water depth. Because continuous and spatially consistent long-term water-depth data and inundation-frequency data were not available for all four lakes from 1999 to 2019, a fixed lake–littoral analysis boundary was used for each lake.

The fixed analysis boundary was designed to include the main open-water body and the surrounding lake–shore transition areas where water, exposed mudflat, field, built-up land, and tree-covered land may interact or transform over time. This boundary was kept unchanged for all years from 1999 to 2019. Therefore, all annual land-cover maps, dry-season water-area statistics, and land-cover transition matrices were calculated within the same spatial extent for each lake.

Under this definition, the littoral zone in this study includes seasonally exposed mudflats, shallow lakebed areas, nearshore field, built-up land, and tree-covered areas adjacent to the lake within the fixed analysis boundary. This operational definition ensures temporal consistency in long-term area statistics and allows land-cover transitions among water, mudflat, field, built-up land, and tree classes to be compared across years. The fixed lake–littoral analysis boundaries of the four lakes are shown by the black outlines in Figures 1c1–c4.

Data sources and preprocessing

Landsat remote-sensing data and the GEE processing platform

This study focused on Dongting Lake, Poyang Lake, Taihu Lake, and Hongze Lake and conducted long-term remote-sensing monitoring and time-series analysis from 1999 to 2019. The remote-sensing data were obtained from the Landsat surface reflectance products provided by the Google Earth Engine (GEE) platform (Collection 2 Tier 1 Level-2), including Landsat 5 TM, Landsat 7 ETM+, and Landsat 8 OLI. After filtering by the study-area boundaries and time range, a total of 1921 Landsat 5 TM scenes (1999–2011), 309 Landsat 7 ETM + scenes (2012–2013), and 1291 Landsat 8 OLI scenes (2013–2019) were acquired, thereby constructing a multi-sensor long-term remote-sensing dataset covering 1999–2019 and providing a consistent data basis for analyzing changes in lake water bodies and littoral-zone structure.

During preprocessing, a Landsat time-series image collection was established in Google Earth Engine for each lake region. All preprocessing steps were implemented in GEE using the same workflow for Taihu Lake, Hongze Lake, Poyang Lake, and Dongting Lake over the entire 1999–2019 study period. The images were first filtered by study-area boundary and acquisition date. To ensure consistency among sensors, common optical bands available across Landsat 5 TM, Landsat 7 ETM+, and Landsat 8 OLI were selected, including BLUE, GREEN, RED, NIR, SWIR1, and SWIR2. The band names were standardized before CCDC time-series modeling so that observations from different Landsat sensors could be jointly used.

Clouds, cloud shadows, saturated pixels, and other low-quality observations were removed using the quality information associated with the Landsat surface reflectance products. In addition, the CCDC temporal masking procedure was applied using GREEN and SWIR2 as TMask bands to further reduce the influence of residual clouds, cloud shadows, and abnormal observations on time-series fitting. For Landsat 7 ETM + observations affected by the scan-line corrector failure, missing scan-line pixels were treated as invalid observations rather than being manually interpolated. The CCDC model used all remaining valid observations in the long-term time series, thereby reducing the influence of individual missing observations and data gaps on temporal segmentation and subsequent classification.

Assessment of Landsat observation availability

To evaluate the temporal support of the Landsat time series for CCDC fitting and dry-season land-cover extraction, Landsat scene availability was summarized for each lake and each year from 1999 to 2019. The statistics were calculated within the fixed lake–littoral analysis boundary of each lake. In addition to the annual Landsat scene count, the number of scenes available within the January–February dry-season window was also counted to assess the temporal support for dry-season map extraction.

This scene-availability assessment was used as a quality-control step rather than as an explanatory factor for land-cover change. Because the CCDC-based classification was derived from long-term temporal segments fitted using all available observations, detected land-cover transitions were not interpreted as being caused by the number of available images. Instead, scene availability and sensor periods were examined to evaluate whether major interannual fluctuations coincided with obvious data gaps or sensor-transition periods. The annual and January–February Landsat scene-availability statistics for representative fluctuation years are provided in Supplementary Table S1.

Field survey and reference sample interpretation

Field surveys were conducted during 2015–2019 to support the interpretation of typical lake–littoral land-cover types and to improve the reliability of reference sample labeling. The surveys covered representative areas around Taihu Lake, Hongze Lake, Poyang Lake, and Dongting Lake, with particular attention to land–water transition zones where spectral confusion is common. During the surveys, field photographs, UAV observations, geographic locations, and local land-cover knowledge were collected to document the visual and spatial characteristics of water, built-up land, mudflat, field, and tree-covered areas. Representative field photographs used to support land-cover interpretation are shown in Figure 2, and the spatial distribution of the field survey sites is shown in Figure 1.

The field survey information was used as auxiliary evidence rather than as the sole source of reference samples. Specifically, field observations helped establish interpretation rules for visually distinguishing spectrally similar classes in remote-sensing imagery. For example, exposed mudflats were identified based on their location near the lake margin, bare or sparsely vegetated surface, and seasonal exposure characteristics; field areas were interpreted based on regular plot boundaries, vegetation texture, and agricultural patterns; built-up land was identified by impervious surfaces, roads, and settlement structures; and tree-covered areas were identified by stable canopy texture and high vegetation signals. These field-supported interpretation rules were then used to guide sample labeling in Google Earth Engine.

Based on these interpretation criteria, reference samples were generated through visual interpretation in Google Earth Engine using Landsat imagery, Google Earth historical high-resolution imagery, available GF imagery, and field-supported knowledge of typical land-cover characteristics. In total, 2686 visually interpreted reference sample points were generated and checked across the four lake regions, including 541 for Taihu Lake, 506 for Poyang Lake, 1024 for Hongze Lake, and 615 for Dongting Lake. The reference sample statistics are summarized in Table 1. These samples covered five land-cover classes: water, built-up land, mudflat, field, and tree. The samples were then divided into training and validation subsets for Random Forest classification and accuracy assessment.

For historical validation years earlier than the field survey period, reference labels were assigned through retrospective visual interpretation. Landsat imagery, Google Earth historical high-resolution imagery, available GF imagery, and field-supported interpretation rules were jointly used to determine the reference class. When interpreting historical samples, priority was given to image evidence closest to the validation year, and samples with ambiguous land-cover history or inconsistent evidence among data sources were excluded from the validation set.

Auxiliary data

To support topographic characterization of the study area, sample selection, and verification of classification results, this study introduced NASA SRTM digital elevation model (DEM) data as auxiliary topographic information. These data were combined with high-resolution Google Earth imagery and GF-1/2/6 imagery for the interpretation and cross-checking of representative land-cover types. The auxiliary data were primarily used to confirm ambiguous classes, verify training samples, and assess the plausibility of classification results, thereby improving the reliability of the annual land-cover maps and the consistency of change interpretation.

Overall framework

Using Taihu Lake, Hongze Lake, Poyang Lake, and Dongting Lake as representative study objects, this study constructed a long-term Landsat observation series for 1999–2019 and adopted the technical route of “Continuous Change Detection and Classification (CCDC)–annual land-cover mapping–area and land-cover transition statistics” to characterize the long-term evolution of land–water patterns in lake and littoral-zone systems. Five land-cover classes were defined within the fixed lake–littoral analysis boundary: water, built-up land, mudflat, field, and tree. Here, “mudflat” mainly represents frequently alternating land–water transition surfaces, including exposed lakebeds, bare or sparsely vegetated shoals, and seasonally exposed littoral areas. All annual classification, dry-season water-area calculation, and land-cover transition analyses were conducted within the same fixed boundary for each lake. The overall methodological workflow is shown in Figure 3.

CCDC temporal segmentation

The Continuous Change Detection and Classification (CCDC) algorithm was proposed by Zhu and Woodcock, (2014). Designed for medium-resolution long-term time-series data such as Landsat, it enables pixel-level time-series fitting and segmentation and can identify both gradual and abrupt changes. The CCDC framework is typically organized as a workflow of data preprocessing, cloud/shadow masking, temporal model construction, change-point detection, and classification mapping, thereby linking change detection with land-cover classification. Given the strong seasonal water-level fluctuations and complex transitional classes in the study lakes, the implementation in this study followed the steps below.

Data preprocessing

The Landsat data sources and preprocessing procedures used in this study were described in Section 2.2. Before applying CCDC, the inputs were organized as quality-controlled surface-reflectance time series, together with necessary spectral indices and composite features, at the pixel level for the four lake study areas, thus providing stable inputs for subsequent model fitting and change detection.

Cloud/Cloud-shadow masking and outlier control

To reduce the influence of clouds, cloud shadows, thin clouds, and occasional abnormal observations on time-series fitting, contaminated observations were removed based on quality bands and cloud-detection results, and residual noise was further reduced using temporal-consistency constraints. This made model fitting more robust and prevented short-term anomalies from being misidentified as real land-surface change.

Harmonic time-series model construction

The Continuous Change Detection and Classification (CCDC) algorithm was used to model the long-term spectral trajectory of each pixel and to identify temporal breakpoints in the Landsat time series. Instead of relying on single-date observations, CCDC fits a harmonic time-series model for each stable temporal segment, which makes it suitable for long-term monitoring under seasonal fluctuations and irregular valid observations. For the i th Landsat band, the fitted reflectance within the k th temporal segment can be expressed as shown in Equation 1: where ρ ^ i x is the predicted surface reflectance of the, i -th Landsat band at time x, x is the Julian date, T is the number of days in 1 year, a 0, i represents the overall spectral level of the segment, a 1, i and b 1, i represent intra-annual seasonal variation, c 1, i represents the interannual trend, and τ k denotes the k -th temporal breakpoint.

Change-point detection and time-series segmentation

Temporal breakpoints were detected by comparing new valid Landsat observations with the values predicted by the fitted model. The deviation between observed and predicted reflectance was evaluated using a normalized residual criterion. A simplified multi-band form can be written as shown in Equation 2: where m is the number of spectral bands used for breakpoint detection, ρ i x is the observed surface reflectance of the i -th band at time x, ρ ^ i x is the model-predicted reflectance, and R M S E i is the root mean square error of the fitted model for the i -th band. In the GEE implementation of CCDC, the final change decision is further controlled by the specified breakpoint bands, the chi-square probability threshold, and the minimum number of valid observations.

In this study, the common Landsat optical bands BLUE, GREEN, RED, NIR, SWIR1, and SWIR2 were used in the CCDC workflow. The breakpoint detection bands were GREEN, RED, NIR, SWIR1, and SWIR2, while GREEN and SWIR2 were used as TMask bands to reduce the influence of residual clouds, cloud shadows, and abnormal observations. The main CCDC parameters used in this study are listed in Table 2.

After CCDC temporal segmentation, the fitted coefficients of each stable segment were used as spectral-temporal predictors for land-cover classification. The classification was therefore conducted at the CCDC segment level rather than on single-date images or annual composite images. For each Landsat spectral band, the retained CCDC coefficients included the intercept, slope, seasonal harmonic coefficients, and RMSE. These variables describe the spectral level, long-term trend, seasonal variation, and fitting uncertainty of each pixel during a stable temporal segment.

Segment-based Random Forest classification

In this study, the input spectral bands included BLUE, GREEN, RED, NIR, SWIR1, and SWIR2. For each band, the CCDC-derived coefficients used as predictors included INTP, SLP, COS, SIN, COS2, SIN2, COS3, SIN3, and RMSE. The final predictor set therefore combined spectral information and temporal information. The spectral information was represented by the Landsat bands, while the temporal information was represented by the CCDC coefficients that characterize the behavior of each pixel over time. This segment-based feature construction helps reduce the influence of single-date noise, temporary water-level anomalies, and cloud-related residual errors, which are common in lake and littoral-zone environments.

A Random Forest classifier was then trained to classify the CCDC segments into five land-cover classes: water, built-up land, mudflat, field, and tree. The classifier was implemented in Google Earth Engine using the smileRandomForest algorithm. Based on the GEE implementation used in this study, the number of trees was set to 300. The training samples were interpreted from field survey data and auxiliary high-resolution imagery, and 70% of the samples were used for classifier training, while the remaining 30% were reserved for independent validation. The trained classifier was applied to all CCDC temporal segments within the fixed analysis boundary of each lake. The resulting classified segments were subsequently used to extract date-specific annual land-cover maps. The predictor variables and Random Forest settings used for segment-based classification are summarized in Supplementary Table S2.

Date-specific annual land-cover map generation

Annual land-cover maps were generated from the classified CCDC temporal segments rather than from conventional annual image composites. After each CCDC segment was assigned a land-cover label by the Random Forest classifier, a date-specific land-cover map was extracted for each target year. In this study, a fixed representative date within the January–February dry-season window was used for all years and all lakes. For a given year, the land-cover label of the CCDC segment valid at the selected dry-season date was assigned to each pixel. In other words, the class of each pixel in a given year was determined by the classified temporal segment covering the fixed dry-season reference date.

This procedure differs from annual compositing because it does not aggregate all observations within a year into a single spectral composite. Instead, it uses the long-term CCDC model to determine the temporally valid segment and then extracts the corresponding classified label at the target date. Therefore, for highly dynamic lakes such as Poyang Lake and Dongting Lake, the annual product used in this study represents land-cover conditions during the same dry-season reference window, rather than a mixture of wet-season and dry-season states.

Because CCDC fits temporal segments using all available valid observations in the long-term Landsat time series, the date-specific dry-season label does not require a cloud-free Landsat image exactly on the selected reference date. Instead, the land-cover label is derived from the classified CCDC segment that is temporally valid at that date. This design reduces the influence of occasional missing observations, cloud contamination, and Landsat 7 ETM + SLC-off gaps on annual dry-season map generation.

Land-cover transition analysis and accuracy assessment

Extraction of dry-season water area

To characterize lake-area change under temporally comparable hydrological conditions, this study used a fixed representative date within the January–February dry-season window to extract dry-season land-cover maps for each year from 1999 to 2019. The dry-season water area was not derived from an independent classification using only January–February images, nor from a temporally generalized annual composite. Instead, after the CCDC temporal segments were classified, the land-cover label of the segment valid at the selected dry-season date was extracted for each year. The water class in this date-specific map was then used to calculate dry-season water area.

Unless otherwise specified, the term “lake area” in this study therefore refers to dry-season water area under the January–February low-water condition, rather than annual mean water area or maximum inundation extent. This design was adopted because the dry season is the period when shallow lakebeds, mudflats, and lake–shore transition zones are more likely to be exposed, making it suitable for identifying water–mudflat–field transitions and littoral-zone restructuring. In addition, using a fixed dry-season reference date improves the temporal comparability of water-area statistics among different years and lakes. However, this study does not represent maximum wet-season inundation, and this limitation is acknowledged in the Discussion. Unless otherwise specified, the term “lake area” in the subsequent analysis refers to dry-season water area. Based on the annual land-cover maps, overlay statistics were used to quantify transitions among major land-cover classes from 1999 to 2019, thereby identifying the dominant pathways of structural reorganization in the littoral zones.

Land-cover transition matrix calculation

Land-cover transition matrices were generated by overlaying the exported dry-season classification maps of 1999 and 2019 in ArcGIS. Pixel-based cross-tabulation was used to quantify the transition area between each pair of land-cover classes. In the transition matrix, rows represent land-cover classes in 1999, columns represent land-cover classes in 2019, and each matrix element represents the area converted from one class in 1999 to another class in 2019. Diagonal elements represent unchanged areas, whereas off-diagonal elements represent land-cover conversions.

The transition area from class i in 1999 to class j in 2019 was calculated as shown in Equation 3: where M i j is the transition area from class i in 1999 to class j in 2019 in km 2, N i j is the number of pixels converted from class i to class j, and 30 × 30/10 6 converts the area of a 30 m Landsat pixel from m 2 to km 2.

To avoid ambiguity, Total outflow and Total inflow were calculated using only off-diagonal elements. For class i, Total outflow refers to the area that belonged to class i in 1999 but changed to other classes by 2019, and Total inflow refers to the area that belonged to other classes in 1999 but changed to class i by 2019. Total outflow, Total inflow, and Net change were calculated as shown in Equations 4-6: where O i is the Total outflow of class i, I i is the Total inflow of class i, and N i is the Net change of class i. The unchanged area M j i was excluded from both Total outflow and Total inflow.

Accuracy assessment

To evaluate the reliability of the long-term land-cover products, three representative years were selected for accuracy assessment: 1999, 2010, and 2019. These years correspond to the beginning, middle, and end of the 1999–2019 study period and were used as temporal checkpoints to examine classification performance across the long-term mapping sequence. The purpose of selecting these 3 years was to provide representative accuracy evidence for the early, intermediate, and final stages of the time series, rather than to imply that classification uncertainty was identical in every year.

Reference information for the selected validation years was obtained by integrating visually interpreted samples, Landsat imagery, Google Earth historical high-resolution imagery, available GF imagery, and field-supported interpretation knowledge. For 2019, contemporaneous high-resolution imagery and field-supported land-cover knowledge provided relatively direct reference support. For 1999 and 2010, reference labels were assigned through retrospective interpretation using historical imagery closest to the validation years. Samples with unclear historical land-cover states or inconsistent evidence among data sources were excluded.

Accuracy indicators, including User’s Accuracy (UA), Producer’s Accuracy (PA), Overall Accuracy (OA), and the Kappa coefficient, were calculated for the three representative years. In addition, class-specific accuracy and the confusion matrix were further summarized for 2019 to illustrate the main sources of classification uncertainty among different land-cover classes (Supplementary Table S3).

Accuracy assessment results

To verify the reliability of the long-term land-cover maps, accuracy assessment was conducted for three representative validation years: 1999, 2010, and 2019. These years were selected as temporal checkpoints representing the beginning, middle, and end of the 1999–2019 study period. The validation results were used to evaluate the overall performance of the CCDC–RF classification framework across different stages of the long-term mapping sequence.

As shown in Table 3, the classification results achieved stable overall accuracy for the three representative years. The OA values were 90.10%, 91.30%, and 90.40% for 1999, 2010, and 2019, respectively, with corresponding Kappa coefficients of 0.872, 0.886, and 0.875. The UA values ranged from 89.80% to 90.60%, and the PA values ranged from 89.30% to 90.40%. These results indicate that the CCDC–RF framework provided reliable overall classification performance for the selected temporal checkpoints.

To further examine class-level uncertainty, the 2019 validation results were summarized by class (Supplementary Table S3). The 2019 confusion matrix in Supplementary Table S3 showed that water, tree, and field generally achieved high classification accuracy, whereas built-up land and mudflat showed relatively lower accuracy. The main confusion occurred among mudflat, field, and built-up land, which is reasonable in lake–littoral transition zones where exposed wet surfaces, bare land, and agricultural land may have similar spectral characteristics.

The years 1999, 2010, and 2019 were selected as representative temporal checkpoints corresponding to the beginning, middle, and end of the 1999–2019 study period.

Temporal changes and structural transformation of typical lakes

All area statistics and land-cover transition results reported in this section were calculated within the fixed lake–littoral analysis boundary of each lake. Therefore, the reported changes represent land-cover dynamics within the predefined lake–shore transition analysis zone rather than changes within administrative regions or year-specific moving lake boundaries.

This study focused on four representative lakes in the middle and lower reaches of the Yangtze River and the adjacent Jianghuai region, namely, Taihu Lake, Hongze Lake, Poyang Lake, and Dongting Lake. To reduce the influence of intra-annual water-level fluctuations on area statistics, the dry-season water area derived from the annual classification results within the January–February window was used to characterize lake-area change. Unless otherwise specified, the term “lake area” in the following sections refers to the dry-season water area.

Overall trends in lake-area change

During 1999–2019, dry-season water-area changes in the four lakes showed pronounced inter-lake differences (Figure 4). Unless otherwise specified, the dry-season water area reported in this section refers to the water area extracted from the date-specific land-cover maps using the fixed January–February dry-season reference date. Overall, Poyang Lake exhibited a clear decrease in water area, from 933.46 km 2 to 796.51 km 2, resulting in a net loss of 136.95 km 2 (−14.67%). In contrast, Taihu Lake, Hongze Lake, and Dongting Lake showed relatively small overall changes, with net increases of 21.07 km 2 (+0.85%), 36.29 km 2 (+3.39%), and 13.00 km 2 (+2.75%), respectively. In stage-specific terms, Poyang Lake experienced the largest fluctuation and a marked downward tendency, Hongze Lake showed a slight upward trend, and Taihu Lake and Dongting Lake fluctuated within a relatively stable range. These results indicate that the dry-season water dynamics of the four representative lakes differed substantially in both trend and magnitude, reflecting heterogeneity in hydrological conditions, human activities, and lake–river interactions.

Some interannual fluctuations were observed in the dry-season water-area series, including relatively clear changes around 2002–2003 for Hongze Lake and around 2013–2014 for Taihu Lake and Dongting Lake. These fluctuations should not be interpreted as being caused by image availability itself. The CCDC-based classification was derived from long-term temporal segments fitted using all available valid observations, rather than from a single image or a simple annual composite. Therefore, variations in the number of available images are unlikely to directly generate land-cover transitions by themselves. Nevertheless, observation availability and sensor periods were examined as quality-control information to ensure that the detected changes were not associated with obvious data gaps, insufficient valid observations, or sensor-transition artifacts.

For the years with relatively clear interannual fluctuations, the Landsat scene-availability check showed that image data were still available (Supplementary Table S1). For Hongze Lake, annual scene counts were 64 and 45 in 2002 and 2003, respectively, and January–February scene counts were 11 and 12. For Taihu Lake, annual scene counts were 50 and 67 in 2013 and 2014, respectively, although the January–February count was relatively low in 2013. For Dongting Lake, annual scene counts were 85 and 76 in 2013 and 2014, respectively. These results indicate that the detected fluctuations should not be interpreted simply as artifacts of image absence. However, because 2013 corresponded to the Landsat 7–Landsat eight transition period and had relatively limited dry-season scenes for some lakes, these years were interpreted cautiously.

Characteristics of internal land-cover transformation within the lakes

In the following transition matrices, rows indicate land-cover classes in 1999 and columns indicate land-cover classes in 2019. The diagonal values represent unchanged areas, while off-diagonal values represent land-cover conversions. Total outflow and Total inflow were recalculated using only off-diagonal values, so unchanged areas were not included in either metric. Net change was calculated as Total inflow minus Total outflow.

From 1999 to 2019, land-cover transitions in the four lakes generally showed a structural reorganization dominated by conversions among water, mudflat, and field (Table 4). At the aggregate level, 141.40 km 2 of water was converted to mudflat and 18.53 km 2 of water was converted to field. At the same time, 72.31 km 2 of mudflat was converted to water and 89.67 km 2 of mudflat was converted to field. Built-up expansion also contributed to littoral-zone restructuring, with 23.44 km 2 of water and 90.83 km 2 of mudflat converted to built-up land. After excluding unchanged diagonal areas, mudflat had the largest Total outflow among all classes, reaching 271.70 km 2, while built-up land and field showed positive net changes of 71.34 km 2 and 27.44 km 2, respectively. These results indicate that the lake–littoral transition zones underwent both hydrologically related water–mudflat conversions and land-use-related conversions toward built-up land and field.

To further reveal inter-lake differences, the corrected land-cover transition matrices for each lake are presented in Supplementary Appendix Tables A1–A4. Poyang Lake showed the most pronounced water-to-mudflat transition, with 117.99 km 2 converted from water to mudflat, and also exhibited substantial conversions from mudflat to built-up land and field. Dongting Lake mainly exhibited bidirectional transitions between mudflat and water, with 26.47 km 2 converted from mudflat to water and 9.04 km 2 from water to mudflat. Hongze Lake showed prominent field-related transitions, including 37.89 km 2 converted from field to water and 21.59 km 2 from mudflat to field. Taihu Lake exhibited more complex shoreline adjustment, with field-to-water conversion of 13.99 km 2 and mudflat-to-built-up conversion of 11.26 km 2. These inter-lake differences provide quantitative evidence for interpreting lake-specific littoral-zone restructuring.

The long-term CCDC–RF classification results revealed substantial changes in dry-season water area and lake–littoral land-cover structure among the four lakes from 1999 to 2019. These changes were quantified using a temporally consistent fixed lake–littoral analysis boundary and a fixed representative date within the January–February dry-season window, which improved the comparability of interannual and inter-lake comparisons under low-water conditions. However, the remote-sensing results primarily document spatial and temporal patterns of land-cover change. They should therefore be interpreted as evidence of lake–littoral dynamics rather than direct proof of specific causal mechanisms. In the following sections, hydrological variability, shoreline land use, and possible governance-related influences are discussed as plausible explanations, but causal attribution should be interpreted cautiously.

Hydrological variability and dry-season water–mudflat transitions

Hydrological variability is an important background factor affecting dry-season water extent and lake–littoral exposure. In the study area, dry-season water levels and boundary positions may be influenced by interannual precipitation variability, seasonal recession processes, river–lake connectivity, shallow-basin topography, and local water regulation. Because the water class in this study represents dry-season water extent, its interannual variation reflects not only water-storage conditions but also the exposure of shallow lakebeds and littoral flats during low-water periods.

When dry-season water levels remain relatively high or recession is delayed, seasonally exposed shoals may remain inundated, resulting in larger dry-season water area. Conversely, when dry-season water levels are relatively low or recession occurs earlier, shallow flats are more likely to be exposed, which may appear as mudflat or low-vegetation transition surfaces in the classification results. Therefore, the observed water–mudflat transitions are consistent with the influence of hydrological variability, but they should not be interpreted as being caused by hydrological processes alone without additional water-level, discharge, or precipitation evidence (Sogno et al., 2022).

The representative case in Dongting Lake illustrates a typical water–mudflat transition detected by the CCDC-based workflow (Figure 5). The NDVI time series shows a clear breakpoint, and the Landsat image comparison and classification results indicate a shift from an exposed or mudflat-dominated dry-season surface to a more water-dominated condition. At the lake scale, Dongting Lake showed a small net increase in dry-season water area between 1999 and 2019, while mudflat area decreased. This pattern is consistent with the influence of hydrological fluctuation on dry-season exposure and inundation. However, because no independent water-level or discharge series was directly used in the attribution analysis, this case is interpreted as a hydrologically consistent pattern rather than definitive proof of hydrological causality.

Possible effects of shoreline land use and human modification

In addition to hydrologically related water–mudflat fluctuations, lake–littoral land-cover transitions may also be associated with shoreline land use and human modification. Reclamation, cultivation, urban expansion, shoreline engineering, returning farmland to lakes, and wetland rehabilitation can alter the stability and direction of land-cover transitions in nearshore areas. In the classification and transition matrices, these processes may appear as conversions among mudflat, field, built-up land, and water. However, because this study did not directly incorporate detailed engineering records, land-management documents, or socioeconomic datasets into a quantitative attribution model, these interpretations should be regarded as plausible mechanisms rather than confirmed causal relationships.

The representative case in Hongze Lake appears consistent with a transition from field to water or wetter land-cover conditions (Figure 6). The NDVI time series, image comparison, and classification results together indicate a shift from a vegetation- or field-dominated surface toward a more water-influenced condition. At the lake scale, the corrected transition matrix shows that field-to-water conversion exceeded water-to-field conversion. This pattern may reflect local shoreline adjustment, restoration-related land-cover change, or hydrological exposure–inundation variability. However, without independent project-level restoration records or water-management data, the result should be interpreted as a possible restoration-related or hydrologically related pattern rather than a confirmed possible governance-related outcome.

For Poyang Lake, the representative case shows a transition from water or exposed mudflat to built-up land in the lake–shore transition zone (Figure 7). The NDVI time series, image comparison, and classification results together indicate that part of the lake–littoral surface became a more stable artificial surface during the study period. At the lake scale, the corrected transition matrix also shows substantial conversions from mudflat to built-up land. This pattern may be associated with shoreline development or local land-use expansion. However, because planning, construction, and land-management records were not directly incorporated into the analysis, the specific human drivers of these transitions require further confirmation.

The observed water–mudflat–field transitions also align with regional and international wetland transformation studies. Previous studies have shown that wetland transformation in floodplain river basins can affect both ecological conditions and local livelihoods, while remote-sensing-based monitoring can reveal substantial changes in wetland extent, lake boundaries, and surrounding land cover. These studies support the interpretation that lake and wetland margins are often affected by interacting hydrological and land-use processes. In this study, the prominent transitions among water, mudflat, and field in Poyang Lake, Hongze Lake, and Dongting Lake similarly indicate that lake–littoral zones are not static boundaries but dynamic transition systems (Singha and Pal, 2023; Acar and Zengin, 2025; Zengin et al., 2025).

Remote-sensing-detected structural change and possible restoration-related signals

The representative Taihu Lake case shows a clear CCDC breakpoint around 2016, together with corresponding changes in Landsat image texture and classification results (Figure 8). This indicates that the land-cover state of the selected lake–shore area changed during the study period. The observed shift may be consistent with shoreline rehabilitation, ecological buffer-zone construction, or other local management and restoration activities. However, because this study did not directly use project implementation records, policy timelines, or shoreline-management documents to verify the timing and location of specific interventions, the 2016 breakpoint should be interpreted as a remote-sensing-detected land-cover change signal rather than direct evidence of a specific possible governance-related event.

Therefore, this case is used to illustrate how CCDC can detect persistent structural changes in lake–shore transition zones, while the possible governance- or restoration-related explanation remains interpretive and should be further tested with independent management records.

Methodological contribution and transferability

This study contributes to long-term lake and lake–littoral monitoring in three aspects. First, annual dry-season land-cover products were generated from classified CCDC temporal segments rather than from conventional annual image composites. This segment-based approach uses long-term spectral–temporal information and reduces the influence of single-date noise, temporary cloud contamination, and short-term hydrological anomalies. Second, a fixed representative date within the January–February dry-season window was used to extract dry-season land-cover maps, improving temporal comparability among years and lakes under low-water conditions. Third, all area statistics and transition matrices were calculated within fixed lake–littoral analysis boundaries, allowing water, mudflat, field, built-up land, and tree transitions to be compared over a consistent spatial extent.

These methodological choices are particularly useful for shallow lakes and lake–shore transition zones, where land-cover states can vary strongly within a year. By combining CCDC-based temporal segmentation, Random Forest classification, fixed dry-season extraction, and transition-matrix analysis, this study provides a reproducible framework for detecting long-term lake–littoral restructuring in complex hydrological environments.

The results of this study are consistent with broader evidence that surface-water and wetland systems are highly dynamic under climatic variability, hydrological regulation, and land-use pressure. Global studies have shown widespread changes in surface-water extent and occurrence over recent decades, while review work has emphasized that surface-water dynamics are closely connected to climate variability, human water use, and land-cover transformation. Compared with global water-occurrence datasets that mainly emphasize open-water dynamics, the present study further distinguishes transitions among water, mudflat, field, built-up land, and tree-covered areas within fixed lake–littoral analysis boundaries. This provides additional information on how lake margins are reorganized under dry-season low-water conditions (Pekel et al., 2016; Donchyts et al., 2016; Sogno et al., 2022).

The use of CCDC-derived temporal segments and Random Forest classification also connects this study with broader LULC time-series and machine-learning change-detection research. Previous studies have demonstrated the applicability of machine-learning models for LULC performance modeling, the production of time-series LULC maps from Landsat and Sentinel imagery, and temporal change detection using high-resolution imagery. Compared with these studies, the present work focuses specifically on dry-season lake–littoral transition zones and integrates CCDC-based temporal segmentation with annual dry-season extraction and transition-matrix analysis (Aryal et al., 2023; Acar and Zengin, 2023; Acar et al., 2025).

Uncertainty and limitations

Several uncertainties should be acknowledged. First, although CCDC reduces the influence of occasional missing observations by fitting temporal segments from long-term Landsat time series, classification uncertainty may still arise from residual clouds, mixed pixels in land–water transition zones, Landsat 7 ETM + SLC-off gaps, and spectral differences among Landsat sensors. The Landsat scene-availability assessment was therefore used as quality-control information rather than as a direct explanation for land-cover change. Years around the Landsat 7–Landsat eight transition were interpreted cautiously because sensor transition and uneven dry-season image availability may contribute to classification uncertainty.

Second, this study used a fixed representative date within the January–February dry-season window to characterize dry-season water area. This design improves temporal comparability under low-water conditions, but it does not represent annual mean water area or maximum wet-season inundation. Therefore, the results should be interpreted as dry-season lake–littoral dynamics rather than full hydrological variability throughout the year.

Third, 1999, 2010, and 2019 were selected as representative accuracy checkpoints for the beginning, middle, and end of the study period. Although these years provide evidence for the overall reliability of the classification framework, they do not fully characterize year-to-year variation in classification uncertainty. Class-specific accuracy was further examined for 2019, while class-level confusion for historical years was mainly interpreted based on overall accuracy indicators and visual inspection of historical imagery. Therefore, the long-term trends and land-cover transition matrices should be interpreted together with the representative accuracy assessment, spatial pattern consistency, and visual comparison with remote-sensing imagery.

Fourth, the discussion of potential drivers is interpretive. This study did not directly build a quantitative attribution model using long-term water-level, discharge, sediment, engineering-operation, or socioeconomic datasets. Therefore, the relationships between observed land-cover transitions and possible drivers, such as hydrological fluctuation, shoreline development, agricultural use, water regulation, and restoration activities, should be interpreted as plausible associations rather than direct causal conclusions.

Broader water-related land-cover change studies also indicate that inundation, shoreline change, and water-related land loss can be monitored using remote-sensing and machine-learning approaches. Reviews of flood-inundation mapping have shown that machine learning and deep learning methods are increasingly used to detect inundated areas and water-related hazards. Studies on sea-level-rise-related land loss further demonstrate the importance of linking remote-sensing observations with risk and management contexts in water–land transition zones. In addition, projected vegetation change and terrestrial water loss highlight that land-cover change and water availability are closely coupled under environmental change. Although the present study focuses on inland lakes rather than coastal or flood-hazard systems, these studies provide broader context for interpreting water-related land-cover transitions and the need for cautious driver attribution (Silwal et al., 2026; Zengin, 2023; Chai et al., 2025).

This study focused on Taihu Lake, Hongze Lake, Poyang Lake, and Dongting Lake and developed a technical framework for “CCDC-based continuous change detection, annual land-cover mapping, and area and land-cover transition statistics” using Landsat time-series data from 1999 to 2019. Under a unified classification system and a fixed dry-season observation window, the long-term evolution of water bodies and littoral-zone land–water structure in typical lakes was quantitatively characterized.

The results showed clear inter-lake differences in dry-season water-area change. Poyang Lake exhibited an overall decline, whereas Taihu Lake, Hongze Lake, and Dongting Lake remained relatively stable, although all three showed interannual fluctuations. In terms of structural transformation, the evolution of lake–littoral zones was mainly characterized by mutual conversions among water, mudflat, and field, indicating that shallow-water areas and land–water transition zones were sensitive to hydrological variability and shoreline land-use change.

Combined with the threefold evidence chain of breakpoint detection, image comparison, and classification validation in representative areas, the observed changes in the four lakes suggest possible associations with hydrological fluctuation, shoreline land use, and local governance or restoration activities. Natural hydrological processes may explain part of the water–mudflat oscillation, human land use may contribute to the conversion of transition zones into field or built-up land, and restoration or shoreline rehabilitation may be associated with the reconversion of some nearshore land-cover types. However, these interpretations should be regarded as plausible explanations rather than direct causal attribution, because this study did not include a quantitative driver-attribution model.

Overall, the long-term remote-sensing analysis framework developed in this study can effectively support the identification of land–water pattern evolution and inter-lake comparison for typical lakes and can provide quantitative evidence for shoreline regulation, wetland restoration, and zoned management in lake regions of the middle and lower reaches of the Yangtze River and the adjacent Jianghuai region.