Anzali Wetland Crisis: Unraveling the Decline of Iran's Ecological Gem

The wetland loss rate in Iran is faster than the global average. Comprehending the shrinkage rate in Iranian wetlands and identifying the underlying drivers of these changes is essential for safeguarding their ecosystems' health and services. This study proposes a novel gray‐box modeling framework to quantify the effects of climate change and anthropogenic activities on the wetlands, by combining process‐based and machine learning models. The developed model is utilized to project the Anzali coastal wetland shrinkage by simulating the complex interaction between the meteorological, hydrological, anthropogenic and sea water level characteristics, and the changes in wetland water surface area. Our framework aggregates Soil and Water Assessment Tool model, the 12 General Circulation Models of the Coupled Model Intercomparison Project Phase 6, Landsat imagery, and the Long Short‐Term Memory model to project the shrinkage of the wetland till 2100. A comprehensive range of climate and Land Use/Cover change scenarios are analyzed. The results show that wetland will seasonally desiccate in 2058, mainly due to increasing air temperature, reduction in precipitation and inflow, excessive sediment loading to the wetland, and decline in the Caspian Sea level. For optimistic scenarios, where no changes in the Caspian Sea level is considered, the wetland will gradually diminish and become a seasonal waterbody by 2100. The outcomes of this study highlight that the Anzali wetland desiccation has profound implications for the regional‐scale ecological balance, ecosystem health and function, public health, and local economy. Robust environmental interventions and sustainable development strategies are urgently needed to mitigate the detrimental impacts of climate and anthropogenic drivers on the wetland.

10.1029/2023JD039538 2 of 18 these services increases, the rate of wetland loss intensifies in the face of climate change, making wetlands one of the most threatened ecosystems on our planet (Chatterjee et al., 2015;Kingsford et al., 2016;Xi et al., 2021).The rate of wetland loss is estimated to be three times faster than that of forests (Tickner et al., 2020).The rate of wetland loss varies, with some areas experiencing more significant declines (Davidson, 2014).Iran, with six wetlands listed in the Montreux Record (Davis, 1994), ranked as the second country globally in terms of wetland loss, after Greece.Iran experienced ∼14% wetlands loss between 2010 and 2020, a trend that is likely to continue in the future (Wendling et al., 2020).This underscores the pressing need for conservation and sustainable management plans to mitigate the degradation of wetland ecosystems in Iran, and protect their valuable ecological integrity, biodiversity, and ecosystem services.
The Anzali wetland, located in the north of Iran, is hydraulically interconnected with the Caspian Sea and is susceptible to the widespread global decline of wetlands.It holds international recognition as a significant wetland, particularly for its role in supporting migratory bird populations, and maintaining the diversity of flora and fauna habitats, making it a critical ecological hotspot.The Anzali wetland holds immense importance for the local community, regarding ecological well-being and economic activities tied to the wetland's resources (JICA & Moja, 2004).The wetland has essential roles in enhancing water quality, groundwater recharge, saltwater intrusion, and serving as a buffer for flood and erosion (Aghsaei et al., 2020;JICA & Moja, 2004;Mahdian et al., 2023;Naderi & Saatsaz, 2020).Anzali wetland was recorded as a Ramsar site in June 1975 and later was added to the Montreux Record of endangered wetlands, mainly for anthropogenic activities (Davis, 1994;JICA & Moja, 2004).Long-term monitoring shows the Anzali wetland lost ∼80% of its water surface area from 1930 (∼258 km 2 ) to 1989 (∼52 km 2 ) (JICA & Moja, 2004).A similar degradation trend was observed from 1989 to 2020, indicating a concerning trajectory toward complete disappearance in the near future (Aghsaei et al., 2020;Mahdian et al., 2023).
Natural and human-made factors are contributing to the change in the Anzali wetland's water surface area (AWSA).The decline of the Caspian Sea level (−3.6 cm/yr from 1993 to 2019) has introduced severe concerns regarding the desiccation of its coastal water bodies, such as the Anzali wetland, which is hydraulically connected to the sea in the southwest end (Akbari et al., 2020;Kheirabadi et al., 2018;Modabberi et al., 2020).The alteration in water levels stands out as a primary drivers for the degradation of coastal wetlands worldwide (Anderson et al., 2023;Fracz & Chow-Fraser, 2013;Kitula et al., 2015;Kiwango & Wolanski, 2008;Wurtsbaugh et al., 2017).The historical decrease in precipitation and rise in air temperature within the Anzali wetland watershed (AWW) further exacerbate the degradation through increase in evaporation, reduction of inflow to the wetland, and increase in water demand for expanding agriculture in the AWW (Mahdian et al., 2023;Naderi & Saatsaz, 2020).The rise in temperature and the concurrent decline in precipitation have consistently been identified as two major drivers of wetland ecosystem degradation during the Anthropocene (Ashok et al., 2021;Fluet-Chouinard et al., 2023;Liu et al., 2006).Local anthropogenic activities upstream of the wetland can further exacerbate the shrinkage of AWSA via increased sediment yield, mainly due to intense deforestation and urbanization (Aghsaei et al., 2020;JICA & Moja, 2004;Mahdian et al., 2023).This phenomenon aligns with a global trend where such factors contribute significantly to wetland loss (Fluet-Chouinard et al., 2023;Xi et al., 2021).Consequently, understanding the influences of Caspian Sea level fluctuations, climate change, and local anthropogenic activities is crucial for prioritizing conservation and protection efforts for the Anzali wetland.
This study investigates the complex interactions between the underlying parameters underpinning AWSA, including changes in the Caspian Sea's level, precipitation, air temperature, inflow, and sediment load, to quantify the wetland's degradation under climate and land use/cover (LUC) change scenarios.A novel gray-box modeling framework is developed by combining different white-box tools (i.e., Soil and Water Assessment Tool-SWAT, General Circulation Models-GCMs, and Caspian Sea level) and a black-box model (i.e., Long-Short Term Memory-LSTM) to overcome the challenges associated with data needed for the boundary and initial conditions in white-box modeling, and the complexities of the coastal wetland hydrodynamics.

The Anzali Wetland
The Anzali wetland is intricately connected to the southwestern coast of the Caspian Sea (Figure 1).Internationally, the wetland is recognized as a critical habitat for migratory birds, providing a diverse ecosystem encompassing several distinct habitats, including freshwater lagoons, expansive reed beds, shallow impoundments, and seasonally flooded meadows.The ecological elements within the wetland interact intricately, providing crucial habitats for 150 species of birds (resident and migrant), 49 species of fish, 17 species of amphibians and reptiles, and 31 species of mammals (JICA & Moja, 2004).

10.1029/2023JD039538
3 of 18 The AWW covers ∼3,600 km 2 area fed by nine major rivers and the Caspian Sea (Figure 1).The prevalent climate in the AWW is the Caspian or Hyrcanian climate.Based on the long-term historical synoptic station records , annual precipitation across the AWW varies between 934 mm (south) to 1,719 mm (north).Fall and summer are the most and least humid seasons in the AWW, respectively.The temperature in the watershed is mild, with an annual average of 16.5°C.The AWW exhibits a diverse land cover composition, with forests being the dominant land cover type, followed by agricultural areas (mainly rice and orchard), residential and bare lands, pastures, wetlands, and water zones (Aghsaei et al., 2020;JICA & Moja, 2004).

Methodology
A gray-box modeling framework is proposed, linking the complex changes in the AWSA to watershed hydrological processes, meteorological drivers, climate scenarios, and the Caspian Sea level fluctuations via data from 1989 to 2020 (i.e., base period).The proposed framework projects the wetland shrinkage under the simultaneous effects of LUC and SSP1-2.6,SSP2-4.5, and SSP5-8.5 (here, the SSP stands for Shared Socio-economic Pathways) climate change scenarios (IPCC, 2022) from 2021 to 2100 (i.e., projection period).The gray box modeling framework integrates the in-situ measurements, SWAT model (Arnold et al., 1998), the 12 GCMs of the Coupled Model Intercomparison Project Phase 6 (CMIP6) (Eyring et al., 2016), and satellite-based altimetry as whitebox tools, and the machine learning-based LSTM (Hochreiter & Schmidhuber, 1997) as a black-box network to quantify the relationship between the AWSA and its dominant drivers (Figure 2).

Caspian Sea Level
Time series of seasonal and annual water levels at the Caspian Sea, covering the period from 1992 to 2020, have been obtained from Schwatke et al. (2015aSchwatke et al. ( , 2015b)), and summarized in Table S1 in Supporting Information S1.Satellite altimetry of the Caspian Sea was not measured before 1992.Hence, data from 1989 to 1992 were obtained from in-situ measurements of the Iranian National Institute for Oceanography and Atmospheric Science at the Anzali port, given in Table S1 in Supporting Information S1.
We followed Huang et al. (2021) study of the Caspian Sea level decline associated with CO 2 emission during 1993-2019.Using a coupled atmospheric-ocean model, they showed that if CO 2 concentration was doubled (i.e., 734 ppm) or quadrupled (i.e., 1,468 ppm) compared to the 1999 (i.e., 367 ppm), the measured recorded annual historical decline of −3.6 cm  in the annual Caspian Sea's water level during the 1993 to 2019 (i.e., −3.6 cm) reaches −8 and −20 cm, respectively.Utilizing this data set, we developed a linear regression model linking the annual water level in the Caspian Sea to the annual CO 2 concentration during the base period, as described by Equation 1: where, y a is the annual change in the Caspian Sea level, ρ 1 and ρ 2 denote the slope of the lines, β 1 and β 2 are y-intercepts, and x is the annual CO 2 concentration.
With the annual CO 2 emission data for SSP1-2.6,SSP2-4.5, and SSP5-8.5 obtained from Meinshausen et al. (2017aMeinshausen et al. ( , 2017bMeinshausen et al. ( , 2019aMeinshausen et al. ( , 2019b) ) (Table S2 in Supporting Information S1), and leveraging the calibrated Equation 1, the annual Caspian Sea level was projected for SSP1-2.6,SSP2-4.5, and SSP5-8.5 from 2021 to 2100.However, we need the seasonal projection of the water level for further investigation of the decline in the AWSA.For this purpose, we tuned another linear regression model which converted the annual water level to the seasonal one during the base period as Equation 2: where, y s is the seasonal Caspian Sea level, σ denotes the slope of the line, y b denotes the annual Caspian Sea level, and  is the y-intercept.

Precipitation and Air Temperature
Precipitation and air temperature are two primary drivers of the AWSA, with precipitation directly contributing water to the wetland, whereas air temperature influences evaporation-induced water loss.Changes in precipitation and air temperature influence the upstream hydrological processes, impacting inflow and sediment load into the wetland (Aghsaei et al., 2020;Mahdian et al., 2023).Historical records of precipitation and air temperature data  were obtained from the synoptic station nearest to the wetland, that is, C3 (Figure 1), available in Table S1 in Supporting Information S1.The remaining synoptic station data from Iran Meteorological Organization were tuned to the hydrological model for simulation discharge and sediment load process in the AWW.For the investigation of climate change impact on AWSA, we employed the outputs of 12 GCMs of the CMIP6 models (i.e., ACCESS-CM2, CanESM5, CMCC-ESM2, CNRM-CM6-1, FGOALS-g3, GFDL-ESM4, HadGEM3-GC31-LL, IPSL-CM6A-LR, MIROC-ES2L, MPI-ESM1-2-LR, MRI-ESM2-0, and NorESM2-LM), to produce climate variables such as precipitation and air temperature under SSP1-2.6,SSP2-4.5, and SSP5-8.5 from 2021 to 2100.The Euclidean distance and quantile delta mapping methods were used for statistical downscaling and bias correction, respectively.For performance analysis of bias-corrected GCMs relative to the observational climate variables from 1986 to 2014, correlation coefficient (CC), normalized root mean square error (NRMSE), and Kolmogorov-Smirnov statistics were utilized.Further information about downscaling, bias correction methods, and performance analyses was explained in the previous research (Mahdian et al., 2023).The precipitation and air temperature data for the projection period are obtained through the downscaled and bias-corrected outputs of twelve GCMs-CMIP6 models in station C3, given in Tables S3-S20 in Supporting Information S1.

Inflow and Cumulative Sediment Load
Other essential drivers of the AWSA are inflow and sediment load to the wetland through the upstream rivers.
Water and sediment yield in the AWW has recently experienced significant changes because of natural and anthropogenic activities such as deforestation, urbanization, and agricultural expansion (Aghsaei et al., 2020;Mahdian et al., 2023).In this study, we utilized monthly data derived from the robust SWAT model in our previous study (Mahdian et al., 2023), which resulted in a good accuracy based on the Nash-Sutcliffe efficiency (>0.5) and coefficient of determination (>0.5) indexes.This data set includes the sum of the discharge from eight rivers to the wetland and the corresponding aggregated cumulative sediment load.These data were employed to simulate hydrological processes in the AWW, considering the combined effects of climate and LUC changes.
The AWW was divided into 70 sub-basins within the fine-tuned SWAT model, and these sub-basins were further subdivided into 1679 smaller units known as Hydrologic Response Units (HRUs).These HRUs were categorized based on the eight predominant LUC types (i.e., forest, rice paddies, orchard, urban, bare land, wetland, pasture, and water), four soil types as defined by FAO (FAO, 1995), and watershed slope characteristics (i.e., 0%-15%, 15%-30%, and 30%-99%).Discharge and sediment yield computations were performed for each HRU, and the results were subsequently routed through the stream network.The fundamental principle of the SWAT model revolves around the water balance equation, as shown in Equation 3: where SW t denotes the final soil water content, SW 0 is the initial soil water content, and t denotes time.Also, R day , Q surf , E a , and Q gw are the daily amounts of precipitation, surface run-off, evapo-transpiration, percolation and bypass flow exiting in the soil profile bottom, and the return flow, respectively.
The SWAT estimation of sediment yield transported via surface run-off is accomplished by using the modified universal soil loss equation: where Sed is the sediment yield on a given day, Q surf is the surface runoff volume, q peak is the peak runoff rate, area hru is the area of the HRU, K USLE is the USLE soil erodibility factor, C USLE is the USLE cover and management factor, P USLE is the USLE support practice factor, LS USLE is the USLE topographic factor, and CFRG is the coarse fragment factor.
Observed climate data, along with downscaled and bias-corrected outputs of 12 GCMs under climate change scenarios (SSP1-2.6,SSP2-4.5, and SSP5-8.5)and LUC projection are utilized as inputs to the SWAT model to simulate the inflow and sediment load to the wetland during both base (1989-2020) and projection (2021-2100) periods (see details in Tables S3-S20 in Supporting Information S1).

Historical Records of the AWSA
In the absence of direct measurements of the AWSA, satellite imagery from Landsat 5, 7, and 8 (Level 2, Collection 2, and Tier 1) obtained through the Google Earth Engine (GEE) platform was utilized to generate seasonal water surface area in the wetland during the base period.Given the region's characteristics, including numerous built-up areas, high cloud cover particularly during rainy season, and low albedo surfaces, application of conventional indices such as NDWI and MNDWI for water surface area detection is limited and challenging (Frey et al., 2010;Verpoorter et al., 2012;Xu, 2006).Consequently, the automatic water detection index (AWEI shadow ) (Feyisa et al., 2014) was employed for delineating multi-temporal water pixels from other forms of LUC in the study area.This method is robust for water detection in complex aquatic ecosystems such as wetlands, addresses challenges posed by the presence of shadow and low albedo, which may hinder other methods such as NDWI and MNDWI (Doña et al., 2021;Laonamsai et al., 2023;Nguyen et al., 2019;Yulianto et al., 2022;Zhai et al., 2015).
To enhance the robustness of the derived AWSA, Otsu dynamic thresholding (Otsu, 1979) coupled with a Canny edge filter was applied, overcoming the challenges of dynamic thresholding and constraining the number of input pixels for those located near water bodies (Donchyts et al., 2016).This method was applied to the images with less than 10% of cloud-cover with the Canny's threshold of 0.7 and a sigma of 1.For images with over 10% cloud-cover, the long-term seasonal average value was adopted to fill the gaps (Aissia et al., 2017;Sharma et al., 2021).
To evaluate the accuracy of the AWEI shadow method in extracting 118 AWSA maps during the study period, a stratified random sampling approach (Cochran, 1946;Neyman, 1938Neyman, , 1992) ) was applied.First, the pixels were divided into two strata, that is, water and non-water.Subsequently, a total of 100 points were randomly selected from Google Earth as true data for each extracted AWSA maps, totaling 11,800 points for all extracted maps.These 100 points were allocated based on the proportion of water and non-water strata.To better highlight the performance of the AWEI shadow , half of the water stratum points were selected from the boundary regions inside the water area.Finally, these true points taken from water and non-water strata were compared with those assigned by AWEI shadow to determine the accuracy of the method in extracting satellite-based AWSA maps, using the Kappa coefficient and overall accuracy.Both the Kappa coefficient and overall accuracy optimum value were determined at 1, which is above 0.8 indicating the excellent performance of the method (Viera & Garrett, 2005).

Black-Box Network
While the white-box tools used in the study are reliable, they have limitations in simultaneously considering the complex interactions of the AWSA with hydro-meteorological processes, anthropogenic changes in LUC, variations in the wetland boundary, and fluctuations in the Caspian Sea's level.Therefore, LSTM model was employed as a black-box network to robustly establish connection between the AWSA and its primary drivers.The primary rationale for utilizing a sophisticated model such as LSTM in this study relates to its aptitude for predicting time series data (Hochreiter & Schmidhuber, 1997).As depicted in Figure 2, first, the time-series into supervised step attributes past data to the current data, with a three-step strategy is employed, incorporating both input data and AWSA from the three preceding seasons are incorporated into the model.Subsequently, the LSTM model incorporates a unit that transfers pertinent information from past observations to the model's current state, as illustrated in Figure 2. Furthermore, to attain optimal model performance, the hyperparameters of the LSTM model, comprising batch size, dropout, learning rate, neuron, and activation function, are concurrently tuned utilizing the Bayesian optimization algorithm (Khosravi et al., 2023).The white-box tools were employed as sub-models to generate the main drivers of the AWSA for the base and projection periods (Figure 2).Here, the LSTM model enhances the projection of AWSA from 2021 to 2100 by establishing relationship between the AWSA and its main drivers. 10.1029/2023JD039538 7 of 18 Another key benefit of the black box model is its capacity not only to forecast the future but also to estimate the influence of variables (Noori et al., 2012).
In order to accomplish this objective, we employ the permutation feature importance method, which involves shuffling each variable individually for ranking of variables based on their impact on the model's generalizability (Altmann et al., 2010).This approach affords us valuable insights into the parameters that exert an influence on AWSA.

Wetland Shrinkage Scenarios
A total of 18 scenarios were designed to assess the future shrinkage of the Anzali wetland.Two main categories were considered: (a) nine sub-scenarios that accounted for change in the Caspian Sea level (real scenarios), (b) nine sub-scenarios assuming no change in the sea level (conservative scenarios).
Nine sub-scenarios were carefully defined to encompass extreme conditions related to both climatic and human-induced drivers of the AWSA under different SSPs.A wide range of potential future scenarios are captured, allowing for a comprehensive assessment of the wetland's response to various combinations of climatic and human influences.The first three sub-scenarios involve selecting the outputs from a GCM that exhibits the highest average precipitation increase under SSP scenarios investigated and LUC dynamics.The fourth to sixth sub-scenarios utilized the outputs from a GCM that projected the lowest average precipitation decrease and static LUC.The seventh to ninth scenarios were defined based on the mean ensemble of projections from all 12 GCMs under the SSP scenarios, considering static LUC (Table 1).To visualize the uncertainties in the projection of the AWSA, the results of the various scenarios were presented in six bands as shown in Table 1.These bands primarily associated with changes in the Caspian Sea level and further differentiated by other factors such as LUC, GCMs, and SSPs.The Mann-Kendall test (Kendall, 1975;Mann, 1945) and Sen slope estimator (M-S) (Sen, 1968), as two widely used trend detector methods in hydroclimatology studies (Noori, Bateni, et al., 2022;Noori, Woolway, et al., 2022), were employed to identify statistically significant univariate trends in the AWSA for the base and projection periods.

Changes in the Caspian Sea Level
We conducted a thorough tuning of the relationship between the annual Caspian Sea level and annual CO 2 emissions during the base period, achieving a commendable CC of 0.94.The tuned model was then employed to project the annual Caspian Sea level from 2021 to 2100, as shown in Figure 3. Furthermore, our results demonstrate the successful tuning of Equation 2 for conversion of annual water level to the seasonal dimension, with high CC values of 0.98, 0.99, 0.98, and 0.98 for Spring, Summer, Fall, and Winter, respectively.Subsequently, the seasonal water level was projected for SSP1-2.6,SSP2-4.5, and SSP5-8.5 from 2021 to 2100 (Tables S3-S20 in Supporting Information S1).
Measurements highlight a statistically significant decline in the Caspian Sea level during the base period, with an annual rate of −4.4 cm (Figure 3).This decline can be attributed to increased evaporation and decreased precipitation and inflows (Modabberi et al., 2020;Mozafari et al., 2023).The projected data (raw data are given in Tables S3-S20 in Supporting Information S1) shows continuation of this decline with rates varying depending on the CO 2 emission scenarios, that is, SSP1-2.6 (−4.2 cm/yr), SSP2-4.5 (−5.3 cm/yr), and SSP5-8.5 (−7.2 cm/ yr) (Figure 3).Compared with the base period, the Caspian Sea level will faster decrease below mean sea level (MSL) by 2100 in both SSP2-4.5 (∼31.8 m) and SSP5-8.5 (∼33.4 m) scenarios.Optimistically, if the water level decline rate during the base period remains unchanged, our projections suggest that the Caspian Sea level will reduce to ∼31.0 m below the MSL by 2100.Our findings are consistent with the existing research that projected 4-10 m decrease in the Caspian sea level by the end of the 21st century (Huang et al., 2021;Renssen et al., 2007).

Changes in Precipitation and Air Temperature
The statistical-downscaled and bias-corrected results of the 12 GCMs showed that the CNRM-CM6-1 model projected the highest increasing trend in average precipitation.In contrast, the MRIESM2-0 model projected the highest decreasing trend, compared to other GCMs, for all three SSPs in the C3 station.Hence, CNRM-CM6-1, MRIESM2-0, and the mean ensemble of the 12 GCMs were selected for further investigation of AWSA under climate change scenarios.Figure 4 shows the distribution of precipitation and air temperature observed during the base period and those projected using the selected GCMs (i.e., CNRM-CM6-1, MRIESM2-0, and the mean ensemble of the 12 GCMs) for different SSP scenarios.
Relative to the base period, the projected mean seasonal precipitation in the Anzali wetland is expected to decrease in most scenarios, except for SC1-SC4 and SC7-SC10 scenarios.In these both scenarios, extreme precipitation will occur more frequently in the future under the climate change effect, resulting the daily mean precipitation in the C3 station becomes greater than its median value (Figure 4).However, the most decrease in mean precipitation is observed in the SC2-SC5 (−18%), SC8-SC11 (−21%), and SC14-SC17 (−23%) scenarios.On the other hand, in all climate change scenarios, air temperature increases in the C3 station relative to the base period.For example, in scenarios built from SSP1-2.6,SSP2-4.5, and SSP5-8.5 (see, Table 1), air temperature increases between 1.2-1.8°C,1.5-2.2°C,and 2.6-3.9°C,respectively.As shown in Figure 4, frequency of occurrence of the higher air temperature, relative to the base period, can be observed in the high quantiles, mainly due to global warming, particularly in the scenarios built from SSP2-4.5 and SSP5-8.5.These changes will growth in future, especially in far-future, when shrinkage in the AWSA is projected to increase.Additionally, decrease in precipitation will likely increase the actual evaporation rate from the wetland.The decrease in precipitation can be associated with less cloudy days which can lead to enhanced solar radiation and subsequently increased evaporation.The combination of increased air temperature and decreased precipitation in the Anzali wetland can potentially lead to wetland shrinkage through increased evaporation, reduced inflow, declined in direct water flux, and higher water demand due to the climate change impacts and growing population (Mahdian et al., 2023;Naderi & Saatsaz, 2020).It should be noted that the CC (>0.5),NRMSE (<0.85), and Kolmogorov-Smirnov test (<0.2) indicated satisfactory performance for both downscaled and bias-corrected precipitation and air temperature data derived from CNRM-CM6-1, MRIESM2-0, and the mean ensemble GCMs, compared to those measured during the base period in C3 station.Further information about statistical-downscaled and bias-corrected model performance is described in Mahdian et al. (2023).and projection period (2021-2100)."SC" stands for scenarios described in Table 1.The circle and square shapes in the violins represented the median and mean of the data.

Changes in Sediment Load and Inflow
Distribution of inflow and sediment load projected using the selected GCMs (i.e., CNRM-CM6-1, MRIESM2-0, and the mean ensemble of the 12 GCMs) for different SSP scenarios is illustrated in Figure 5. Compared to the base period, almost all scenarios (except SC2-SC5, SC8-SC11, and SC14-SC17) show an increase in the frequency of occurrence of sediment load and inflow to the wetland at the 50% of high quantiles.For example, seasonal inflow and sediment load to the wetland will increase up to 36% and 21% in SC1-SC4, up to 26% and 14% in SC7-SC10, and up to 26% and 13% in SC13-SC16.Such changes can be contributed to the increase in intensity and frequency of extreme climatic events which mainly impacted by the extensive changes in LUC (mainly, an increase up to 18% in deforestation and urbanization).Cumulative sediment load into the wetland will reach 33, 31, and 30 million tons in SC1-SC4, SC7-SC10, and SC13-SC16 scenarios characterized by the dynamic LUC by 2100.In other scenarios, characterized by the static LUC, the mean seasonal inflow will decrease relative to the base period, except for the SC3-SC6 scenario influenced by extreme climatic events, as discussed by Mahdian et al. (2023).The projected results indicate a decrease in the cumulative sediment load under the most static LUC scenarios (SC2-SC5, SC8-SC11, SC14-SC17, and SC15-SC18) and an increase in the remaining scenarios (SC3-SC6 and SC9-SC12) (Figure 5), given that sediment load is more sensitive to the extreme climatic events (Darabi et al., 2022;Mahdian et al., 2023).On the other hand, in SC2-SC5, SC8-SC11, and SC14-SC17 scenarios, precipitation and air temperature in the summer season will decrease and increase, respectively, relative to the base period.These changes have decreased low inflows (i.e., 50% of low quantiles) to the wetland in SC2-SC5 (−22%), SC8-SC11 (−18%), and SC14-SC17 (−29%) scenarios.Similarly, in the summer, sediment load to the wetland decreases by approximately 18%, 20%, and 26% for SC2-SC5, SC8-SC11, and SC14-SC17 scenarios, respectively, at the 50% low quantiles (Figure 5).These projected changes have detrimental impacts on the ASWA, given that the upstream rivers supply fresh water for the Anzali wetland and their inflow reduction and large sediment loads play a key role in shrinkage of the wetland and its ecosystem health (Aghsaei et al., 2020;Aradpour et al., 2021;JICA & Moja, 2004).

Historical Record of the Water Surface Area
The time series of the AWSA extracted through Landsat imagery using the AWEI shadow index (raw data are given in Table S1 in Supporting Information S1), demonstrated a good agreement with the randomly selected points taken from Google Earth (Congalton & Green, 2019), as indicated by Kappa (Fitzgerald & Lees, 1994) and overall accuracy indices (Table 2).Even with half of the points of water strata located at the water's edge, the AWEI shadow index demonstrated excellent agreement with the AWSA from Landsat imagery and Google Earth points (average of Kappa and overall accuracy >0.8) (Viera & Garrett, 2005).Despite the robust accuracy of AWEI shadow , certain challenges may impact its accuracy.The index is susceptible to high cloud cover in the images, a prevalent issue in our study area, especially during the rainy season.Additionally, the relatively coarse spatial resolution of Landsat images (30 m) may introduce some uncertainty or error, particularly in detecting small or narrow (non)water areas.However, the high values of Kappa and overall accuracy (see, Table 2) indicate the minimal impacts of the such uncertainties/errors in our calculations.
The wetland has the lowest and highest surface area in warm (summer) and cold (winter) seasons, respectively.This is due to the air temperature variation and agricultural activities in the AWW.Summer is the peak season for cultivation in the AWW, with a maximum agricultural water demand (Table 2).However, the AWSA has significantly decreased from 1989 to 2020 (−0.75 km 2 /yr, p-value <0.05), resulted in the shrinkage of the wetland to 25.9 km 2 , less than the half of its initial area in 1989.The decline in the Caspian Sea level may contribute to the shrinkage of the wetland.From 1995 to 2020, the Caspian Sea experienced a significant water level drop of approximately −1.5 m, resulting in a weak connection with the Anzali wetland, as the wetland's elevation is now higher than the sea level (JICA & Moja, 2004).Intense deforestation, urbanization, and expansion of agricultural activities in the AWW have led to the introduction of a large sediment load to the wetland (Aghsaei et al., 2020;Mahdian et al., 2023).SWAT simulations show ∼7.6 million tons of sediment have entered the wetland from 1989 to 2020, further reducing the mean depth of wetland and shrinkage of AWSA (Mahdian et al., 2023).In situ measurements of water depths taken in 2003 (eight stations during six months) and 2011 (eight stations during 21 months) indicate a decrease in the wetland's mean depth from 1.7 to 1.1 m.This decline can be attributed to both the decline in the Caspian Sea level and the introduction of a large sediment load to the wetland (JICA & Moja, 2004).
Our findings suggest that the model achieves a lower NMAE by using a lower value for the learning rate, batch size, and units, as well as a higher value for the dropout.Changing the activation function did not have an impact on the error.The model's generalization and accuracy, determined by the validation data set (20% of the data set), were also satisfactory (R 2 = 0.71, NRMSE = 0.15).The model effectively captures the majority of the observed values and accurately follows the changing trends (Figure 6b).
A permutation feature importance analysis is also conducted using the LSTM model to determine the relative importance of the main drivers of the wetland shrinkage.The most significant driver of wetland shrinkage is air temperature, followed by Caspian Sea level, cumulative sediment load, precipitation, and inflow to the wetland (Figure 6c).Our finding is somewhat inconsistent with the Japan International Cooperation Agency's study (JICA & Moja, 2004), which reported the Caspian Sea level is the most important driver of the Anzali wetland shrinkage.Increase in air temperature not only affects the wetland shrinkage, but also dominantly declines the Caspian Sea level and increases the water demand for anthropogenic activities in the AWW because of the sub-tropical area the wetland is located, making the wetland sensitive to the global warming; consequently, inducing air temperature as the most important driver of the wetland shrinkage.

Wetland Shrinkage Projection
Our results indicate that the Anzali wetland will undergo significant shrinkage under all 18 projected scenarios (p-value <0.05), with varying shrinkage rates depending on the scenario.The annual wetland shrinkage rate varies from −0.25 km 2 (in SC17) to −0.47 km 2 (in SC3) in the conservative scenarios, where the Caspian Sea level remains unchanged.The wetland shrinkage accelerates when the changes in the Caspian Sea level are considered, with the annual shrinkage rates ranging from −0.61 km 2 (in SC2) to −0.77 km 2 (in SC13).The Caspian Sea level plays a crucial role in shaping the dynamics of the wetland and directly impacts its water surface area.The projected decrease in AWSA will occur at a slower or equal rate, depending on the scenario implemented, compared to the observed wetland shrinkage rates during the base period.This is attributed to two main factors: (a) the vulnerable parts of the wetland located near the Caspian Sea experienced drying during the base period, while the remaining parts are more resistant to desiccation; (b) as the Anzali wetland shrinks in the future, it will become more saline than in the base period, resulting in a decrease in the evaporation rate from the wetland surface (Al-Shammiri, 2002).
To enhance the projection under climate change and anthropogenic activities, we presented possible scenarios of the wetland desiccation using six uncertainty bands (Figure 7).Bands #1, #3, and #5 represent the projected uncertainties that take into account the changes in the Caspian Sea level, whilst bands #2, #4, and #6 depict the projected uncertainties with no change in the sea water level.The AWSA is projected to reach zero for the first time in different periods depending on the scenarios considered (band #1 = spring 2058, band #3 = summer 2062, band #5 = fall 2059).These results strongly suggest that the decline of the Caspian Sea level is a significant driver for the desiccation of the Anzali wetland and will adversely influence the coastal area of the Caspian Sea, a finding that is in line with the reported results by Akbari et al. (2020).This decline can be associated with the growth of hydraulic conductivity from the wetland to the Sea.In addition, the results highlight that in increasing precipitation scenarios (i.e., SC1-SC4, SC7-SC10, and SC13-SC16); the water surface area of the Anzali wetland is desiccated faster compared to other studied scenarios.In the scenarios with increased precipitation, the sediment load into the wetland is also increased compared to other scenarios, facilitating the wetlands' volume to get filled with sediment, provoking an accelerated declining trend in the surface water area of the wetland.These findings indicate that the Anzali wetland will undergo a significant transformation, transitioning from being a permanent wetland to a seasonal waterbody.This would severely affect the rich biodiversity of the Anzali wetland.When the Anzali wetland transforms into a seasonal water body, many species reliant on wetlands for breeding, nesting, feeding, or migration may face population declines or local extinctions, leading to a complete ecological collapse and eventually the death of the wetland (JICA & Moja, 2004).Additionally, as the Anzali wetland stores a significant amount of carbon, its desiccation could release large quantity of greenhouse gasses into the atmosphere, exacerbating climate change (Zou et al., 2022).
Despite the real-based uncertainties, the wetland will not completely desiccate (see band #4, Figure 7).Conservative-based uncertainties (bands #2 and #6, Figure 7) suggest its disappearance could be delayed until the 2090s (spring 2093 and summer 2098).The maximum rate of wetland shrinkage in conservative-based uncertainties is observed in band #2, where the results of SC3 representing the lower band of uncertainty.In SC3 scenario, the SWAT model considered dynamic LUC, simulating intense deforestation and urbanization, intensifying sediment load to the wetland.In bands #4 and #6, the increase in sediment load is lower compared to the band #2, indicating that other drivers of AWSA including, decreased streamflow, precipitation, and increased air temperature, play a dominant role in the wetland shrinkage.Reduced inflow due to decreased precipitation and increased air temperature will lead to a significant decrease in freshwater flux to the Anzali wetland.The reduction in inflow and precipitation accompanied by an increase in air temperature will intensify evaporation, speeding the wetland shrinkage.Notably, the projected sediment load into the wetland is higher in band #4 compared to band #6, resulting in a faster shrinkage of the wetland in the former scenario.

The Way Forward
Overall, the results obtained from this study (see, Figure 7) show that the desiccation process of Anzali wetland will be severe over the coming decades.In all scenarios, the Anzali wetland will transit a seasonal wetland.These findings underscore that existing management practices in the AWW are not sustainable, and an augmentation in precipitation and discharge into the wetland cannot avert its shrinkage because the mentioned changes are accompanied by an increase in sediment loads to the wetland and increase in air temperature.Notably, neither scenarios exhibited a substantial positive trend in the AWSA, even with increased discharge and precipitation.This declining trend is attributed to the more influential role played by factors such as sediment load, Caspian Sea level, and air temperature, as determined by feature importance analysis (see Figure 6c).Intense deforestation, urbanization, and agricultural expansion have led to increase in erodibility of the AWW, intensifying sediment load to the wetland.To be more specific, forests with deep roots and canopies can stabilize the soil with their deep roots and reduce the shear force of rainfall and runoff, especially in steep-slope fields.Replacing such forests with shallow-rooted vegetation classes such as pasture and agricultural lands increases sediment yield in the AWW.Consequently, the cumulative sediment load to the Anzali wetland will increase, leading to reduction in the wetland's water holding capacity and exacerbating its shrinkage (Aghsaei et al., 2020;Mahdian et al., 2023).In scenarios where the Caspian Sea level remains constant, and the average sediment load to the Anzali wetland decreases compared to the historical period, the diminishing trend in AWSA persists due to an increase in air temperature, leading to elevated evaporation rates.This finding aligns with the existing research indicating that an increase in temperature, even with increased precipitation, results in a loss of wetland water surface area (Xi et al., 2021).
The continuous reduction in the wetland's water surface will introduce multi-faceted detrimental effects for biodiversity, food chain, local, and regional economy, microclimate and the public health.The shrinkage of the Anzali wetland can severely impact agriculture that is one of the key economic activities in the region and crucial to food security for the country.Hence, comprehensive management and conservation strategies must be considered to protect the Anzali wetland from the climate change induced and anthropogenic shrinkage.To be more specific, these conservative strategies, should simultaneously consider the Caspian Sea level's decline, sediment load proliferation, and the reduction of streamflow to save the Anzali wetland from degradation and desiccation.As such, the anthropogenic activities (i.e., deforestation, urbanization, and agricultural expansion) in the upstream of the Anzali wetland's watershed should be controlled and follow a more effective approach to reduce the sediment load through afforestation and increase streamflow via introducing sustainable agricultural practices.In addition, to save the Anzali wetland ecosystem, control of the local anthropogenic activities should coincide with the governance of the Caspian Sea to ensure water level decline in the sea will be managed and mitigated effectively.

Conclusions
This study developed and successfully validated a robust gray-box modeling approach for assessing the combined effects of climate change and LUC on the changes in the water surface levels and shrinkage of wetlands.
The proposed modeling tool uses the capabilities of white-box models (e.g., SWAT) and advanced machine learning-based black-box models (i.e., LSTM) to simulate the effects of hydrological, fluvial hydraulics, climatic, and LUC parameters on wetlands' water surface area and fate.The proposed model was tested for a case study of a coastal wetland located in the north of Iran (i.e., Anzali wetland).In this regard, the Anzali wetland shrinkage under comprehensive range of climate and LUC change scenarios was investigated considering the complex interacting meteorological, hydrological, and hydraulic processes, and the interconnections between the Caspian Sea and the AWSA.18 scenarios based on LUC, GCMs and fluctuation in the Caspian Sea level were investigated

Figure 1 .
Figure 1.Location of the Anzali wetland, synoptic and hydrometric stations, and the main land use/cover types in its watershed.

Figure 2 .
Figure 2. The database preparation and modeling process for the Anzali wetland water surface area projection.(a) Preparing the training data sets from observations, drivers of the Soil and Water Assessment Tool model and outputs from Google Earth Engine.(b) Preparing the wetland's historical data for a memory-based artificial neural network.(c) Long-short term memory architecture and data flow through its forget gate (g1), input gate (g2), block input (g3), output gate (g4), and its units: input x(t), cell state C(t), output y(t).(d) Preparing the projection data sets from SWAT's inputs and outputs from the 12 General Circulation Models of the Coupled Model Intercomparison Project Phase 6.

Figure 4 .
Figure 4. Violin plots of daily average precipitation and air temperature during both the base period (1989-2020) and projection period (2021-2100)."SC" stands for scenarios described in Table1.The circle and square shapes in the violins represented the median and mean of the data.

Figure 6 .
Figure 6.Predictive results obtained from the Long-Short Term Memory (LSTM) model for the Anzali wetland water surface area (AWSA).(a) Learning curves: train and test losses, (b) observed and predicted values of the AWSA using the LSTM model during the training and validation steps, and (c) results of the permutation feature importance analysis performed by the LSTM model.

Figure 7 .
Figure 7. Projection of the Anzali wetland water surface area under 18 scenarios and 6 bands of climate change and anthropogenic activities.The raw data are available in Tables S3-S20 in Supporting Information S1.

Table 1 Scenarios
Defined Based on Land Use/Cover (LUC), General Circulation Models (GCMs), Climate Change Emission Scenarios, and the Caspian Sea Level

Table 2
Summary of the Statistical Performance for Annual and Seasonal Water Area in the Anzali Wetland