Interannual Responses of Common Reed (Phragmites australis) to Fluctuating Water Flows Entering the Ili River Delta, Kazakhstan

Abstract

Kazakhstan’s Ili River delta nourishes a unique wetland ecosystem in arid Central Asia. The delta is dominated by common reed [Phragmites australis (Cav.) Trin. ex Steud.], an ecologically and economically significant species that is sensitive to water levels. We used machine learning methods, including the Random Forest algorithm, to classify common reed-containing wetland vegetation and water surfaces in the delta based on satellite data from 2000 to 2023. The remote sensing results were integrated with field-based geobotanical studies conducted between 2017 and 2019. These studies, which facilitated identification of three classes of vegetation: meadow, marsh and aquatic, also revealed a good correspondence between predicted and measured common reed biomass per unit area. The long-term dynamics of four wetland communities with a common reed content of ≥25% were analyzed with respect to significant interannual variability in the flow of the Ili River. The highest water inflows and water surface areas in the delta were recorded in 2002, 2010 and 2016, and in each case, a statistically significant expansion of common reed was observed one year later. Common reed areas subsequently declined—rapidly or after a lag of several years. Temporal expansion and contraction of these areas following pulses of water differed substantially from that of overall wetland vegetation as measured previously. The identified patterns have important scientific and practical significance for assessing the stability of the surrounding wetlands, preservation of the delta environment, and sustainable use of common reed.

1. Introduction

Wetlands are among the most valuable ecosystems on earth because they regulate water quality, retain pollutants, accumulate and transform organic matter, support high biodiversity, and provide a wide range of ecosystem services [1,2,3,4,5,6,7,8,9]. These services include shoreline protection, groundwater recharge and storage, climate-regulating functions, and recreational value [5,6,7,8,9,10]. Nevertheless, wetlands worldwide are undergoing extensive degradation due to hydrological alteration, habitat destruction and fragmentation, overexploitation of biological resources, pollution, and climate change [11,12,13,14,15,16]. According to global assessments associated with the Ramsar Convention, wetland area declined by approximately 35% between 1970 and 2015 [9,14], creating a major challenge for the conservation and sustainable management of freshwater ecosystems [15,17,18,19].

These problems are particularly acute in arid and semi-arid regions, where ecosystem structure and functioning are strongly constrained by water availability. In Central Eurasia, large river systems support ecologically important inland wetlands within predominantly dry landscapes [20,21]. One of the most representative examples is the Ili River, which flows from China into southeastern Kazakhstan and forms the largest delta in Eurasia before entering the endorheic Lake Balkhash [22,23]. Wetland ecosystems of the Ili–Balkhash basin have developed through long-term interactions among variable soil conditions, river discharge, and surface- and groundwater dynamics, creating a complex mosaic of azonal habitats [24,25,26,27]. Construction of the Kapchagay Dam and reservoir, which disrupted the timing and flow of the river, was the first major anthropogenic hydrological disturbance in the delta [28,29,30]. In recent decades, increasing water abstraction for irrigation in both the Chinese and Kazakh portions of the basin has further reduced downstream inflow and intensified pressure on deltaic wetlands [30,31,32].

The Ili River Delta is characterized by a heterogeneous combination of desert landscapes and floodplain habitats that include open water bodies, marshes, wet meadows, and tugai vegetation [32]. The dominant structural component of these wetlands is common reed, Phragmites australis (Cav.) Trin. ex Steud. This species, which is hereafter referred to as reed or P. australis, occupies large areas of the delta and forms extensive beds. It has high biomass-production potential [33,34] and has long supported traditional use of the delta’s natural resources, including animal husbandry, hunting, and fishing [34,35]. Reed often forms dense monodominant stands that lack diversity but play an important role in habitat formation, primary productivity, and overall ecosystem functioning [36,37].

The spatial and temporal dynamics of reed are unquestionably important for wetland ecology and conservation, biodiversity management, and sustainable use of the delta’s biological resources. Previous studies of the Ili–Balkhash system have nevertheless focused on wetland dynamics at the level of broad classes of land cover [20,27,38,39,40], leaving a gap in our knowledge of reed, the dominant species [40,41,42,43,44]. We hypothesized that reed’s response to interannual water availability in the delta may differ from that of broader classes of wetland vegetation. We consequently integrated the results of remote sensing tools such as Landsat imagery, spectral indices, and the Random Forest algorithm with field geobotanical observations. The response of reed to water entering the delta between 2000 and 2023 was characterized by a distinct time lag that differs from the response of broader classes of wetland vegetation, as revealed in previous studies. This underscores the importance of species-specific analysis of reed, the key structural and functional component of the delta ecosystem.

2. Materials and Methods

2.1. Study Area

The Ili River delta (Figure 1) lies near the southeastern shore of Lake Balkhash in a flat, arid depression near the western terminus of the Ili–Balkhash basin [23]. The area is classified as a cold desert and has a harsh continental climate characterized by large seasonal and diurnal temperature differentials such that the average monthly high temperature in July is 25 to 27 °C, and the average monthly low temperature in January is −13 to −15 °C [45]. The annual potential evaporation of about 1200 mm greatly exceeds the average annual precipitation of 135 to 150 mm. This conditions a moisture deficit that increases sharply in early spring and dominates during the summer, when the upper soil profile becomes desiccated [46].

The delta covers roughly 800,000 ha and is a mosaic of the zonal desert ecosystem that typifies the area and interzonal floodplains and valleys that are continually being shaped by erosion-accumulative and scouring processes [32,34]. The interzonal areas support high species biodiversity of plants and animals and consequently are key targets for conservation [42]. As indicated in Figure 1, the delta can be divided into three separate systems, one of which, the central Ili system, contains the main stem of the Ili River [36]. Due to sedimentation of the main stem, the Zhideli system, which lies north of the river, currently receives about 90% of the river’s runoff and encompasses the delta’s most complicated hydrography. In contrast, the southern Topar system is only flooded when the river’s flow is high. Haloxeromorphism is a key feature of the natural vegetation due to the delta’s hydromorphic soils, especially near the lake, where groundwater lies just beneath the surface [42].

2.2. Geobotanical Methods

Field geobotanical surveys of sites containing reed in the Ili River Delta (Figure 1) were conducted during the period of maximum growth and standing biomass of reed vegetation (mid-June to late July) in 2017 and 2019. The work followed standard methods for large-scale geobotanical surveys of natural lands in Kazakhstan [47]. The sampling design was stratified-random and was intended to represent the three main delta systems (Ili, Zhideli, and Topar), the full gradient of hydrological conditions (from drier areas to shallowly flooded zones), and the main types of wetland vegetation in which reed occurs. In total, 102 georeferenced field plots were recorded using Garmin Oregon 750 GPS receivers (Garmin, Olathe, KS, USA) and a pre-prepared map of the study area: 50 plots were located in the Zhideli system, 20 in the Ili system, and 32 in the Topar system.

At each plot, vegetation cover was photographically documented, and, where necessary, plant specimens were collected for species identification using standard floristic keys [48]. Particular attention was paid to relationships among water availability, reed cover, and the composition of accompanying wetland vegetation, especially on small islands, shallow waters, and along the shores of Lake Balkhash and river channels. For each plot, percentage reed cover was estimated, information on accompanying vegetation was recorded, and local habitat characteristics were described. Each field plot was matched to the corresponding Landsat pixel footprint, and field records were used to calibrate and validate the remote sensing analysis and to define five cover classes.

Aboveground reed biomass was assessed by the harvest method at 58 reed-dominant plots during the vegetation period. On typical sites of each plant community, 1 m² sampling quadrats were established with 3–5 replicates. Within each quadrat, the number of reed stems was counted before vegetation was cut at a height of 4–6 cm above the soil surface. The harvested biomass was weighed immediately after cutting to determine fresh mass. To determine dry aboveground biomass, plant samples were dried at 105 °C to absolute dryness, after which the results were converted to t ha⁻¹. These field measurements were used to characterize the spatial variability of reed productivity and to assess the relationship between biomass and remote sensing variables.

To maximize temporal consistency, all field observations used for calibration and validation of the remote sensing analysis were collected within the same seasonal window (June–July), thereby reducing phenological differences unrelated to hydrological forcing. In addition, the temporal transferability of the classifier was assessed by training on one survey year and testing on the other (temporal transfer validation).

2.3. Remote Sensing Methods

To assess the dynamics of wetland vegetation cover in the Ili River Delta, Landsat TM, ETM+, and OLI satellite images with a spatial resolution of 30 m were used. For the period 2000–2023, 60 scenes corresponding to the peak growing season and maximum standing biomass of reed communities (June–July) were selected from the archive of the U.S. Geological Survey (USGS) [49]. To improve interannual comparability, Landsat Collection 2 Level-2 surface-reflectance products were used. Clouds, cloud shadows, and low quality observations were removed using the QA layers supplied with the Level-2 products. All images were processed using a consistent workflow: radiometric and geometric correction, alignment to a common spatial grid, mosaicking, and clipping to the study-area boundary [49,50]. For Landsat 7 ETM+ images acquired after the failure of the scan line corrector (SLC-off), gap masks were used to exclude missing-data stripes from area calculations. Annual composites were generated only from valid pixels within the study area. As a result, 24 annual datasets were produced and used to analyze interannual changes in open water, wetland vegetation, and reed communities [51].

To characterize vegetation cover and moisture conditions, four spectral indices were calculated: the normalized difference vegetation index (NDVI), normalized difference water index (NDWI), enhanced vegetation index (EVI), and normalized difference moisture index (NDMI) [52,53,54,55,56]. The indices were calculated using the following equations:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{\rho NIR-\rho R}{\rho NIR+\rho R}}$$

(1)

$$\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}={\displaystyle \frac{\rho G-\rho NIR}{\rho G+\rho NIR}}$$

(2)

$$\mathrm{E}\mathrm{V}\mathrm{I}=2.5\ast {\displaystyle \frac{\rho NIR-\rho R}{(\rho NIR+6\ast \rho R-7.5\ast \rho B+1)}}$$

(3)

$$\mathrm{N}\mathrm{D}\mathrm{M}\mathrm{I}={\displaystyle \frac{\rho NIR-\rho SWIR}{\rho NIR+\rho SWIR}}$$

(4)

where pNIR is the pixel value in the near-infrared range of the spectrum; pSWIR is the pixel value in the shortwave range of the spectrum; pR is the pixel value in the red range of the spectrum; pG is the pixel value in the green range of the spectrum; and pB is the pixel value in the blue range of the spectrum.

The selected indices reflect complementary properties of wetland vegetation and its hydrological context. NDVI was used as a general indicator of green biomass and vegetation-cover density. EVI was included because it is less sensitive to soil background and saturation effects in dense vegetation canopies. NDMI was used to characterize the moisture status of vegetation cover and substrate, which is particularly important for emergent wetland macrophytes. NDWI was used primarily to identify open-water surfaces and to exclude water pixels before subsequent vegetation analysis. The combined use of these indices accounted for the fact that reed beds in the Ili Delta differ not only in vegetation density but also in their association with shallow flooding, wet substrates, and shallow groundwater.

In addition, a 16 day net primary production (NPP) product (NPP MOD17A3HGF v6.1, 500 m) was used as a productivity predictor [57]. To align NPP with the seasonal Landsat composites, all 16 day NPP layers overlapping June–July of each year were aggregated into a single seasonal NPP layer by calculating the mean value. Because NPP is a continuous variable, the resulting seasonal NPP raster was then reprojected and aligned to the 30 m Landsat spatial grid using bilinear interpolation. The NPP layer was subsequently used together with the spectral indices NDVI, NDWI, EVI, and NDMI as an additional input predictor in the classification of reed and non-reed vegetation.

Mapping was performed as a sequential hierarchical workflow in which thresholding, Random Forest classification, and reed-area calculation had different roles. In the first step, an open-water mask was generated using NDWI; pixels with NDWI ≥ 0.1 were classified as open water and excluded from subsequent vegetation analysis. This threshold was used as a conservative value based on previously published approaches to open-water mapping with Landsat data and was additionally checked by visual comparison with false-color composites and field points. In the second step, within the non-water mask, wetland vegetation was separated from non-wetland vegetation using NDVI thresholds and additional spectral features (NDMI and EVI). These thresholds were used only to preliminarily restrict the analysis to areas with distinct vegetation cover and wet habitats, whereas the final separation of reed and non-reed vegetation was performed using a Random Forest algorithm trained on field data.

In the third step, communities dominated by P. australis were separated from non-reed vegetation using a Random Forest algorithm trained on field geobotanical plots [58,59,60,61]. Training labels were obtained from 102 georeferenced field plots of 100 m² surveyed in 2017 and 2019. At each plot, percentage reed cover was assessed in the field, accompanying vegetation was recorded, and local habitat characteristics were described. Because reed phytocoenoses in the delta are often monodominant but may include meadow, marsh, and aquatic species, field cover estimates were used to define classes of reed participation in the vegetation cover.

The Random Forest model was implemented in Python using the scikit-learn library [59]. NDVI, NDWI, NDMI, EVI, and NPP were used as predictors. The robustness of the key hyperparameter was assessed using stratified five-fold cross-validation. Values of n_estimators = 100, 200, 300, 500, 700, and 1000 were tested while keeping other model settings fixed. Overall accuracy (OA) and F1-score for the reed class stabilized as the number of trees increased; therefore, n_estimators = 700 was selected for the final model and used the following parameters: max_features = “sqrt”, class_weight = “balanced”, max_depth = None, min_samples_leaf = 1, random_state = 42, and n_jobs = −1.

Random Forest performance was assessed, not on the training set, but by stratified five-fold cross-validation. For each fold, the model was trained on 80% of the field plots and tested on the remaining 20%. Thus, each of the 102 georeferenced plots was predicted only in an out-of-fold mode, that is, outside of the training set. This reduced the risk of inflated accuracy estimates and provided a more realistic assessment of the model’s generalization ability.

Model performance was evaluated using overall accuracy (OA), precision, recall, and F1-score for each class; where necessary, Cohen’s kappa coefficient was additionally calculated from the error matrix [62,63]. Because the number of reed and non-reed plots was imbalanced, particular attention was paid to the F1-score for the reed class, which is a more informative metric for binary classifiers under uneven class distributions [64].

To assess temporal transferability, an additional temporal transfer validation was performed between the two years with available field data. This analysis used the same 102 georeferenced field plots and their retained field class labels, whereas spectral predictor values were extracted separately from the annual composites of the corresponding year. In other words, in the Train 2017 → Test 2019 scheme, the model was trained on spectral features extracted for the field plots from the 2017 composite and tested on spectral features extracted for the same plots from the 2019 composite. The reverse Train 2019 → Test 2017 scheme was performed in a similar manner. This approach tested whether the relationship between spectral predictors and reed/non-reed vegetation classes was maintained when transferring the model between years and assessed the feasibility of applying a single classification scheme to the multi-year time series.

Separately from the classification of reed and non-reed vegetation, the relationship between field measurements of aboveground reed biomass and satellite predictors was evaluated. For this purpose, 58 reed plots with field-determined biomass were used, and the corresponding values of NDVI, NDWI, NDMI, EVI, and NPP were applied as explanatory variables. Based on these data, a Random Forest model was built to evaluate how well spectral and productivity features reflected the spatial variability of reed-bed biomass. This model was used to compare measured and predicted biomass and was not used to calculate the area of reed communities.

At the final stage, field reed-cover classes were used to convert classified maps into an estimate of actual reed-bed area (

Table 1. Generation of coefficients to determine reed areas in classified remote images. The coefficients are based on analysis of reference plots in 2019.

Classification Cluster	Total Area (km2)	Reed Area (%)	Coefficient	Reed Area (km2)
Class 00 Waterbody	667.0	0	0	0
Class 01 Non-reed	1004.9	<25	0	0
Class 02 Reed	971.6	25–55	0.25	242.9
Class 03 Reed	987.6	55–75	0.55	543.2
Class 04 Reed	773.3	75–90	0.75	580.0
Class 05 Reed	298.6	>90	0.90	268.7
Reed classes 02–05	3031.1			1634.8

). It is important to emphasize that the classes <25%, 25–55%, 55–75%, 75–90%, and >90% represent, not Random Forest probabilities, but field estimates of the proportion of reed cover. For classes with reed cover of at least 25%, conservative conversion coefficients corresponding to the lower bounds of the cover intervals were assigned as shown in Table 1. This approach reduces the risk of overestimating reed area in mixed pixels. Vegetation with reed cover below 25% was treated as non-reed or weakly represented by reed and was excluded from the final reed-area estimate. The overall approach to mapping reed and wetland vegetation was based on previously applied remote sensing methods for wetlands and reed communities [65,66,67].

The reed area for each year was calculated as the sum of class areas multiplied by the corresponding cover coefficients:

$${RB}_{area,y}={\sum }_{i=1}^n{(A}_{i,y}\ast C_i)$$

(5)

where ${RB}_{area,y}$ is the estimated area of reed beds in year y, $A_{i,y}$ is the area of class i in year y, and $C_i$ is the reed-cover coefficient for that class. Thus, NDWI and NDVI thresholds were used for preliminary masking of water and wetland vegetation, Random Forest was used to distinguish reed from non-reed vegetation, and field cover coefficients were used to convert classified classes into an estimate of reed area.

Classification accuracy was evaluated using 102 field plots surveyed in 2017 and 2019. Classification reliability was determined using an error matrix [62], overall accuracy, precision, recall, and F1-score; overall agreement was additionally assessed using the kappa coefficient [63]. Because the class set was imbalanced, special attention was paid to the F1-score for the reed class, which is more informative when evaluating binary classifiers with uneven class distributions [64].

2.4. Uncertainty Assessment of Reed-Bed Area

Uncertainty in reed-bed area estimates was associated with two main sources: classification error in distinguishing reed and non-reed vegetation and uncertainty in reed-cover coefficients used to convert classified classes into estimates of actual reed area. Classification error was assessed using stratified five-fold Random Forest cross-validation, the error matrix, overall accuracy, precision, recall, F1-score, and kappa coefficient [62,63,64].

To evaluate the sensitivity of the final reed area to the selected cover coefficients, a scenario analysis was performed. The baseline scenario used conservative coefficients corresponding to the lower bounds of field-observed reed-cover intervals: 0.25 for the 25–55% class, 0.55 for the 55–75% class, 0.75 for the 75–90% class, and 0.90 for the >90% class. Reed area was additionally recalculated by varying the cover coefficients by ±10% relative to the baseline values. This analysis allowed us to assess how sensitive the absolute estimates of reed-bed area were to uncertainty in the cover coefficients.

The main interpretation of interannual dynamics was based not only on absolute area values but also on the stability of temporal patterns under changes in the cover coefficients. If the direction of changes and the years of maximum reed expansion were retained under alternative coefficients, the dynamics were considered robust to uncertainty associated with the conversion of mixed vegetation classes.

2.5. Hydrological Data and Statistical Analysis

Hydrological data for the Ili River were obtained from the Kazhydromet archive [68]. To characterize the hydrological regime of the lower Ili River during 2000–2023, observations from the Kapchagay-37 gauging station (44°7′48.21″ N, 76°59′12.98″ E; Figure 1), located 37 km downstream of the Kapchagay Reservoir dam, were used. Mean annual river discharge (Q) was used as the main hydrological indicator and was compared with open-water area (OW) and reed-bed area (RB) calculated from remote sensing data. Pearson’s linear correlation coefficient was used to assess relationships among river discharge, open-water area, and reed-community area [69].

To identify a possible delayed response of reed communities to changes in river discharge, the analysis was performed for three temporal variants: no lag, a one-year lag, and a two-year lag. In the no-lag variant, river discharge in year t was compared with open-water and reed area in the same year. With a one-year lag, river discharge in year t was compared with open-water and reed area in the following year (t + 1). With a two-year lag, river discharge in year t was compared with the corresponding areas two years later (t + 2). This scheme allowed us to test whether the response of reed to hydrological pulses was immediate or delayed.

Because annual hydrological and vegetation time series may contain temporal autocorrelation, lagged relationships were assessed not only using Pearson’s r. To test the robustness of the lag effect, three additional procedures accounting for temporal dependence were applied: (i) the significance of Pearson correlations was recalculated using the effective sample size (Neff) computed from the lag-1 autocorrelation of the compared series; (ii) generalized least-squares models with AR(1) errors [GLS-AR(1)] were additionally fitted, with reed-community area (RB) as the dependent variable and mean annual river discharge (Q) with lags of 0, 1, or 2 years as the predictor; and (iii) 95% confidence intervals for lagged correlations were calculated using moving block bootstrap. The lag effect was considered robust only if the sign and strength of the relationship were retained under different ways of accounting for temporal dependence.

All statistical calculations were performed in Python (v3.9.17) using standard libraries for statistical analysis and machine learning. Results were presented as correlation coefficients, significance levels, p values adjusted for effective sample size, parameters of GLS-AR(1) models, and confidence intervals obtained by moving block bootstrap.

3. Results

3.1. Reed Communities of the Ili River Delta

The overall landscape of the Ili River delta is a complex patchwork of plant communities, almost all of which contain reed in some proportion, often as the monodominant species. Detailed field surveys were conducted over the period from 2017 to 2020 to facilitate accurate extraction of reed areas from classified remote images and provide complementary ground-based information on the relationship between water, reed, and other wetland vegetation. A total of 37 species in different categories of rarity was observed during our field surveys of the delta. They represent nearly 10% of the total number of rare and endangered plant species in Kazakhstan [70,71]. The following three classes of vegetation were identified:

Meadow-type vegetation occurs on marsh and marsh-meadow soils and is influenced by water availability and salinity. The delta’s marshy meadows are formed under conditions of temporary and prolonged flooding and usually appear in flat, wide depressions and low areas between ridges. Areas of meadow-type vegetation are sometimes hand cut and harvested for use as fodder (Figure 2A). They also serve as pastures for horses and other farm animals (Figure 2B). Plant communities are usually monodominant with reed, cattail (Typha angustifolia L.), and saltmarsh bulrush [Bolboschoenus maritimus (L.) Palla] when moisture is adequate. Under suboptimal conditions for reed, bushgrass [Calamagrostis epigejos (L.) Roth] and coastal small reed [C. pseudophragmites (Haller.f) Koeler] become subdominant, usually along the drying periphery of reed beds. Regularly flooded depressions containing alluvial-meadow soils support grassy oligohalophytes that prefer fresh water but can tolerate salinity. Couch [Elytrigia repens (L.) Gould], bushgrass, and reed dominate in these areas, which form pastures. The distribution of grasses changes according to the water levels in a given year, whereby the roots and rhizomes survive drier years and therefore can support the expansion of a grass species under wetter conditions. The most common associated species, which sometimes appear as subdominants, include Chinese licorice [Glycyrrhiza uralensis Fisch ex DC], alkali swainsonpea [Sphaerophyza salsula (Pall.) DC], manystem wild rye [Leymus multicaulis (Kar. & Kar.) Tzvelev], banat sweetclover [Melilotus dentatus (Waldst. & Kit.) Desf.], perfoliate gypsophila (Gypsophila perfoliate L.), barnyard grass (Panicum crus-galli L.), Dodartia orientalis L., and Hordeum bogdanii Wilensky.

Marsh-type vegetation is formed on stable combinations of non-saline marsh and meadow marsh soils and is typical for sites along the southern shore of Lake Balkhash in the Topar and Ili systems (Figure 1). Hydrophilic vegetation thrives in relief depressions of these areas under conditions of almost constant flooding and consists of beds of reed and cattail with associated saltmarsh bulrush, as well as other bulrush and sedge species (Scirpus and Carex spp.) (Figure 2C). The combination of dense vegetation and high moisture levels leads to the accumulation of decaying root material and surface litter, which when thick enough, can interfere with the growth of reed. The coverage area of marsh-type vegetation varies greatly from year to year depending on the availability of water.

Aquatic-type vegetation is characteristic of the lower reaches of the delta. Vast reed beds cover the entire southern shore of Lake Balkhash and form ribbons along channels where the water depth is 20–60 cm (Figure 2D). Insufficient water can cause these beds to recede and even disappear. In addition to reed, rush, and cattail, the following rooted species are widely represented in standing water of the delta: flowering rush (Butomus umbellatus L.), spikerush [Eleocharis acicularis (L.) Roem. & Schult.], spiny water nymph (Najas marina L.), Eurasian water milfoil (Myriophyllum spicatum L.), water smartweed (Persicaria amphibia (L.) Gray], and marsh fern (Thelypteris palustris Schott), as well as pondweed (Potamogeton spp.), arrowhead (Sagittaria spp.), softstem bulrush (Scirpus spp.), and bur reed (Sparganium spp.). Rare dwarf white water lily (Nymphaea candida C. Presl) and threatened species such as fringed water lily [Nymphoides peltata (S.G. Gmel.) Kuntze] waterwheel plant (Aldrovanda vesiculosa L.), and Scirpus kasachstanicus Dobrochot. are also occasionally present. Floating aquatic vegetation in these areas consists of duckweed (Lemna minor L.), bladderwort (Utricularia vulgaris L.), lesser bladderwort (U. minor L.), yellow water lily [Nuphar lutea (L.) Sm.], and hornwort (Ceratophyllum demersum L.).

3.2. Spatiotemporal Relationships Between Runoff, Open Water, and Reed Area in the Ili River Delta

Analysis of the satellite image series for 2000–2023 showed that the applied mapping scheme consistently distinguished open-water surfaces, non-reed vegetation, and communities dominated by reed. Classification accuracy was assessed using 102 georeferenced field plots surveyed in 2017 and 2019. Overall classification agreement, expressed by the kappa coefficient, was k = 0.86, indicating high reliability of information extracted from remote sensing data [39,72]. Producer’s and user’s accuracies were 87% and 82%, respectively. The Random Forest model was used for additional separation of reed and non-reed vegetation based on spectral and productivity predictors. Classification performance metrics for 58 reed and 44 non-reed plots are presented in

Table 2. Quality indicators for the classification of 58 reed and 44 non-reed communities in the Ili River delta using the Random Forest algorithm. Reed and non-reed refer to aboveground vegetation.

Indicator	Precision	Recall	F1-Score
Reed (N = 58)	0.81	0.93	0.86
Non-reed (N = 44)	0.89	0.70	0.78
Macro average (N = 102)	0.85	0.82	0.82
Weighted average (N = 102)	0.85	0.83	0.83

.

The results of temporal transfer validation showed that the model retained acceptable robustness when transferred between the two years of field observations. When the model was trained on 2017 data and tested on 2019 data, OA = 0.802 and the F1-score for the reed class = 0.808 were obtained. In the reverse transfer, that is, when trained on 2019 data and tested on 2017 data, the values were OA = 0.781 and F1-score for the reed class = 0.828. In both cases, the assessment was based on 102 georeferenced field plots. These results confirm that the relationship between spectral predictors and reed/non-reed vegetation classes is maintained between years and supports the use of a single classification scheme for retrospective analysis of the 2000–2023 period.

Overall, the model provided reliable separation of reed and non-reed vegetation. The reed class had a particularly high recall value, indicating that most reed communities were correctly identified. The lower recall for the non-reed class was probably related to the greater spectral and structural heterogeneity of non-reed vegetation, as well as the presence of transitional ecotone communities. Nevertheless, F1-score values and averaged metrics indicate balanced overall classification performance.

Variable importance in the Random Forest model was evaluated using permutation importance on held-out data within stratified five-fold cross-validation (mean ± SD across folds). NDVI was the most informative predictor, indicating the key role of vegetation-cover density and condition in distinguishing reed communities. NDWI and NPP had moderate contributions, whereas NDMI and EVI, in this feature configuration, had near-zero mean importance and high variability across folds (

Table 3. Cross-validated permutation importance of predictors in the Random Forest model for distinguishing reed and non-reed communities (mean ± SD across folds).

Variable	Mean Permutation Importance	SD Across Folds
NDVI	0.208	0.051
NDWI	0.038	0.037
NPP	0.029	0.025
NDMI	0.006	0.031
EVI	0.005	0.032

).

Permutation importance was calculated as the mean decrease in accuracy after random permutation of a predictor’s values in the test subset of each fold. Values are reported as mean ± SD across five stratified folds (20 permutations per fold).

The additional uncertainty assessment showed that absolute reed-bed area values depend on the cover coefficients used to convert mixed vegetation classes. When cover coefficients were varied by ±10%, the estimated reed area changed proportionally; however, the main interannual dynamics were preserved, i.e., the years of minimum and maximum area and the pronounced expansion of reed communities after high-water periods remained unchanged. Therefore, the detected temporal pattern, including the lagged reed response after peak river-discharge values, was not an artifact of the selected cover coefficients. Absolute area values should nevertheless be interpreted as conservative estimates because the lower bounds of field cover intervals were used for mixed classes.

Modeling of aboveground reed biomass using spectral and productivity predictors was also conducted. Field biomass measurements from reed plots and corresponding satellite variables (NDVI, NDWI, NDMI, EVI, and NPP) were used to build a Random Forest model linking ground observations to remote sensing data. Figure 3 presents the relationship between measured aboveground reed biomass and biomass predicted from remote sensing data using the Random Forest model. In all three delta systems, the relationship was strongly positive, confirming high agreement between field observations and satellite predictors. This indicates that the combination of spectral indices and NPP adequately reflects the spatial variability of reed community productivity. The strongest relationship between measured and predicted biomass was obtained for the Topar system, but values for the Zhideli and Ili systems were also high, with R-square values exceeding 0.81. These results indicate good explanatory capacity of the model and confirm that differences in the spectral characteristics of reed communities are associated, not only with their presence or absence, but also with quantitative differences in productivity.

Reed biomass was characterized by high spatial variability within the delta. The mean value of field measurements was 8.43 t ha⁻¹, and the range varied from 1.16 to 22.31 t ha⁻¹, corresponding to an almost 19-fold difference among plots. The highest biomass values were recorded in the Zhideli system (Figure 3). High values were also recorded in the Ili system, although they did not exceed 20.44 t ha⁻¹. By contrast, in deeply flooded habitats of the Topar system, biomass was generally lower, around 2.0–3.0 t ha⁻¹. These differences show that reed productivity is maximized not under extreme flooding, but under optimal moisture conditions, when sufficient water supply is maintained without excessive inundation of the root zone.

Temporal relationships among mean annual river discharge, area of open water, and area of reed community are shown in Figure 4 and Figure 5. Mean annual discharge over the study period was 498 m³ s⁻¹, but interannual variability was high. The lowest discharge values occurred in 2014, 2015, and 2020, with the absolute minimum of 350 m³ s⁻¹ recorded in 2020, whereas the maximum mean annual discharge was 719 m³ s⁻¹ in 2010 [68]. Reed area varied from 1122 km² in 2009 and 1144 km² in 2000 to a maximum of 2260 km² in 2017. Open-water area generally followed changes in river discharge; the most pronounced flood events occurred in spring and summer of 2002, 2010, and 2016, when mean discharge during April–September was 852, 998, and 869 m³ s⁻¹, respectively [68]. These years were accompanied by noticeable expansion of the channel network and an increase in the area of small lakes within the delta.

The positive relationship between peak flow years and reed area expansion was evident but appeared only with a time lag (Figure 4). In each case, reed area increased substantially in the year following a high-water period and then declined as water availability decreased. Correlation analysis confirmed this pattern (Figure 5). Within the same year (t), river discharge was strongly related to open-water area but only weakly related to reed area. When a one-year lag was introduced, with river discharge in year t compared with open water and reed area in year t + 1, the correlation between discharge and open water weakened, whereas the correlation between discharge and reed area strengthened. With a two-year lag, correlations again became weak.

The lag-1 relationship “discharge in year t → reed area in year t + 1” retained a positive sign and statistical significance after accounting for temporal dependence: for lag-1, r = 0.668 (p = 0.0018); after correction for effective sample size, padj (Neff) = 0.0025; and in the GLS-AR(1) model, the discharge coefficient also remained statistically significant (p = 0.00046). The moving block bootstrap confidence interval for the lag-1 correlation was 0.18–0.84 and did not include zero. By contrast, same-year (lag 0) and two-year lag (lag 2) relationships remained unstable. Thus, reed communities respond to hydrological pulses primarily with a delay of one year.

The magnitude and duration of reed bed expansion varied depending on the specific flood event. Expansion after the high-water years of 2002 and 2016 was more pronounced than after 2010, and the duration of the expanded area also varied. For example, the 2016 flood was followed by an approximately 50% increase in reed area in 2017, but by 2018 the area had already returned to a level close to that of 2016. In contrast, expansion after earlier flood events persisted longer before the area again declined to near pre-flood levels. Field observations suggest that this variability is related to differences in the rate of post-flood drying.

Maps of reed distribution over a four-year period from 2015 to 2018 (Figure 6) illustrate the response of reed areas, which were mainly concentrated around small water bodies and the channel network in the northern part of the Zhideli delta, to favorable and unfavorable hydrological conditions. Relatively modest reed areas in 2015 corresponded to restricted water availability in this and several previous years (Figure 4). These areas expanded somewhat in 2016, a flood year. The largest reed area of the 21st century was recorded in 2017, one year after the flood peak, when dense reed beds occupied a substantial portion of the Lake Balkhash shoreline, and peripheral zones of reed communities expanded noticeably into wetter parts of the Ili and Zhideli systems (Figure 6). Contraction in reed areas was evident in 2018, when reed area was only slightly greater than that recorded in 2016.

4. Discussion

4.1. Status of Reed in the Ili River Delta

The status of the geographically isolated population of reed in the Ili River delta is dependent on variable and unstable water inflows from the river [43,73] and governed by a set of complex factors. Some, such as the basin’s severe climate, are long-term natural phenomena. Others, including diversion of water behind upstream dams, tourism and recreation in the delta [74,75], and reed harvests [76,77,78], represent more recent human impacts. Factors such as these have received significant attention in Europe and North America, where researchers have documented fluctuations in reed populations, identified the causes of increases and declines, and formulated measures to prevent and mitigate damage to reed beds [79,80,81]. In comparison, our understanding of reed dynamics in the arid, ecologically fragile Ili River delta is fragmentary.

Analysis of reed-containing vegetation in the Ili River delta has increasingly relied on remotely sensed imagery [20,65,66,67,82] and tended to emphasize broad classes of wetland flora rather than individual species [27,39,40,83,84]. Contractions and expansions of wetland areas, changes in patterns of vegetation across the landscape, and relationships between wetland coverage and the availability of water have all been investigated. Remote imagery has also been used to calculate the extent of various reed-containing assemblages in the delta and map their locations [43]. Sivanpillai et al. [41], for example, determined that the area of reed beds, reed interspersed with other vegetation, and reed along water totaled about 424,000 ha as of 2001. The area of submerged reed, non-submerged reed, and open reed and shrubs was later estimated to be about 350,000 ha [42]. Other estimates of the area of reed ecosystems in the delta are slightly lower, ranging from 243,000 to 303,000 ha [39].

These and other studies [20,33,43,44,85] demonstrate that the extent of reed in the delta varies in response to changing hydrological conditions, but definitions of reed-containing areas lack uniformity, and time intervals between measurements are variable and often encompass multiple years. This complicates assessment of the rapidity and magnitude of reed’s responses to the well-known interannual flooding and drying events that characterize the delta. We addressed these dynamics with a time series of data spanning more than two recent decades. Ground-based surveys resolved the delta’s reed-containing plant communities [27,38] into three categories that are distinguishable by the patterns and extent of reed cover, as well as the identity and distribution of reed-associated plant species. Relatively large areas of meadow- and marsh-type vegetation containing reed as a minor component (class 01 in Table 1) were excluded from our calculations, thereby lowering overall estimates of the delta’s reed area in comparison to those published previously, while focusing analysis on vegetation containing significant amounts of the target species.

4.2. Spatiotemporal Responses of Reed to Inflows from the Ili River

Evidence that very substantial increases in the delta’s reed area closely track high Ili River inflows is the first of two key findings of our study. Reed responded rapidly and uniformly to three annual flooding events with a temporary pulse of areal expansion in the following year. The pulses were of variable amplitude, and subsequent reductions in reed area occurred at varying rates as hydrological conditions became less favorable, primarily in marginal areas prone to desiccation [86]. The extent of reed areas peaked in 2017 and varied almost 2-fold over a 21-year interval, but in contrast to observations from Europe [87], these fluctuations did not result in a net reduction in reed area over the time interval. Thus, in spite of significant year-to-year fluctuations over a period of several decades, reed populations in the delta remained relatively stable.

Flooding is known to trigger rapid growth of reed rhizomes and belowground buds that have remained dormant due to unfavorable environmental conditions [88,89,90]. These structures can survive unfavorable periods and then form shoots and spread laterally following inundation, providing an ecological explanation for the observed rapid increases [91,92]. Plasticity of this sort has limits, because excessively deep or prolonged flooding can reduce substrate aeration and limit growth [82]. Buds and rhizomes also eventually die during periods of prolonged drought. The sequence of “flood pulse → post-flood moisture → expansion in the following season” is nonetheless consistent with the observed one-year lag and with the retreat that follows a pulse of growth conditioned by flooding.

Elevated inflow likely acts through several connected pathways: (i) expansion of shallow flooding and increased area of wet substrates, (ii) recharge of near-surface and root-zone horizons due to rising groundwater levels, (iii) reduction in salt stress in floodplain soils following flushing of salinity, and (iv) redistribution of fine sediments and nutrients [93,94]. The magnitude of the response is also likely influenced by climatic, geomorphological, and land-use factors that include weather events, reorientation of surface channels, grazing by livestock, fires, and reed harvesting [32,78].

Our data also reveal a lack of correspondence between the onset and magnitude of the flood-associated pulses of reed areal expansion in 2003, 2011–2012 and 2017 and peak years of wetland expansion, which occurred in 2000–2001, 2008–2009, 2015–2016, and 2019 [73]. Fluctuations in the delta’s reed areas, which as shown here reflect aboveground reed biomass per unit area, thus can be masked by overall fluctuations in wetland areas. This distinction, which represents the second key finding of our study, is fundamental. P. australis is a structurally dominant but ecologically selective component of the wetland mosaic, whereas aggregated wetland vegetation classes include a mixture of habitats, transitional ecotones, and life forms that differ in hydrological sensitivity and recovery time. Importantly, we used a scale of the proportion of reed in the community (from <25% to >90%) as an operational description of dominance, which makes it possible to distinguish reed-dominant areas from mixed communities. This helps explain why the lagged response is more clearly expressed here at the species/dominance level than in studies based on broader wetland categories—and why comparisons among studies require caution because of differences in class definitions and reed delineation thresholds.

4.3. Limitations to the Current Study and Future Perspectives

Although the classification approach used in the present study reliably distinguished reed-dominated and non-reed communities, some uncertainty remains in transitional ecotone zones where reed occurs in mixed communities with other wetland vegetation species. The reed-bed area calculated in this study is not a true subpixel reed fraction, but rather a reed-dominance-weighted area, i.e., an area weighted by field classes of reed dominance. The cover coefficients used here allow mixed Landsat pixels to be accounted for conservatively, but they do not replace direct subpixel modeling. Hence absolute area values should be interpreted with caution, especially for transitional communities containing reed. The main interannual trends, years of maximum expansion, and the detected lagged response were nevertheless robust to changes in cover coefficients, because they were retained in the sensitivity analysis.

The environmental and economic importance of reed in the delta provides a compelling rationale for additional direct measurements to understand the ecological basis for reed’s dominant position in the area’s wetland vegetation of reed. Our analysis captured large-scale interannual changes in reed area, but it did not include direct quantitative assessment of additional structural characteristics such as stand density, fragmentation, and within-stand heterogeneity, which may also respond substantially to hydrological changes. In addition, and although the detected time delay strongly indicates hydrological control, reed dynamics also depend on several covariates that were not examined. These include local water depth, salinity, sedimentation, flow velocity, and rhizome-system condition [93,95,96,97,98]. These factors warrant future study.

The above limitations point to several priorities for future research. Reed-stand density, fragmentation, and habitat configuration could be assessed in greater detail using satellite data with higher spatial resolution and landscape metrics, extending approaches previously applied to general wetland vegetation of the delta [40]. Long-term integration of remote sensing with repeated field surveys would also improve understanding of how changes in inundation depth, drying rate, and connectivity of the channel network affect reed productivity and community resilience. Such information is important not only for ecological interpretation but also for practical resource use such as livestock grazing and harvesting reed as fodder or construction material [32,78].

5. Conclusions

Reed dynamics in the Ili River delta are significant not only for the ecology of wetlands but also for the future economic sustainability of the Lower Ili–Balkhash region. Our analysis confirms that reed-dominated communities in the Ili River delta are highly and predictably responsive to interannual variability in river inflow. This observation is of direct relevance to management of the delta’s wetlands in an era of increasing competition for water resources. Any strategies for regulating river flow and flooding regimes will inevitably affect reed distribution, which in turn will disproportionately influence wetland stability. Given the high sensitivity of reed to fluctuations in river inflow, continuous monitoring that combines field observations and remote sensing would benefit both conservation and sustainable resource use [87,99,100,101]. Specifically, and in the context of Kazakhstan’s “green” development priorities [75,76,79], a data-based, balanced approach would reduce ongoing conflicts between conservation goals and exploitation of the delta’s resources.