Long-Term Dynamics and Transitions of Surface Water Extent in the Dryland Wetlands of Central Asia Using a Hybrid Ensemble–Occurrence Approach

Highlights

What are the main findings?

• First seasonally explicit, multi-decadal (2000–2024) mapping of Ile River Delta (IRD) wetlands
• Multi-index ensemble with dynamic thresholds improves water classification accuracy
• Periodic and ephemeral surface water extent (SWE) dominates; stable SWE covers only 12% of IRD
• Significant summer–fall declines driven by reservoir regulation and warming

What are the implications of the main findings?

• Findings support transboundary water cooperation that explicitly includes environmental flow allocations to protect downstream wetland extent, connectivity, and biodiversity.
• Evidence indicates reservoir and irrigation management should be coordinated basin-wide to better mimic natural seasonal flow regimes and reduce summer to fall wetland contraction.
• Results highlight urgent need for Conservation and Transboundary Cooperation in Arid Central Asia (ACA)

Abstract

Wetlands in dryland regions are rapidly degrading under the combined effects of climate change and human regulation, yet long-term, seasonally resolved assessments of surface water extent (SWE) and its dynamics remain scarce. Here, we map and analyze seasonal surface water extent (SWE) over the period 2000–2024 in the Ile River Delta (IRD), south-eastern Kazakhstan, using Landsat TM/ETM+/OLI data within the Google Earth Engine (GEE) framework. We integrate multiple indices using the modified Normalized Difference Water Index (mNDWI), Automated Water Extraction Index (AWEI) variants, Water Index 2015 (WI2015), and Multi-Band Water Index (MBWI) with dynamic Otsu thresholding. The resulting index-wise binary water maps are merged via ensemble agreement (intersection, majority, union) to delineate three SWE regimes: stable (persists most of the time), periodic (appears regularly but not in every season), and ephemeral (appears only occasionally). Validation against Sentinel-2 imagery showed high accuracy F1-Score/Overall accuracy (F1/OA ≈ 0.85/85%), confirming our workflow to be robust. Hydroclimatic drivers were evaluated through modified Mann–Kendall (MMK) and Spearman’s (r) correlations between SWE, discharge (D), water level (WL), precipitation (P), and air temperature (AT), while a hybrid ensemble–occurrence framework was applied to identify degradation and transition patterns. Trend analysis revealed significant long–term declines, most pronounced during summer and fall. Discharge is predominantly controlled by stable spring SWE, while discharge and temperature jointly influence periodic SWE in summer–fall, with warming reducing the delta surface water. Ephemeral SWE responds episodically to flow pulses, whereas precipitation played a limited role in this semi–arid region. Spatially, area(s) of interest (AOI)-II/III (the main distributary system) support the most extensive yet dynamic wetlands. In contrast, AOI-I and AOI-IV host smaller, more constrained wetland mosaics. AOI-I shows persistence under steady low flows, while AOI-IV reflects a stressed system with sporadic high-water levels. Overall, the results highlight the dominant influence of flow regulation and distributary allocation on IRD hydrology and the need for ecologically timed releases, targeted restoration, and transboundary cooperation to sustain delta resilience.

1. Introduction

Anthropogenic climate change is significantly altering the water balance in dryland settings worldwide [1,2,3]. Rising temperatures, shifts in precipitation patterns, and large-scale land-use modifications such as damming, irrigation, and channel diversions have disrupted natural drainage systems, sediment and nutrient fluxes, and hydro-geomorphological processes [3,4,5]. These disruptions affect river connectivity, alter microclimates, and degrade ecosystem functions, ultimately constraining the movement of organisms, nutrients, and biologically mediated energy flows essential for ecosystem services [5,6,7].

Drylands cover ~41% of the earth’s terrestrial surface and support over 38% of the world’s human population [1]. Despite their aridity, many drylands host wetlands that maintain a positive water balance for part or all of the year [5,6]. These wetlands play a disproportionately large ecological role, supporting biodiversity, regulating hydrology, and contributing an estimated 12% to the global carbon pool [5]. However, nearly half of the world’s wetlands have been degraded or lost in the past 100 years due to the combined effects of climate change and human activities [8,9]. While climate change is increasingly recognized as a global security issue, its impact on surface water dynamics in dryland wetlands remains poorly quantified, with few studies addressing their critical interface [10,11,12]. Surface water dynamics are fundamental to regulating wetland structure, function, and ecosystem service provision [5,7,13]. Many wetlands exhibit highly variable hydrological patterns, with surface-water coverage fluctuating across seasonal and interannual timescales [12,14]. Deviations from average precipitation and temperature, common in arid regions, further intensify these dynamics, influencing surface water persistence and hydrological connectivity [10,15]. Accurate mapping of surface water extent (SWE) and its spatio–temporal variability is therefore essential for assessing wetland health, ecosystem stability, and resilience to climate variability [7,14].

Remote sensing offers a practical and cost-effective means for monitoring wetland hydrology, particularly in large or inaccessible areas with limited in situ data [16,17]. While both active and passive sensors have been successfully used, mapping small or semi-arid wetlands at high spatial and temporal resolution remains challenging, as no single sensor offers both [12,18]. In this context, the long-term Landsat archive, with its moderate spatial resolution (30 m) and consistent temporal coverage, provides a valuable resource for detecting, classifying, and mapping SWE with high accuracy [9,14]. Furthermore, its dense temporal record allows the identification of disturbance-driven changes and recovery trends through breakpoint analysis in long-term time-series data [9,12]. Extraction of SWE from optical imagery has often relied on single-band thresholding or spectral indices, with the Normalized Difference Water Index (NDWI) [19] and modified Normalized Difference Water (mNDWI) [20] providing complementary spectral sensitivities to better discriminate water from non-water features [17].

Building on these core indices, studies have shown that combining multiple water-related indices can further improve surface-water delineation in data-scarce regions worldwide. For example, combining the Tasselled Cap Wetness index (TCW) and NDWI enhances shoreline mapping [21]. Integration of NDWI, mNDWI, AWEI_sh (shadow/with shadow), and AWEI_nsh (no shadow/non-shadow) yielded Kappa values (κ) exceeding 0.9 for river and lake extraction in China [22]. Furthermore, the Open Water Likelihood (OWL) index, a regression-based model blending NDVI and mNDWI, enhances temporal stability in time-series water monitoring [23]. In the Nepalese Himalaya, Acharya et al. (2018) evaluated NDVI, NDWI, MNDWI, and AWEI using Landsat 8 imagery and demonstrated that elevation-based segmentation and combined indices substantially improved surface-water detection accuracy (OA = 0.96; κ = 0.89) across diverse environments, ranging from lowland ponds to high-altitude glacial lakes [24]. Similarly, Davranche et al. (2013) mapped flooding regimes in the Camargue wetlands (Rhône Delta, France) using seasonal multi-spectral SPOT-5 data and mid-infrared-based indices, achieving up to 83% classification accuracy and predicting shallow-marsh water levels with ~6 cm precision [17]. In arid and semi-arid regions, Hemati et al. (2023) demonstrated that hybrid combinations of NDWI, MNDWI, AWEI, and Normalized Difference Turbidity Index (NDTI) enhanced water extraction accuracy by reducing confusion with bare soil and saline crusts [25] in wetlands and waterbodies of North America. In Central Asia, Zan et al. (2022) used NDVI and NDWI from multi-decadal Landsat imagery in the Amu Darya River Delta, combined with wetland occurrence mapping and landscape metrics, to quantify spatial and temporal variation in wetland extent and its driving climatic and anthropogenic factors [26]. Chen et al. (2020) applied a similar Landsat-based time-series approach that combined multiple spectral indices, including NDWI, mNDWI, and the Enhanced Vegetation Index (EVI), to quantify changes in permanent and seasonal open-surface water bodies in the arid Tarim River Basin (north-western China) under intensive flow regulation and ecological water transfers [27]. Collectively, multi-index frameworks offer greater robustness, minimize misclassification, and enable consistent detection across diverse hydroclimatic and geomorphic settings [16,17,28]. Together, these indices capture a full range of water conditions, i.e., from deep lakes to shallow, turbid, or vegetation-covered wetlands, thereby providing a comprehensive basis for mapping SWE [17,28].

Central Asia is one of the driest regions on the planet, yet it hosts numerous natural wetland systems [6,26,27]. Following the desiccation of the Aral Sea caused by the 1960s diversion of the Amu Darya River [26], Lake Balkhash in south-eastern Kazakhstan, with its natural delta and wetland complex, is one of the few remaining areas in Central Asia that still receives permanent inflow [6,29]. However, extensive dams and reservoirs constructed in the upper Ile River Basin (China) in recent years have modified runoff, water-level variability, and reduced sediment delivery to downstream wetlands [29,30,31]. As a consequence, the delta has experienced substantial wetland contraction, channel desiccation, and expansion of desert-like ecosystems [6,32]. These changes have disrupted the downstream supply of nutrients essential for maintaining the ecological balance of the river–delta–lake continuum [7,13]. With ongoing climate change and glacier retreat projected to further reduce inflows, competition for water between ecosystems and human uses is expected to intensify [7,30,33]. Crucially, the Ile–Balkhash system remains poorly investigated, despite its ecological significance: long-term, seasonally explicit assessments of surface-water dynamics remain scarce [13,33].

This study investigates SWE variability across the Ile River Delta (IRD) at seasonal resolution for the period 2000–2024, a period for which Landsat satellite data are available, and includes the period of substantial human land-use changes in the catchment. This region is remote and lacks systematic direct observational data; therefore, a remotely sensed approach is required. We recognize a number of potential limitations in such an investigation. We note that the 30 m Landsat imagery may underestimate very narrow channels, small ponds, and water bodies that are obscured by vegetation or residual cloud cover. Our focus on the period 2000–2024 is constrained to these years because the pre-2000 archive for the region is sparse and spatially incomplete. The time series, therefore, characterizes the hydrologically regulated phase of the Ile–Balkhash system rather than pre-dam conditions. The attribution of SWE changes to hydroclimatic drivers is limited by the availability and spatial distribution of discharge, water–level, and meteorological stations, and therefore, some local variability in forcing cannot be resolved. Finally, we quantify changes in SWE and its transitional patterns but do not explicitly address groundwater interactions or water quality.

The IRD and its wetland complexes represent a valuable natural laboratory for understanding the hydrological and ecological functioning of dryland wetlands [6,7]. To better understand the critical role of diverse wetland systems and their dynamics in sustaining resilient ecosystems, it is necessary to generate spatially explicit, temporally resolved, and statistically validated SWE maps that provide a comprehensive assessment of wetland hydrology. Accordingly, this study aims to (a) map spatially and temporally explicit time series of SWE across the IRD from 2000 to 2024 using the Landsat archive and a multi-index framework, with dynamic Otsu thresholding; (b) conduct statistically robust accuracy and sensitivity analyses to identify the best-performing spectral indices for ensemble agreement; (c) assess the role of hydrological regulation and climatic variability on seasonal SWE dynamics using modified Mann–Kendall (MMK) and Spearman’s correlation (r), and (d) classify long-term wetland degradation and transition patterns using a hybrid ensemble–occurrence framework.

2. Materials and Methods

2.1. Study Area

The IRD (Figure 1a) represents one of the world’s largest preserved endorheic delta and wetland complexes found in a dryland setting [6,34]. It forms a transboundary river system extending across western China and the south-eastern part of the Republic of Kazakhstan (Figure 1b). Originating from three glacial-fed rivers, the Tekes, Kunes, and Kash in the Tianshan mountains of China (Figure 1c), the tributaries converge to form the main Ile River, which serves as the primary inflow to Lake Balkhash [6,35]. The IRD has a total area of c. 8000 km² and a broad triangular region comprising a number of highly diverse habitats [6,32].

The IRD experiences an arid continental climate, characterized by cold winters, hot summers, low humidity, and large seasonal and diurnal temperature fluctuations. Annual precipitation averages 137.6 to 141.6 mm. The highest monthly precipitation occurs in spring (April–May) and the lowest in late winter (February). Mean annual temperature varies from 6 °C to 9 °C, with July maxima of 24 °C to 26 °C and January minima of −8 °C to −14 °C [31,36]. Snow cover typically persists between mid-November and early December, but it does not occur every year. Near Araltobe village (N75°24′49.253″ E45°6′55.189″), the Ile River divides into three primary branches: the Zhideli to the east, Ile in the centre, and Topar to the west. The Zhideli branch now conveys 80% of the total discharge and features a complex hydrography with numerous channels, lakes, and spills [31,37]. The central Ile River branch, once the primary watercourse, now accounts for only 4% of runoff due to sediment accumulation and progressive infilling of its channel [31,37]. To the west, the Topar river system has bifurcated into two channels, Topar I and Topar II. Currently, the Suminka stream feeds the system only when the Ile River is in a high flow regime [31,37]. The complex network of streams on the large, extensive delta acts as a natural regulator, buffering hydrological extremes and maintaining ecological stability in the Ile Balkhash basin [37,38].

The IRD and southern Lake Balkhash have been designated as a Ramsar Site of International Importance (Figure 1a), defined within an area of 976,630 hectares (ha). The Ramsar site lies within the borders of the three State Nature Sanctuaries of Republican Importance—Balkhash, Karroy, and Kukan—which together cover an area of 1,061,100 ha [34]. The IRD hosts diverse ecosystems, including 427 plant species, 345 terrestrial vertebrates (282 birds, 39 mammals, 19 reptiles, and 3 amphibians), and 25 fish species [34].

2.2. Datasets

2.2.1. Satellite Imagery

Landsat imagery was obtained from the Google Earth Engine (GEE) platform (https://earthengine.google.com/), which provides long temporal coverage, 30 m spatial resolution, and spectral bands sensitive to water features. We accessed all Landsat Collection 2 Level 2 (surface reflectance) scenes from Landsat-5 TM, Landsat-7 ETM+ (1999 onward), and Landsat-8/9 OLI/TIRS over the IRD (Figure S1; Table S1). The final analysis focuses on the 2000–2024 period to map and evaluate the dynamics of SWE across the IRD (Figure S1; Table S1).

2.2.2. Hydrological Datasets

Discharge (D) and water-level (WL) data were obtained from seven gauging stations within the IRD (Figure 1a), maintained by the Kazakhstan Hydrometeorological Service (www.kazhydromet.kz; Table S1). For each station, we list location, variable(s), period of record, and data availability in Table S1.

2.2.3. Meteorological Datasets

Seasonal air temperature (AT) and precipitation (P) records were acquired from three meteorological stations: Kapchagay, Bakanas, and Kuigan (Figure 1a; Table S1). These datasets were sourced from Kazakhstan’s hydrometeorological organization (https://www.kazhydromet.kz/) and evaluated for their relevance to hydro-climatic conditions in the IRD.

2.3. Methods

The overall methodological workflow is illustrated in Figure 2. We first defined a design rationale for data selection and the temporal analysis window based on sensor availability, spatial coverage, and seasonal representativeness of the datasets used in the present study.

2.3.1. Rationale for Data Selection and Temporal Window

Each year was divided into three hydrological seasons: spring (March–May), summer (June–August), and fall (September–November) (Figure S1; Table S1). All Landsat Collection 2 Level 2 (surface reflectance) scenes were filtered by defined seasons and cloud cover (<85%) to ensure data quality [39]. In addition, cloud masking was applied using the “QA_PIXEL” band to exclude pixels affected by clouds and shadows [25] (Figure 2). For all sensors, pixels flagged as fill, cloud, cirrus, cloud shadow, snow/ice, or radiometric saturation in the QA_PIXEL and QA_RADSAT bands were removed prior to index computation. In Landsat-7 ETM+, scan line corrector (SLC)-off gaps are encoded as fill pixels in QA_PIXEL. These pixels are masked in the same step and do not contribute to the seasonal composites. Therefore, SLC-off stripes are treated as missing data rather than as land or water and were not gap-filled. Seasonal composites of all spectral indices were then generated by merging all valid observations from Landsat-5, Landsat-7, and Landsat-8/9 within each 3-month season using the 25th percentile value per pixel in the final workflow [40,41]. The percentile images appear to better describe water dynamics than the mean and median images used in other studies [40].

For each composite, we also recorded the number of contributing images (nimgs) per pixel. If, after QA masking, no valid observations were available for a given year season (nimgs = 0), the composite remained fully masked, and those pixels/seasons were excluded from subsequent classification and trend analysis. The spatial and temporal distribution of nimgs is summarized in Figure S1 and Table S2, allowing explicit assessment of the influence of missing data on surface-water dynamics. The availability of cloud-filtered images spanning 1992–2024 is summarized in Figure S1. Between 1992 and 2024, 5409 scenes were retained after filtering, including 1700 fall, 1700 spring, and 2009 summer images. However, the pre-2000 archive over the IRD is sparse and spatially incomplete, with several seasons lacking usable imagery and many composites failing to cover the full delta (Table S2). To ensure that each seasonal composite has complete spatial coverage and a sufficient number of observations, we restricted the long-term surface-water analysis to 2000–2024. This period coincides with denser and more consistent acquisitions, following the launch of Landsat-7 and Landsat 8/9, and provides a robust baseline for monitoring wetland dynamics (Figure S1 and Table S2).

Discharge (D) and water-level (WL) records were compiled from all available hydrometric stations within and upstream of the IRD. A preliminary assessment revealed substantial variability in record length, gaps, and inconsistencies among stations, limiting their suitability for characterizing the delta-scale hydrological regime. To overcome these limitations, the Kapchagay station, constructed in 1970 on Kapchagay Reservoir, was selected as the primary hydrological reference (Figure 1a). Its strategic location enables direct regulation of downstream flows to Ile, Zhideli, and Topar distributaries, making it the most reliable indicator of delta inflow dynamics [13,37]. Kapchagay station provides the most continuous and consistent dataset over the study period, thereby enhancing the robustness of the analysis (Table S1). Similarly, the available Precipitation (P) and air temperature (AT) records were evaluated for the Kuigan, Kapchagay, and Bakanas meteorological stations. The Kuigan station, located downstream near the Lake Balkhash shoreline, does not adequately capture the climatic conditions influencing its upstream delta (Figure 1a). By contrast, the Kapchagay and Bakanas stations, situated in the upper and middle reaches of the Ile River corridor, respectively, better represent the local climatic variability that controls the IRD’s hydrological regime. Accordingly, these two stations were emphasized when interpreting seasonal hydro-climatic drivers of surface water dynamics.

2.3.2. Index-Wise Water Mapping with Dynamic Otsu Thresholding

To enhance the separation between water and non-water pixels, seven spectral indices, i.e., mNDWI, NDWI, AWEI_nsh, AWEI_sh, TCW, Water Index 2015 (WI2015), and Multi-Band Water Index (MBWI) (Table S3) were computed for each seasonal composite between 2000 and 2024. Each index captures distinct spectral properties relevant to hydrological mapping: mNDWI is sensitive to open surface water [20], AWEI variants help reduce confusion between shadows and built-up features [42], TCW captures soil and canopy moisture [43], WI2015 was designed to detect shallow and seasonally inundated water [44], and MBWI integrates multiple spectral bands to enhance detection of small or fragmented patches [28]. The combination of these indices provides complementary spectral sensitivity and has demonstrated high accuracy in detecting water features across arid and deltaic environments similar to the IRD [16,28].

We applied dynamic Otsu thresholding to each seasonal composite within the study area to classify water and non-water pixels for all spectral indices (mNDWI, NDWI, ${\mathrm{A}\mathrm{W}\mathrm{E}\mathrm{I}}_{\mathrm{s}\mathrm{h}}$, ${\mathrm{A}\mathrm{W}\mathrm{E}\mathrm{I}}_{\mathrm{n}\mathrm{s}\mathrm{h}}$, WI2015, TCW, MBWI) [45,46]. This method calculates the optimal threshold by minimizing intra-class variance on the histogram of index values, after cloud/shadow/snow masking and fulfilling ‘No Data’. For each image–index pair, we obtained a season-adaptive threshold (T) and generated a binary mask using index-specific polarity as documented in Table S4. For positively oriented indices (e.g., mNDWI, NDWI, WI2015, MBWI, TCW), pixels with index $\ge \mathrm{T}$ were labelled water = 1, otherwise non-water = 0; for inversely oriented indices (AWEI variants, where applicable), the inequality was reversed, i.e., index ≤ $\mathrm{T}$. Pixels without valid observations at classification time were flagged as ‘No Data’ to distinguish them from class 0. The resulting per-index binary masks were further assessed for water classification accuracy based on actual conditions for each sample period, and the most suitable indices for mapping SWEs were identified.

2.3.3. Accuracy Assessment of Selected Spectral Water Indices

To validate the effectiveness of the generated binary index maps and identify the most suitable indices for mapping SWE, we conducted a point-based accuracy assessment of the seven indices across three representative seasons and years, i.e., spring-2023, summer-2019, and fall-2017, within the selected Central IRD region (Figure 1a). This validation region encompasses the main distributary channels, floodplain wetlands, lakes, and adjacent dryland surfaces, and is therefore representative of the dominant land–water configurations present across the delta. For each season, a year was randomly selected to reduce subjective bias in data selection [17].

Cloud-free Sentinel-2 Level-2A images (COPERNICUS/S2_SR_HARMONIZED; 10 m resolution) for spring-2023, summer-2019, and fall-2017 were visually interpreted to delineate reference water/non-water masks over the validation area. After applying the QA60-based cloud and cirrus mask, the corresponding Sentinel-2 collections contained 141(spring-2023), 237 (summer-2019), and 120 (fall-2017) scenes over the study area, providing dense seasonal coverage for reference mapping and validation. To mitigate class imbalance, particularly the dominance of non-water pixels in heterogeneous dryland wetlands, validation datasets were balanced by selecting equal numbers of water (=1) and non-water (=0) samples. These validation samples were drawn using a stratified random design from the Sentinel-2 reference masks, resulting in 299 reference samples per season and totaling 897 samples. This balanced approach ensures that accuracy metrics were not inflated by majority-class dominance, as the IRD region is mostly dry, thus providing a robust assessment of index performance [28,47]. All seven spectral indices (Table S3) were assessed for accuracy separately for each of the three seasons, enabling index performance to be evaluated in a seasonally explicit framework. Evaluation metrics such as Precision (P), Recall (R), Overall Accuracy (OA), F1-score (F1), and Cohen’s Kappa (κ) were computed from binary index maps (see Supplementary Section S1.1 for details). Season specific index performance was evaluated using OA, κ, and F1-score. To ensure robustness of the SWE ensemble mapping, only indices that consistently achieved high accuracy (F1 and OA ≥ 0.85 and 85%) in at least two of the three seasons were retained for application across seasonal spectral variability [18,47,48]. Sentinel-2 data were used exclusively for accuracy assessment and not for index computation or threshold tuning, ensuring independence between validation and classification.

2.3.4. Ensemble Agreement Logic for Present-Day SWE Mapping

Building on the complementary strengths of best-performing indices, we applied an ensemble agreement approach to generate spatio-temporal SWE. This integration reduces the dependency on any single index and minimizes index-specific misclassification [28]. We implemented three complementary ensemble logics (Figure 2).

• Intersection (AND) rule

A conservative method that retains only those pixels consistently classified as water across all selected indices. This criterion isolates stable water pixels and minimizes commission errors.

$${\mathrm{W}}_{\mathrm{i}\mathrm{n}\mathrm{t}}\left(\mathrm{x},\mathrm{y}\right)=\left\{\begin{array}{c}1, \mathrm{i}\mathrm{f} {\sum }_{\mathrm{t}=1}^5{\mathrm{W}}_{\mathrm{x},\mathrm{y}}^{\mathrm{t}}=5\\ 0, \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\hfill \end{array}\right.$$

(1)

2.

Majority voting (t-of-5, t = 3) rule

A balanced method in which a pixel is classified as water if it is detected by the majority (≥3 out of 5 indices in our case). This logic captures periodic inundation dynamics while reducing the influence of any single index, particularly useful in floodplain wetlands and delta regions [28].

$${\mathrm{W}}_{\mathrm{m}\mathrm{a}\mathrm{j}}\left(\mathrm{x},\mathrm{y}\right)=\left\{\begin{array}{c}1, \mathrm{i}\mathrm{f} {\sum }_{\mathrm{t}=1}^5{\mathrm{W}}_{\mathrm{x},\mathrm{y}}^{\mathrm{t}}>3\\ 0, \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\hfill \end{array}\right.$$

(2)

3.

Union (OR) rule

An inclusive method that classifies a pixel as water if at least one index detects it. This maximizes sensitivity to ephemeral water features but may also introduce false positives.

$${\mathrm{W}}_{\mathrm{u}\mathrm{n}\mathrm{i}}\left(\mathrm{x},\mathrm{y}\right)=\left\{\begin{array}{c}1, \mathrm{i}\mathrm{f} {\sum }_{\mathrm{t}=1}^5{\mathrm{W}}_{\mathrm{x},\mathrm{y}}^{\mathrm{t}}\ge 1\\ 0, \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\hfill \end{array}\right.$$

(3)

Here, ${\mathrm{W}}_{\mathrm{x},\mathrm{y},}^{\mathrm{t}}$ are the binary masks of identical dimensions indexed (water or non-water) from the t^th index mask at the given pixel.

From these three ensembles, we mapped three primary SWE regimes in the IRD region for each year from 2000–2024 as follows:

• Stable SWE (${\mathrm{W}}_{\mathrm{s}}={\mathrm{W}}_{\mathrm{i}\mathrm{n}\mathrm{t}}$): pixels that belong to the intersection ensemble (W_int), i.e., classified as water by all selected indices in a given season. These represent the most persistent and spectrally robust surface water areas.
• Periodic SWE (${\mathrm{W}}_{\mathrm{P}}={\mathrm{W}}_{\mathrm{m}\mathrm{a}\mathrm{j}}-{\mathrm{W}}_{\mathrm{i}\mathrm{n}\mathrm{t}}$): pixels that occur in the majority ensemble (W_maj) but not in the intersection (W_int). They are mapped as water by the majority of indices, indicating regularly inundated zones with higher spectral variability (e.g., partly vegetated or shallow water).
• Ephemeral SWE (${\mathrm{W}}_{\mathrm{E}}={\mathrm{W}}_{\mathrm{u}\mathrm{n}\mathrm{i}}-{\mathrm{W}}_{\mathrm{m}\mathrm{a}\mathrm{j}}$): pixels that occur only in the union ensemble (W_uni) and not in the majority (W_maj), i.e., detected as water by at least one index but not by most. These correspond to occasional or short-lived inundation, or spectrally ambiguous fringe waters.

This hybrid ensemble logic provides a robust and transparent framework for mapping three primary surface water extents (SWE), i.e., stable, periodic, and ephemeral. The reliability reflects the validated accuracy of the underlying seasonal water/non-water classifications, rather than separate class-specific confusion matrices for the derived temporal classes. For each season, total SWE (km²) was computed annually and summarized as mean, maximum, and minimum values to capture interannual variability (Table S5).

2.3.5. SWE Seasonal Trends and Correlation with Hydroclimatic Parameters

We assessed monotonic trends in core SWE and associated hydroclimatic variables using the MMK non-parametric test, which is robust to non-normal distributions, outliers, and missing values [49]. To account for serial autocorrelation, which is common in hydrological time series, we adopted a trend-free pre-whitening (TFPW) approach [50,51]. The procedure involves (i) estimating an initial Sen’s slope (β); (ii) detrending the series $(x_t-\mathrm{\beta }t)$; (iii) estimating lag-1 autocorrelation (ρ1) of the detrended residuals; (iv) pre-whitening residuals via AR(1) adjustment if, ∣ρ1∣ ≥ 0.05, and (v) reintroducing the estimated trend, yielding a TFPW series on which MK was reapplied. In parallel, we used the variance-corrected MK formulation with tie correction [49] to stabilize ρ-values in the presence of autocorrelation (see Supplementary Section S1.2 for details). The magnitude of each trend was quantified using Sen’s slope (β), representing the annual rate of change. Negative slopes indicate declining water extent (drying), while positive slopes reflect increasing extent (wetting). We computed Spearman’s (r) to explore the relationship between seasonal SWEs (stable, periodic, and ephemeral) and hydrological drivers such as D, WL, AT, and P. Spearman’s (r) was defined as the Pearson correlation applied to rank-transformed values of two variables, with corresponding two-sided ρ-values [52,53]. Only pairs with at least two overlapping years of valid data were included, summarized by significance level (ρ < 0.05, 0.01, 0.001) [54] (see Supplementary Section S1.2). To account for potential delays in the hydrological response of the wetlands, we repeated the correlation analysis using lagged hydroclimatic predictors. For each seasonal SWE time series, we computed interannual lag by correlating SWE for a given season with the same season’s hydrological drivers shifted by 1–3 years (see Supplementary Section S1.3).

2.3.6. Hybrid Ensemble–Occurrence Analysis of SWE Transitions

We analyzed wetland degradation and transition dynamics using a hybrid ensemble-occurrence approach that integrates our ensemble agreement logic with long-term water occurrence analysis. The presence of different inundation types within clustered pixels characterizes wetland areas. In dryland settings, most streams are ephemeral in nature, which makes them difficult to distinguish [2,55]. However, under regulated flow conditions, these ephemeral channels can sometimes maintain hydrological connectivity between upstream and downstream areas, and even support wetlands such as isolated pools that appear within channels through rational water allocation [55,56,57].

To capture these dynamics, we combined the primary SWE classes obtained from three-core ensemble agreement logic of Intersection (I, Equation (1)), Majority (M, Equation (2)), and Union (U, Equation (3)) with a water occurrence framework to map long-term SWE changes. Each ensemble cube stored binary information (0/1) for every seasonal timestep (spring, summer, fall) from 2000–2024. Pixels were considered valid only where at least 50% of timesteps contained cloud-free observations. This minimum fraction of valid observations requires a pixel to be kept in the analysis (MIN_VALID_OBS_FRAC = 0.50), which represents a compromise between retaining spatial coverage and avoiding poorly sampled pixels with uncertain temporal behaviour [10,40,58]. For each valid pixel, we computed (i) the overall occurrence of each ensemble SWE class, defined as the fraction of valid timesteps classified as water in the union ensemble water occurrence (O_ccU), majority ensemble water occurrence (O_ccM) and Intersection ensemble water occurrence (O_ccI), (ii) the frequency of periodic and ephemeral SWE class occurrence (where ≥10% of timesteps were classified as water) from the ensemble core, such that isolated, one-off detections are not misclassified as hydrologically meaningful features [40]; (iii) epochal changes by comparing early (2000–2008) and late (2016–2024) periods for the remaining pixels in the O_ccU to detect step-change dynamics. Using two balanced 9-year windows as seasonal endmembers helps to normalize short-term variability (such as extreme events) while maintaining sufficient granularity to detect significant long-term hydrological changes [58,59,60]. These changes include (a) the flowing phase, when water is maintained due to rainfall, snowfall or from baseflow contribution; (b) the non-flowing phase when interrupted flow creates connected or isolated pools, and (c) the dry phase, when surface water is absent, but significant hydrological processes occur in the hyporheic zone [61,62]; and (iv) long-term trends using linear slopes of water presence (0/1) against normalized time, where slope ± 0.10 or ±10% was used to flag monotonic drying or wetting in a dryland settings, where small fractions fail to capture weak surface water changes [40,58]. Based on these metrics, we classified each pixel into one of seven mutually exclusive categories: permanent, seasonal, temporary, lost, new, diminishing, or intensifying. Classification thresholds were derived from occurrence fractions and ensemble overlap brackets, as shown in

Table 1. Description of degradation and transition patterns based on hybrid ensemble occurrence analysis of the SWE between 2000 and 2024.

Code	Surface Water Patterns	Definition	Remarks
0	No Water	Outside AOI or <50% valid timesteps, or fails all criteria below	Unclassified as water over the analysis window
1	Permanent	OccI ≥ 75%	Represents the most hydrologically stable regions of the wetland, providing reliable water sources and habitat even during dry spells
2	Seasonal	Not Permanent and (OccM − OccI) ≥ 10%	Reflects predictable intra-annual variability such as floodplain inundation or seasonally filled lakes.
3	Temporary	Neither 1 nor 2 and (OccU − OccM) > 10%	Captures irregular, short-lived water bodies such as ephemeral streams, ponds, or depressions triggered by rainfall or short hydrological events.
4	Lost	OccU, early ≥ 50% and OccU, late ≤ 12.5%	Indicates a shift toward drier regimes and contraction of long-term water bodies.
5	New	OccU, early ≤ 12.5% and OccU, late ≥ 50%	Represents expansion or reactivation of wetlands, reflecting wetter local hydrological conditions.
6	Diminishing	Remaining pixels over the entire time series with trend slope ≤ −0.10 (OccU)	Indicates progressive wetland degradation and contraction of water bodies, often linked to reduced inflows, climate stress, or anthropogenic water extraction
7	Intensifying	Remaining pixels over the entire time series with trend slope ≥ +0.10 (OccU)	Reflects gradual recovery or expansion of water presence, potentially due to improved inflows, localized restoration, or wetter climatic phases.

Observation period: Pre-period (2000–2008) and post-period (2016–2024); O_ccU = Union ensemble water occurrence; O_ccM = Majority ensemble water occurrence; O_ccI = Intersection ensemble water occurrence.

. This hybrid ensemble–occurrence framework integrates frequency-based analysis with ensemble agreement rules, ensuring robust detection of both stable and transitional surface water patterns across space and time.

3. Results

3.1. Accuracy Assessment of Seasonal Indices

Table 2. Seasonal accuracy assessment of seven water indices (AWEI_nsh, AWEI_sh, MBWI, NDWI, TCW, WI2015, and mNDWI) against balanced binary reference samples mapped from Sentinel-2 imagery in the Central Ile region. Reported metrics include True Positives (TP), True Negatives (TN), False Positives (FP), False Negatives (FN), Precision, Recall (Sensitivity), F1-score, and Overall Accuracy (OA) and Cohen’s Kappa (κ), based on N = 299 validation points per season (spring-2023, summer-2019, and fall-2017).

Index	TP	TN	FP	FN	Precision	Recall	F1-Score	OA	κ	N
Spring (2023)
AWEInsh	148	123	12	16	0.925	0.902	0.914	0.906	0.811	299
AWEIsh	141	128	7	23	0.953	0.860	0.904	0.900	0.800	299
MBWI	124	135	0	40	1.000	0.756	0.861	0.866	0.737	299
NDWI	128	132	3	36	0.977	0.780	0.868	0.870	0.742	299
TCW	151	121	14	13	0.915	0.921	0.918	0.910	0.818	299
WI2015	149	125	10	15	0.937	0.909	0.923	0.916	0.832	299
mNDWI	128	132	3	36	0.977	0.780	0.868	0.870	0.742	299
Summer (2019)
AWEInsh	163	88	40	8	0.803	0.953	0.872	0.839	0.662	299
AWEIsh	137	126	2	34	0.986	0.801	0.884	0.880	0.762	299
MBWI	125	128	0	46	1.000	0.731	0.845	0.846	0.699	299
NDWI	134	127	1	37	0.993	0.784	0.876	0.873	0.749	299
TCW	168	58	70	3	0.706	0.982	0.822	0.756	0.466	299
WI2015	144	125	3	27	0.980	0.842	0.906	0.900	0.800	299
mNDWI	134	127	1	37	0.993	0.784	0.876	0.873	0.749	299
Fall (2017)
AWEInsh	145	106	43	5	0.771	0.967	0.858	0.839	0.679	299
AWEIsh	138	143	6	12	0.958	0.920	0.939	0.940	0.880	299
MBWI	126	147	2	24	0.984	0.840	0.906	0.913	0.826	299
NDWI	135	146	3	15	0.978	0.900	0.938	0.940	0.880	299
TCW	149	50	99	1	0.601	0.993	0.749	0.666	0.330	299
WI2015	142	138	11	8	0.928	0.947	0.937	0.936	0.873	299
mNDWI	135	146	3	15	0.978	0.900	0.938	0.940	0.880	299

presents user, producer, and overall accuracies, derived from 299 validation points collected across three representative seasons.

In spring 2023, all indices performed strongly. WI2015, TCW, AWEI_nsh, and AWEI_sh form the top group (Table 2). WI2015 achieved the highest score (F1 = 0.92, OA = 0.92) with κ ≈ 0.83 (almost perfect agreement), closely followed by TCW (F1 ≈ 0.92, OA ≈ 0.91; κ ≈ 0.82), AWEI_nsh (F1 ≈ 0.91, OA ≈ 0.91; κ ≈ 0.81), and AWEI_sh (F1 ≈ 0.90, OA ≈ 0.90; κ ≈ 0.80). NDWI and mNDWI also showed high precision but with somewhat lower recall (F1 ≈ 0.87, OA ≈ 0.87; κ ≈ 0.74), similar to MBWI (F1 ≈ 0.86, OA ≈ 0.87; κ ≈ 0.74), which achieved perfect precision but missed more water pixels. Overall, κ values ranged from ≈0.74 to 0.83, indicating substantial to almost perfect agreement across all indices in spring (Table 2, Figure 3a).

In summer 2019, WI2015 again achieved the highest overall performance (F1 ≈ 0.91, OA ≈ 0.90; κ ≈ 0.80, substantial), followed closely by AWEI_sh and NDWI/mNDWI (F1 ≈ 0.88, OA ≈ 0.87–0.88; κ ≈ 0.75–0.76, substantial) (Table 2; Figure S2a). AWEI_nsh remained strongly recall-dominant (recall ≈ 0.95) but with slightly lower overall accuracy (F1 ≈ 0.87, OA ≈ 0.84; κ ≈ 0.66). MBWI again exhibited perfect precision but reduced recall (F1 ≈ 0.85, OA ≈ 0.85; κ ≈ 0.70). TCW showed the weakest summer performance (F1 ≈ 0.82, OA ≈ 0.76; κ ≈ 0.47, moderate agreement) due to numerous false positives, despite very high recall (≈0.98).

In fall 2017, AWEI_sh, NDWI/mNDWI, and WI2015 formed a clear top-performing group, all reaching F1 ≈ 0.94 and OA ≈ 0.94 with κ ≈ 0.87–0.88 (almost perfect agreement; Table 2; Figure S2b). MBWI also performed strongly (F1 ≈ 0.91, OA ≈ 0.91; κ ≈ 0.83). AWEI_nsh remained acceptable (F1 ≈ 0.86, OA ≈ 0.84; κ ≈ 0.68, substantial). By contrast, TCW was strongly recall-dominant (recall ≈ 0.99) but highly imprecise (precision ≈ 0.60), yielding the weakest fall agreement (F1 ≈ 0.75, OA ≈ 0.67; κ ≈ 0.33, fair).

When synthesizing across the three seasons, the seasonal accuracy metrics of WI2015, AWEI_sh, MBWI, mNDWI, NDWI, and AWEI_nsh demonstrate that the dynamic Otsu thresholding approach is effective for water delineation across contrasting radiometric and hydrological conditions. For all indices, F1-scores and overall accuracies are generally ≥0.85 in at least two of the three seasons, and κ values mostly fall between ≈0.66 and 0.88, indicating substantial to almost-perfect agreement with the reference data (Table 2; Figure 3b). This consistently high performance across spring (high flow), summer (peak vegetation and turbidity), and fall (declining water, exposed floodplains) shows that the Otsu-derived thresholds successfully adapt to seasonal changes in background reflectance and avoid the subjectivity of manually fixed thresholds. TCW, in contrast, is strongly recall-dominant but suffers from excessive false positives in summer and fall, with κ dropping to moderate or fair agreement (≈0.33–0.47) (Table 2).

Taken together, WI2015, AWEI_sh, MBWI, mNDWI, NDWI, and AWEI_nsh emerged as the most robust indices. Each consistently achieved F1 ≥ 0.85 and OA ≥ 0.85 in at least two seasons of three seasons, with κ typically ≥0.70, demonstrating robustness and complementary strengths in detecting clear open water, shallow and turbid water, and small or fragmented patches. Between NDWI and mNDWI, both yielded nearly identical accuracy metrics (Table 2 and Figure 3b); however, mNDWI was retained to avoid redundancy and because its Green–SWIR formulation is better suited to suppressing soil, shadow, and built-up effects typical of dryland deltas. We therefore retain WI2015, AWEI_sh, MBWI, mNDWI, and AWEI_nsh as the core index set for the ensemble analysis, ensuring the ensemble benefits from high individual accuracy, seasonal robustness of the Otsu thresholds, and diverse sensitivity to the full spectrum of water conditions (Figure 3b and Table 2).

3.2. Temporal Dynamics and Long-Term Trends of Seasonal SWEs (2000–2024)

The combined analysis of areal statistics and the MMK test with Sen’s slope estimation revealed clear but seasonally asymmetric changes in SWE across the IRD between 2000 and 2024 (

Table 3. Modified Mann–Kendall (MMK) test showing trend (Z), significance (ρ), and Sen’s slope (β) estimate with 95% confidence intervals for stable, periodic, and ephemeral SWE across spring, summer, and fall (2000–2024).

Component	Z	ρ	Sen’s Slope (β)	Status
Spring
Stable	−3.001	0.003	−25.175	Pre-whitened (rho = 0.523)
Periodic	−0.471	0.637	14.421	Pre-whitened (rho = 0.602)
Ephemeral	−0.620	0.535	−3.402	Pre-whitened (rho = 0.230)
Summer
Stable	−3.200	0.001	−9.451	Pre-whitened (rho = 0.640)
Periodic	−3.246	0.001	−13.803	Pre-whitened (rho = 0.342)
Ephemeral	−3.433	0.001	−29.651	Pre-whitened (rho = 0.356)
Fall
Stable	−3.060	0.002	−7.387	Pre-whitened (rho = 0.601)
Periodic	−3.340	0.001	−22.068	Pre-whitened (rho = 0.085)
Ephemeral	−1.051	0.293	−2.408	Pre-whitened (rho = 0.429)

Note: Units of Sen’s slope (β) to surface water extent (SWE) area = km² yr⁻¹.

; Table S4, and Figure 4).

In spring, stable SWE exhibited a significant long-term decline (Z = −3.00, ρ = 0.003), with Sen’s slope of −25.2 km² yr⁻¹. Their extent peaked in 2004 (2529.6 km²) before contracting steadily to <1000 km² by 2024, highlighting the persistent loss of stable SWE in early growing seasons. Periodic waters showed no significant trend (Z = −0.47, ρ = 0.637, β = +14.4 km² yr⁻¹) but fluctuated strongly, with peak values ranging from 2408 km² in 2013 to 1122 km² in 2024, indicating interannual variability without directional change. Ephemeral SWE also declined slightly (β = −3.4 km² yr⁻¹), though not significantly (ρ = 0.535), varying between ~925 km² (2001) to ~289 km² (2014). Together, these results show that spring is historically the wettest season and experienced a long-term decline of stable SWE despite temporary expansions of periodic SWE (Table 3; Table S4 and Figure 4a).

In summer, all three SWE components declined significantly, making summer the most vulnerable season. Stable SWE significantly decreased at β = −9.5 km² yr⁻¹ (Z = −3.20, ρ = 0.001), falling from ~1582 km² in 2004 to <900 km² in 2024. Periodic waters showed an even steeper contraction (Z = −3.25, ρ = 0.001, S = −13.8 km² yr⁻¹), shrinking from >1100 km² in 2010 to just 458 km² in 2024. Ephemeral SWE declined most sharply (Z = −3.43, ρ = 0.001, β = −29.7 km² yr⁻¹), collapsing from a maximum of 3655 km² in 2017 to 1783 km² in 2024. The synchronized reductions across all classes, coupled with sharp peaks in 2006, 2010, and 2017, indicate that while ephemeral SWE expanded under occasional high inflows, these gains have not been sustained, leaving summer water bodies highly sensitive to flow reductions and evaporative stress (Table 3; Table S4 and Figure 4b).

In the fall, the declining signal persisted. Stable SWE continued to decline significantly (Z = −3.06, ρ = 0.002, β = −7.4 km² yr⁻¹). The SWE decreases from ~1491 km² in 2004 to ~839 km² in 2024, reflecting a steady contraction of stable SWE inundation late in the season. Periodic waters showed the strongest fall trend among all classes (Z = −3.34, ρ = 0.001, β = −22.1 km² yr⁻¹), declining from ~1741 km² in 2016 to ~515 km² in 2024, highlighting their role as the most vulnerable fall component. Ephemeral SWE, by contrast, remained highly variable without a significant trend (ρ = 0.293), oscillating between a high of ~1771 km² in 2017 and a low of ~623 km² in 2023. This seasonal behaviour suggests that fall water extent depends on residual inflows and short-lived pulses, but the long-term trajectory is dominated by the sharp loss of periodic SWE (Table 3; Table S4 and Figure 4c).

To further disentangle the hydrological and climatic controls underlying these declines, additional trend analyses were performed on D, WL, AT, and P records from the Kapchagay and Bakanas stations. The detailed results presented in Supplementary Section S2.1 and Table S6 provide additional context on how flow regulation and climate variability jointly influence the seasonal dynamics of surface water extent across the IRD.

3.3. Correlation Between Seasonal SWE and Hydroclimatic Drivers

The Spearman’s correlation (r) analysis provides further insights into the relationship between hydroclimatic variables and different SWE components (Figure 5;

Table 4. Spearman’s rank correlation coefficient (r) and statistically significant (ρ) between SWE (stable, periodic, ephemeral) and hydrological parameters, i.e., discharge (D), water level (WL), air temperature (AT), precipitation (P) at the IRD region for (a) spring, (b) summer, (c) fall between 2000–2024.

No of Obs.	SWE Cores	Hydro Variables	Spring		Summer		Fall
			Spearman’s (r)	ρ Value	Spearman’s (r)	ρ Value	Spearman’s (r)	ρ Value
25	Stable	D (Kapchagay)	0.472	0.020	−0.012	0.955	0.317	0.123
25		WL (Kapchagay)	0.487	0.016	−0.010	0.961	0.004	0.984
25		AT (Bakanas)	−0.125	0.560	−0.101	0.630	−0.014	0.946
25		AT (Kapchagay)	−0.340	0.104	−0.196	0.348	−0.048	0.818
25		P (Bakanas)	0.247	0.245	−0.122	0.561	−0.002	0.991
25		P (Kapchagay)	0.328	0.117	−0.185	0.375	−0.329	0.108
25	Periodic	D (Kapchagay)	0.131	0.541	0.577	0.003	0.639	0.001
25		WL (Kapchagay)	0.132	0.538	0.562	0.003	0.545	0.005
25		AT (Bakanas)	0.018	0.932	−0.468	0.018	−0.020	0.926
25		AT (Kapchagay)	−0.150	0.485	−0.640	0.001	−0.048	0.821
25		P (Bakanas)	0.126	0.557	0.095	0.651	0.144	0.493
25		P (Kapchagay)	0.320	0.128	0.406	0.044	−0.017	0.935
25	Ephemeral	D (Kapchagay)	0.161	0.451	0.312	0.129	0.430	0.032
25		WL (Kapchagay)	0.304	0.149	0.325	0.113	0.027	0.900
25		AT (Bakanas)	0.308	0.143	−0.317	0.122	0.050	0.811
25		AT (Kapchagay)	0.313	0.136	−0.328	0.109	0.036	0.864
25		P (Bakanas)	0.411	0.046	−0.133	0.527	0.235	0.259
25		P (Kapchagay)	−0.075	0.727	0.169	0.421	−0.130	0.537

).

In spring, stable SWE showed significant positive correlations with both D (r = 0.47, ρ = 0.020) and WL (r = 0.49, ρ = 0.016) at Kapchagay station, suggesting that stable SWE is strongly maintained by upstream releases during the early growing season. By contrast, no significant associations were found with AT at Bakanas (r = −0.13, ρ = 0.560) or Kapchagay (r = −0.34, ρ = 0.104), nor with P at Bakanas (r = 0.25, ρ = 0.245) or Kapchagay (r = 0.33, ρ = 0.117), reflecting the dominance of hydrological regulation over climatic controls in this period. For periodic waters, all correlations were weak and non-significant, including D (r = 0.13, ρ = 0.541) and WL (r = 0.13, ρ = 0.538). Ephemeral SWE also showed no significant links with D (r = 0.16, ρ = 0.451) or WL (r = 0.30, ρ = 0.149), although a marginally positive correlation was observed with P at Bakanas (r = 0.41, ρ = 0.046). This highlights the limited role of short-term climatic variability, which exerts only a limited influence on SWE during early spring (Figure 5a; Table 4).

In summer, stable SWE exhibited no significant relationships with hydrological or climatic variables, with very low correlations for D (r = −0.01, ρ = 0.955) and WL (r = −0.01, ρ = 0.961). By contrast, periodically flooded water areas showed strong positive correlations with both D (r = 0.58, ρ = 0.003) and WL (r = 0.56, ρ = 0.003) at Kapchagay, consistent with the expansion of seasonal inundation under peak flow conditions. At the same time, negative correlations were observed with AT at Bakanas (r = −0.47, ρ = 0.018) and Kapchagay (r = −0.64, ρ = 0.001), indicating that higher summer temperatures reduce surface water extent, likely through increased evapotranspiration and reduced upstream inflows. Precipitation showed weak to moderate positive associations with periodic and ephemeral SWE, but only P at Kapchagay reached significance (r = 0.41, ρ = 0.044), suggesting a limited direct role of rainfall in sustaining SWE compared to regulated discharge. Ephemeral SWE showed no significant relationships with AT and P, but exhibited moderate, non-significant positive correlations with D (r = 0.31, ρ = 0.129) and WL (r = 0.33, ρ = 0.113). These results suggest that periodic water in summer is strongly controlled by inflows and reservoir levels but is negatively affected by elevated temperatures, reflecting high evaporative stress (Figure 5b; Table 4).

In fall, stable SWE again showed no significant relationships with D (r = 0.32, ρ = 0.123), WL (r = 0.00, ρ = 0.984), or climate variable AT at Bakanas: r = −0.01, ρ = 0.946; Kapchagay: r = −0.05, ρ = 0.818). By contrast, periodic waters maintained strong positive correlations with both discharge D (r = 0.64, ρ = 0.001) and WL (r = 0.55, ρ = 0.005), highlighting the continued dominance of hydrological regulation in controlling late-season inundation. Ephemeral SWE also exhibited a significant positive correlation with discharge D (r = 0.43, ρ = 0.032), but no consistent associations with AT or P, suggesting that short-lived SWE are sustained by flow variability and backwater effects during the late growing season. Therefore, in fall, climatic variables remained largely non-significant across all SWE classes, further reaffirming that discharge and water-level regulation are the primary controls of seasonal SWE dynamics in IRD (Figure 5c; Table 4).

Beyond this same year relationship, we also evaluated interannual lag effects between SWE and the main hydroclimatic drivers, considering lags of −3 to +3 years (Table S7). The optimal-lag analysis shows that stable SWE is strongly controlled by multi-year D and WL conditions at Kapchagay, with peak correlations typically at 1–2-year lags in all seasons (e.g., spring D lag 1, r ≈ 0.64; summer D lag 2, r ≈ 0.73; fall D lag 2, r ≈ 0.66). This indicates that core SWE integrates hydroclimatic anomalies over several preceding years rather than responding solely to current year forcing. Periodic SWE exhibits a shorter response time, i.e., summer and fall periodic SWE are most strongly related to current-year or 1-year-lag D and WL (lag 0–1, r ≈ 0.56–0.70), whereas spring periodic SWE shows stronger links at 2–3-year lags, reflecting accumulated effects of antecedent flow and temperature. Ephemeral SWE generally shows weaker and more variable interannual relationships, with significant positive correlations with D and WL confined to lags 0–1 year, consistent with its sensitivity to recent flow pulses and backwater effects. Correlations with AT or P at negative lags are interpreted as causal rather than controlling factors and are therefore not discussed further for physical attribution (see Supplementary Section S2.2 and Table S7).

3.4. Spatial Patterns of Wetlands and Their Main Hydrological Controls

Based on the hybrid ensemble-occurrence approach, the spatio-temporal dynamics of SWE in the IRD were classified into a comprehensive spatial wetland map comprising three primary wetland classes: ‘Permanent’, ‘Seasonal’, and ‘Temporary’, as well as four transitional classes: ‘Lost’, ‘New’, ‘Diminishing’, and ‘Intensifying’ (Table 1). Each of these categories presents stable hydrological occurrence and its transitions over time (

Table 5. Inventory of degradation and transitions categories based on hybrid ensemble occurrence analysis of the SWE between 2000 and 2024.

Area of Interest (AOI)/Area (km2)	Polygon Area	Total Area	Degradation and Transitions Classes
			Permanent	Seasonal	Temporary	Lost	New	Diminishing	Intensifying
At the delta scale	27,791.37	6921.00	828.55 [11.97]	3411.45 [49.29]	1435.70 [20.74]	98.46 [1.42]	49.56 [0.72]	768.68 [11.11]	328.59 [4.75]
AOI-I	2749.57	839.32	239.85 [28.58]	379.98 [45.15]	48.33 [5.76]	9.94 [1.18]	12.97 [1.55]	19.78 [2.36]	129.46 [15.42]
AOI-II	5357.18	2925.23	250.07 [8.55]	1568.74 [53.68]	756.41 [25.86]	25.11 [0.86]	10.89 [0.37]	262.04 [8.96]	51.96 [1.78]
AOI-III	778.24	819.62	167.29 [20.41]	496.64 [60.59]	68.60 [8.37]	8.53 [1.04]	6.92 [0.84]	43.55 [5.31]	28.68 [3.43]
AOI-IV	1441.60	217.01	18.82 [2.41]	99.21 [12.72]	48.28 [6.19]	11.47 [1.47]	0.06 [0.01]	37.73 [4.84]	1.45 [0.19]

The inventory is in the format Count (AREA), where count represents the number of wetlands and areas are in square kilometres (km²). AOI-I: Lake Itishpes and wetlands (2749.57 km²); AOI-II: Central Ile River Delta (5357.18 km²); AOI-III: Wetlands near Tonap (778.24 km²); AOI-IV: Wetlands west of Saryesik Peninsula (1441.60 km²).

).

At the delta scale, a total wetland area of 6921 km² was mapped, accounting for ~25% of the study area (Figure 6, Table 5). Seasonal water coverage dominates the region (3418.8 km²; 49.4%), followed by temporary (1449.6 km²; 20.9%) and permanent (828.6 km²; 12.0%). Transitional dynamics were also notable with diminishing water coverage over 768.7 km² (11.1%), while intensifying class occupied 328.6 km² (4.8%). In contrast, new (49.6 km²; 0.7%) and lost classes (98.5 km²; 1.4%) made up smaller fractions. Overall, these results confirm that the IRD is dominated by highly dynamic seasonal and temporary water classes, with relatively few permanent waters. The spatial distribution reflects a clear ongoing hydrological change over the time frame, modulated by upstream inflows and regulated discharge (Figure 6; Table 5).

At the area of interest (AOI) level, distinct spatial contrasts emerged across the four representative zones. AOI-I (839.3 km²; 30% of polygon area) contained a balanced mix of permanent (239.9 km²; 28.6%) and seasonal wetlands (380.0 km²; 45.2%). Transitional dynamics were marked by a substantial share of intensifying class (129.5 km²; 15.4%), suggesting localized recovery, while new (13.0 km²; 1.6%) and lost (9.9 km²; 1.2%) classes were minor. This indicates a resilient hydrological regime sustained by steady, low-magnitude inflows (Figure 6(I) and Table 5). AOI-II, the largest and most water-rich zone (2925.2 km²; 55% of area), was dominated by seasonal (1568.7 km²; 53.7%) and temporary (756.4 km²; 25.9%) classes, with permanent class forming only 8.6%. Transitional features included diminishing (262.0 km²; 9.0%) and intensifying (52.0 km²; 1.8%) classes, indicating strong hydrological variability but limited resilience (Figure 6(II) and Table 5). AOI-III (819.6 km²; 62% of area) also showed strong seasonal dominance (497.2 km²; 16.7%), supported by permanent (167.3 km²; 20.4%) and temporary (68.6 km²; 8.4%) wetlands. Transitional classes were also notable with diminishing (43.6 km²; 5.3%) and intensifying (28.7 km²; 3.4%), indicating localized resilience amid widespread seasonal variability (Figure 6(III) and Table 5). AOI-IV recorded the lowest wetland coverage (217.0 km²; 19% of area), dominated by seasonal water class (99.2 km²; 12.7%) and temporary (48.3 km²; 6.2%), while permanent wetlands accounted for only 2.4% (18.8 km²). Transitional dynamics were proportionally important, with 37.7 km² diminishing (4.8%) and lost (11.47 km²; 1.47%) class, reflecting a stressed hydrological regime with limited recovery capacity (Figure 6(IV) and Table 5).

Given that D and WL exert primary control on the dynamics of wetlands across the IRD, hydrological observations from Kazhydromet gauging stations were examined to explain these spatial variations (Figure 6; Table S8). At the Kapchagay apex (Site 1), discharges peaked in summer (~675 m³/s) compared to spring/fall (~430 m³/s), accompanied by a ~0.4 m seasonal rise in water level. This upstream pulse sustains delta-wide connectivity but is unevenly distributed across distributaries. In AOI-II, fed by the Zhideli branch (Site 6), summer discharges exceed 650 m³/s before dropping to ~440 m³/s in fall, driving large seasonal expansions and contractions and explaining the dominance of seasonal and temporary water patterns alongside high loss rates. AOI-III, although not directly gauged, reflects the downstream propagation of these pulses and backwater effects, where summer levels rise (~2.0 m), but fall drawdowns drive seasonal dominance and loss. In contrast, AOI-I (Site 8) experiences much smaller but steady flows (6.9–8.6 m³/s), supporting a balanced mix of permanent and seasonal waters with localized intensification. AOI-IV near the Suminka branch (Site 4) experiences very low discharges (6–14 m³/s) despite relatively high water levels (~3.1–3.6 m) and sustains only a fragmented seasonal and temporary class, with little evidence of recovery (Figure 6; Table S8).

Taken together, these patterns demonstrate that AOI-II, fed by the main distributary, sustains the largest wetland areas but remains highly dynamic, with large seasonal expansions and contractions tied to summer inflows and fall drawdowns. By contrast, AOI-I and AOI-III, supplied by smaller distributaries, maintain smaller yet more stable wetland mosaics, with AOI-I showing persistence under low yet steady flows, while AOI-III indicates a stressed system where high WL cannot compensate for low D. At the basin scale, this confirms that regulation at Kapchagay reservoir and the allocation of discharge among distributaries are the dominant controls of SWE and variability across the IRD region (Figure 6; Table S8).

4. Discussion

4.1. Progress and Challenges in Spatio-Temporal Monitoring

Accurate spatio-temporal mapping of surface-water dynamics is essential for understanding hydrological variability and ecosystem resilience, particularly under growing pressures of water scarcity and ecosystem degradation [7,11]. Traditional monitoring has relied on multitemporal satellite imagery, such as MODIS, which offers high temporal resolution but lacks sufficient spatial granularity to capture small, ephemeral SWE [15,63]. The release of free-access Landsat data by the USGS in 2008 provided an important advancement, combining 30 m spatial resolution with global temporal continuity [21,64]. This has enabled consistent mapping of seasonal and interannual water extent variations and the detection of long-term ecological change [12,14,64].

The present study presents a seasonally explicit approach for mapping SWE dynamics across the IRD using 30 m Landsat data from 2000 to 2024. A key strength of our approach is moving beyond single-index classifications toward a multi-index framework [16,28]. Individually, each index, such as mNDWI, AWEI variants, WI2015, and MBWI, is sensitive to specific water conditions, e.g., open water, shadows, shallow inundation, or fragmented patches [42,44,65]. However, relying on a single index risks systematic misclassification, especially in heterogeneous environments where soil, vegetation, turbidity, and seasonal shadows complicate detection [25,65]. To improve index-specific classification, we applied dynamic thresholding (Otsu’s method) rather than relying on single global thresholds, which often fail under variable seasonal and spectral conditions [46,66]. Dynamic thresholds allow local adjustment of water/non-water separation, increasing sensitivity to shallow or turbid waters that are common in the IRD [46,66,67]. By combining the best-performing indices, using OA, Kappa index, and F1-score, which balance omission and commission errors [18,47], we leveraged their complementary strengths, reduced index-specific biases, and achieved consistent performance across arid, semi-arid, and deltaic landscapes [12,14]. The ensemble consensus produced three seasonal SWE classes, i.e., stable, periodic, and ephemeral, representing a persistent to short-lived water regime. “Consensus” or “voting” approaches of this type are widely used to reduce model bias and increase robustness in environmental mapping [68,69]. Unlike advanced machine learning or probabilistic methods that often generate complex subclasses and decision rules, our ensemble-occurrence framework deliberately avoids over-fragmentation [40,58]. Instead, it provides a simple yet transparent categorization of degradation and transition states. The choice of the thresholds (e.g., MIN_VALID_OBS_FRAC, minimum occurrence for seasonal and temporary classes, slope limits for diminishing and intensifying classes) was guided by exploratory sensitivity tests informed by global studies in comparable climatic and geomorphological settings [13,14,27,40]. This balance between scientific rigour and practical usability is crucial for wetland management and conservation, where decision-makers often require interpretable and actionable outputs rather than data-hungry models that may be difficult to apply in resource-limited contexts [5,56].

Our findings highlight both the ecological significance and vulnerability of dryland deltas such as the IRD. Despite their role in biodiversity conservation and regional water security, these systems remain underrepresented in global surface-water datasets, which largely focus on temperate and humid regions [14,69]. Furthermore, there are few seasonally explicit studies for the large region that comprises Arid Central Asia [13,26,27]. In the Tarim River Basin and the Amu Darya River Delta, for example, Landsat-based time-series analyses have shown that ecological water transfers and flow regulation can substantially reshape the spatial distribution of permanent and seasonal surface waters, with local wetland gains occurring alongside downstream degradation and increased hydrological fragmentation [26,27]. Our results for the Ile–Balkhash system are consistent with this broader Central Asian pattern, in that regulated inflows and distributary allocation strongly govern the persistence and loss of seasonal wetlands, but they also highlight a particularly pronounced decline in summer and fall periodic SWE, indicating high sensitivity to warm-season water shortages. By tailoring our method to the hydrological complexity of an IRD region, we capture seasonal variability and multi-decadal transitions with high spatial fidelity. This dataset not only provides a more ecologically meaningful understanding of water dynamics but also highlights the importance of locally calibrated, seasonally explicit approaches, rather than relying solely on global products [17,70].

4.2. Multi-Scale Factors and Controls of SWE Dynamics

The spatio-temporal dynamics of SWE in the IRD reveal a complex interplay of hydrological, climatic, and anthropogenic factors [29,37]. While earlier studies have examined long-term surface-water change at annual or decadal scales, seasonal variability has often been overlooked. By incorporating seasonal variations into our analysis, we observed significant differences in SWE and its pattern across space and time, consistent with other dryland studies emphasizing the importance of intra-annual and interannual hydrological fluctuations [10,12,15,26].

The MMK trend analysis reveals long-term declines in both stable and periodic SWE, consistent with broader evidence from regulated dryland deltas where upstream inflows have diminished wetland extent [11,14]. In the IRD, stable SWE showed a significant decline (β = −25.2 km² yr⁻¹), while periodic classes contracted most sharply in summer and fall. Similar seasonal asymmetry has been reported in other regulated basins, where summer emerges as the most vulnerable period due to low baseflows coinciding with high evaporative demand [10,12,57].

Correlation results reinforce these patterns. Stable SWE were strongly linked to upstream D and reservoir WL in spring (r = 0.47–0.49, ρ < 0.02), emphasizing the role of Kapchagay reservoir operations in sustaining wetland cores during the early growing season [29], who showed that Kapchagay reservoir releases largely determine delta inundation patterns. By contrast, periodic SWE were most responsive to D and WL fluctuations in summer (r = 0.56–0.58, ρ = 0.003) and fall (r = 0.54–0.64, ρ ≤ 0.005), echoing earlier findings that floodplain inundation depends on peak snow and glacial melt [31,56,63].

Although temperature and precipitation are widely recognized as key climatic drivers controlling wetland dynamics globally [8,11], their influence in the IRD was asymmetric. Periodic SWE showed significant negative correlations with summer AT (r = −0.47 to −0.64, ρ < 0.05), confirming that warming intensifies evapotranspiration and reduces surface water extent. This aligns with regional climate analyses documenting steady warming in Central Asia [31,35,56] with evidence that increased evaporation has already amplified water stress in the Balkhash basin [13,30]. Ephemeral SWE, by contrast, displayed episodic positive correlations with D in fall (r = 0.43, ρ = 0.032), suggesting short-term inundation driven by flow pulses and backwater effects rather than direct rainfall [55,57]. Precipitation at Bakanas displayed a modest spring correlation (r = 0.41, ρ = 0.046), but no other significant relationships were detected.

Overall, local precipitation showed little influence on SWE across seasons, despite being a globally recognized wetland driver [11,71]. This weak signal is not unexpected in semi-arid contexts, where low annual rainfall and high evapotranspiration limit direct precipitation impacts on wetland hydrology [29,37,56]. Instead, wetland persistence depends on upstream inputs from snow and glacier-fed tributaries such as the Ile, Tekes, and Kunes Rivers [13,31,56,72]. Peak discharge in July (~950 m³/s), and sharp declines in September, indicate expansion during spring–summer and contraction in fall [32]. The Kapchagay Reservoir, built in 1970, has since attenuated these pulses and redistributed flows for irrigation and hydroelectric power [13,32,37].

The optimal-lag analysis further refines this picture by showing that these seasonal controls operate within a multi-year hydrological memory [26]. In our case, the stable SWE is most strongly related to D and WL at 1–2-year lags, whereas the periodic and ephemeral SWE classes respond mainly to current-year to 1-year antecedent flow and storage (Table S7). This hierarchy is consistent with regulated river–wetland systems in dry and cold regions, where antecedent hydrological conditions over several years shape downstream water availability and wetland extent [26]. In contrast, AT and P act mainly as modulating climatic drivers, amplifying or dampening these hydrological controls in specific seasons, rather than as primary determinants of SWE.

Spatial heterogeneity across AOIs further mirrors these patterns. AOI-II and AOI-III, fed by the Zhideli branch, host the largest seasonal and temporary wetlands but also showed the strongest evidence of loss and diminishing classes, reflecting the instability of distributary-fed deltas [2,36]. AOI-I, supplied by a smaller but steadier branch, sustained more persistent wetlands, while AOI-IV, despite higher water levels, supported only fragmented wetlands due to a weak discharge connectivity pattern consistent with reported declines in regulation resilience of smaller distributaries [32,37].

These multi-scale patterns underscore the IRD’s “boom-and-bust” hydrology, whereby wetland expansion in wet years alternates with rapid contraction during droughts [32,36]. The 2012–2014 drought, affecting more than half of Kazakhstan, corresponded with major surface-water losses [73]. Human pressures, including irrigation, land-use conversion, and population growth, have intensified these fluctuations by reducing natural cover and sediment supply, weakening floodplain resilience [6,13,56]. Together, these findings highlight the urgent need for integrated, transboundary water-management strategies to sustain the ecological and hydrological integrity of the Ile–Balkhash system [7,30,56].

4.3. Implications for the Management and Restoration of the Wetlands

Surface water dynamics have been identified as one of the ten recommended satellite monitoring variables necessary for systematically tracking and achieving global biodiversity targets [74]. The IRD, a Ramsar-designated wetland of international importance, requires systematic monitoring to sustain biodiversity, regulate hydrology, and maintain ecosystem services [6,7,34]. Continuous observation of hydrological changes aligns with Ramsar’s conservation objectives by informing science-based policy and management interventions [34].

Our results highlight major hydrological transformations, with seasonal wetlands dominating the delta (3418.8 km²; 49.4%) followed by temporary wetland (1435.7 km²) classes. This is further accompanied by significant fragmentation into diminishing (768.68 km²) and lost (98.46 km²). AOI-level contrasts are particularly useful. AOI-II, despite being the most water-rich (2925.2 km²) and crucial for connectivity, is highly vulnerable, with more than 25% of its wetlands classified as temporary and 30% as diminishing. This indicates a net reduction in seasonal water availability, largely influenced by drier climatic trends and anthropogenic modifications [2,71,75]. Furthermore, the impact of controlled flow has led to fragmentation of wetland areas, with a sharp decline in floodplain connectivity and expansion of desert-like ecosystems [32,34]. Historically, tributaries such as Topar-I and II contributed significantly to the wetland hydrology but now contribute only during high flow regimes [34,35]. This trend is particularly evident in the AOI-IV region with the lowest wetland coverage (19% of its polygon area) and the highest share of diminishing waters (37.7 km²; 4.84%), reflecting severe hydrological stress [32,34]. In contrast, AOI-I has shown resilience, with intensifying wetlands covering 129.46 km² (15.42%), primarily due to its proximity to active channels and sustained long-term water levels [35,73].

These results emphasize the urgent need for an Integrated Water Resource Management (IWRM) approach to balance the competing demands of agriculture, industry, and conservation [7,13]. Coordinated water allocation strategies are required to preserve seasonal flow patterns and upstream–downstream connectivity. Reservoir operations, particularly at Kapchagay, should aim to mimic natural flow regimes to sustain wetland hydrology and habitat diversity [37]. Current water-sharing frameworks between China and Kazakhstan remain inadequate for balancing economic and ecological needs [2,6]. Strengthening these agreements to include ecological flow requirements is essential for long-term basin stability.

Given the significant wetland losses (Table 5), targeted restoration should prioritize areas experiencing the most severe degradation (e.g., AOI-IV). Conservation efforts such as reforestation, riparian vegetation recovery, and hydrological rehabilitation can enhance water retention, reduce soil salinity, and restore habitat quality [6,29,31]. Adaptive management strategies to mitigate climate variability and anthropogenic modifications must be a priority [9,57]. Signs of natural intensification present opportunities to reinforce resilience by maintaining active-channel flows and protecting vegetated corridors.

Recognizing the extensive functional shifts over recent decades, restoration targets should be realistic. Returning the delta to its pristine state is no longer feasible [5,9,57]. Instead, new reference conditions should be guided by social–ecological sustainability principles, balancing historical baselines with contemporary constraints (e.g., climate variability, water-use competition). Rising temperatures and increased evaporative losses, particularly in summer and fall, pose additional threats to wetland persistence (9,57). Strategies such as improved irrigation efficiency, groundwater recharge initiatives, and systematic monitoring of extreme events are critical for maintaining long-term wetland stability [7,12,75]. As the Ile River Basin is shared between Kazakhstan and China, coordinated transboundary management is imperative. Extensive dam networks supporting upstream hydropower and irrigation have altered flow timing and reduced sediment delivery, often conflicting with downstream agricultural and ecological needs, exacerbating wetland degradation risks [27,29,37]. Strengthening bilateral agreements to ensure environmental flows and harmonized monitoring should be central to the resilience of the Ile–Balkhash system [56,76].

Overall, this study underscores the intricate feedback between hydrology, climate, and human intervention shaping the IRD. Addressing these interlinked pressures requires a collaborative, science-based, and adaptive governance framework that balances socio-economic needs with ecological integrity. Such integration across research, policy, and local practice is critical to sustaining the long-term functionality and biodiversity of this globally significant dryland delta [4,56].

5. Conclusions

This study presents the first seasonally explicit, multi-decadal assessment of surface water extents in the IRD, southeast Kazakhstan, a Ramsar-designated wetland of international importance in ACA. Using 30 m Landsat data (2000–2024) and an ensemble multi-index framework with dynamic thresholding, we mapped stable, periodic, and ephemeral SWE to capture the overall water dynamics. This approach avoided the over-fragmentation typical of advanced classifiers while ensuring transparency and scalability for long-term monitoring.

Across the 2000–2024 period, the IRD experienced a significant decline in SWE, with the strongest reductions in summer and fall. Seasonal wetlands dominate the delta (~3419 km²), reflecting the intermittent nature of water availability. However, their contraction, together with the loss of permanent waters, indicates severe disruption of hydrological connectivity following the construction of the Kapchagay Reservoir. Spatial analysis revealed clear contrasts. AOI-II is the most hydrologically active zone due to inflows from the Zhideli branch, whereas AOI-IV remains the driest, supporting only minimal and fragmented wetlands.

Trend analysis confirmed significant drying trends, particularly for seasonal and ephemeral SWE, while correlation results showed that upstream discharge and reservoir water levels are the dominant hydrological controls of SWE dynamics. Rising summer air temperatures intensified evapotranspiration, amplifying seasonal contractions, whereas local precipitation showed weak or non-significant influence. These findings highlight that wetland sustainability in the IRD is primarily dependent on upstream snowmelt and glacial runoff and strongly shaped by flow regulation at the Kapchagay Reservoir.

Taken together, our results show that the IRD wetlands are highly vulnerable to combined climatic and regulatory pressures. Enhancing resilience requires an Integrated Water Resources Management approach that balances agricultural, industrial, and ecological needs. Ecologically timed releases, more efficient irrigation, groundwater recharge, and climate-adaptive management are all critical to alleviate summer water stress. Restoration should prioritize vulnerable AOIs with high shares of diminishing wetlands, and transboundary cooperation between Kazakhstan and China must incorporate ecological flow provisions to sustain downstream connectivity.

Beyond its regional relevance, the ensemble–occurrence framework applied here provides a scalable, cost-effective methodology for seasonally resolved wetland monitoring in other dryland deltas worldwide. Integrating long-term earth observation with hydrological and climatic drivers, our workflow contributed to both scientific understanding and the management of vulnerable wetlands under changing climate and regulatory regimes.