Water Indices and Machine Learning Classification of Al-Razzaza Lake in Iraq: A Sustainable Perspective

In areas with little or scarce rain, water sources are being strained by climate change, a rise in population, and more human activities [1-4]. It is important to track the flow of water on the surface for proper water management, and this applies most to sensitive zones [5-7]. Combining Geographic Information Systems (GIS) and Remote Sensing (RS) technologies makes it possible to follow and examine how hydrological systems change over the years [8]. Landsat missions supply satellite images that can be used regularly to examine both water elements and the land near them [9]. Researchers in the field of sustainability have devoted themselves to developing quantitative indicators of sustainable development [10]. Normalized difference water index (NDWI) improves the outline of open water using a contrast to land and plants [11], but the normalized difference vegetation index (NDVI) gives information on variations in vegetation and ecology [12]. Using these indices jointly helps give a better picture of changes in both hydrology and the environment [13]. The modern era uses RS and GIS technologies together as efficient spatially inclusive tools to observe water surfaces and detect their changes economically [14-16]. Also, deep and machine learning models have potential significance in the classification processes [17-20]. Land cover mapping in GIS depends importantly on satellite image classification, which gives clear results on Earth’s surface features. Among the different methods of classification, Random Forest is considered a strong machine learning algorithm because it is resilient, produces reliable results, and can handle complex data easily [21].

The largest natural lake in Iraq exists between Karbala and Anbar provinces and carries the dual names of Al-Razzaza Lake and Lake Milh [22, 23]. The reservoir maintains three vital functions in the environment, which include irrigation for farming, support of biodiversity, and development of tourist sites [24]. The Euphrates River controls water flow to the lake through managed canal systems, yet both weather changes and human activities endanger its water balance [25]. The combination of lower rainfall amounts together with higher evapotranspiration rates from climate change and expanding irrigation and water usage for people represents significant threats [26, 27]. The surface water area of Al-Razzaza Lake underwent substantial modifications through decades of dynamic changes. Therefore, demanding urgent scientific investigations to understand its complete transformation [28]. The continuous observation of surface water bodies, especially Al-Razzaza Lake, enables people to understand environmental transformations. Besides, it’s significant in water resource management as well as sustainable development planning in Iraq's arid and semi-arid regions [29].

The integrative nature of GIS and RS is essential since it allows the spatial-temporal analysis of the hydrology shift on a more accurate location [30]. RS and GIS were used in research analyzing Lake Urmia in Iran to reveal the trends of drastic shrinkage relative to climate and human activities [31]. Similarly, RS-GIS applications of Lake Hamrin in Iraq demonstrated that the lake surface mapping was useful in the management of the lake during a period when rainfall reduced [26]. The examples affirm the applicability of such techniques when monitoring sensitive water bodies.

In this study, we used GIS and RS for mapping the changes of Al-Razzaza Lake in different time periods. NDWI and NDVI were combined, which made it possible to spot changes in the lake’s surface and detect any related environmental changes in the nearby land. Some reasons are excessive water usage, less rainfall, and increased damage to the environment. RS supplies foundational knowledge for developing these strategies through evidence-based decision-making processes.

The analysis used RS and GIS methods to explore the environmental dynamics in the Al-Razzaza Lake, specifically the variability of water and vegetation. The level 2 Multi-temporal Landsat 8-9 data on surface reflectance were tested to determine the change across time. Four clear clouds were picked, with the date of the acquisitions being 19 August 2022, 23 September 2023, 17 September 2024, and 19 August 2025. The selection criteria of these photos included low cloud cover 0–0.3. All datasets were handled in the WGS-84 coordinate reference system and had the same spatial resolution, which guaranteed spatial consistency of the time series. Table 1 presents the processed dataset used in this study.

Table 1. The study dataset and metadata

To reduce seasonal differences in inter-annual surveys, the majority of the images were chosen during the dry season (August-September) when vegetation cover is minimal, and water content also remains relatively constant.

Detection of temporal change was done by using time-series and spatial comparison of NDVI, NDWI, and land-cover maps, which were classified.

The resulting outputs were visualized and interpreted to determine long-term environmental trends that influence Al-Razzaza Lake.

Moreover, Figure 2 illustrates the data collection and processing workflow, including the application of corrections and subsequent software-based analysis. Preprocessing of downloaded images was done through Environment for Visualizing Images (ENVI) and ArcGIS software and involved the radiometric correction and cloud detection processes to improve the quality of the data.

Figure 2. Flowchart of the methodology conducted

Spectral indices based on the Landsat bands were used to study vegetation and water dynamics. The near-infrared and red bands were used to derive the NDVI, which was used to assess the state of vegetation, whereas the green and near-infrared bands were used to obtain an NDWI that was used to determine the extent of surface water.

These indices presented a stable performance of monitoring time variations in vegetation density and surface area of lakes.

Random Forest classification was further used to characterize land-cover conditions using multispectral Landsat bands. The Random Forest classifier, based on a combination of randomly built decision trees based on randomly chosen feature subsets, was used to provide land-cover classifications by majority voting. In this way, the discrimination between water, moist soil, salt crust, shrubs, and bare land was enhanced.

Spatial distribution of training samples was done to minimize bias and enhance classification accuracy.

The systematic pre-processing workflow was applied to all the satellite images in order to ensure radiometric and geometric consistency, followed by analysis.

After the pre-processing, the NDVI and NDWI indices were obtained based on the corrected surface reflectance images to examine the change in the vegetation and water bodies over time.

An analysis of the NDWI allowed researchers to improve surface water detection while extracting their extent. The equation that represents NDWI can be set as [10, 37]:

$N D W I=(G-N I R) /(G+N I R)$                   (1)

where, G is the Landsat green band reflectance, and NIR is the near-infrared reflectance.

The NDVI detects vegetation class through red and NIR band combinations and provides a secondary environmental indicator in semi-arid areas where vegetation depends on accessible water resources. The research incorporated NDWI together with NDVI to monitor lake water extent changes while clearly associating these impacts with regional ecological functions. The evaluation technique proves essential in central Iraq because water scarcity, desertification, and climate variability are intensifying in this region. The equation that represents NDVI can be set as:

$N D V I=(N I R-R) /(N I R+R)$                 (2)

where, R is the red band.

The Random Forest classification of Al-Razzazza Lake in 2025 was based on a single-stage, supervised, pixel-based classification. Manual digitization of the training samples was made to reflect the five classes of land cover, namely, water, wet soil, salt crust, vegetation, and bare land. In order to minimize bias on dominant classes, a balanced sampling strategy was also employed by restricting the maximum number of training samples per class to about 1000 pixels to keep minority classes, e.g., water and salt crust, adequately represented. Classification was done using spectral behaviors and field information: water was considered to have low reflectance in the near-infrared; vegetation had a high spectral near-infrared reflectance, wet soil had an intermediate spectral reflectance with moisture effects; salt crust was high-brightness; and dry land had intermediate and dry spectral reflections.

Figure 3 illustrates the applied classifier with the number of trees and samples per class.

Figure 3. Random Forest classification input samples

The classification feature set was based on multi-spectral Landsat 8-9 surface reflectance bands alone without the addition of spectral indices and texture measures to ensure model simplicity and minimize redundancy. The Random Forest model was fit with 50 decision trees, each with a maximum tree depth of 30, which was sufficiently complex yet not overly complex to overfit. The importance of the features was implicitly measured by the Random Forest internal importance, which confirmed the superiority of near-infrared and shortwave infrared bands in separating classes. The effects of spatial autocorrelation were reduced through the distribution of training samples spatially in the study area as opposed to their local clustering, which enhanced the generalization potential of the model.

Multi-temporal evaluation of the Al-Razzaza Lake in the NDVI, NDWI, and Random Forest classification demonstrated that there is a significant and consistent decrease in the surface water cover of the lake from 2022 to 2025, indicating that the lake is embedded in a larger trend of decline in the inland lakes across various regions of the world.

As observed in the NDWI analysis, the open water area progressively decreased, with 2025 having the lowest water area, a pattern which is similar to that reported in Lake Urmia in Iran [38], where NDWI-based research reported a rapid shrinkage of the lake due to low inflow, an extended drought, and intensive extraction of water. Similar NDWI decreases include those found in the Basin of the Aral Sea, where surface water loss increased after massive hydrological changes and climate fluctuations were experienced [39].

NDVI values of Al-Razzaza Lake showed that there were spatially consistent patterns of vegetation, where large values were only found in agricultural areas and in the margins of the lakebed and barren areas, and low values prevailed in the rest of the exposed lakebed.

The time-varying decrease in moderate-high NDVI regions was observed to be the alteration of the vegetation distribution with the decrease in water availability, instead of the ecological degradation. The interpretation can be compared with the results of the studies in the Dead Sea [40], where NDVI variability was attributed to the shoreline recession and the variation in soil moisture as opposed to ecosystem wellness. Similar studies on the lake shrinkage and the redistribution of vegetation are also reported in Lake Urmia, where the shrinking water tables also shifted the surrounding agricultural and natural vegetation distribution [38].

Random Forest classification offered a more detailed description of land-cover transitions than index-based methods alone.

The Random Forest map 2025 indicated disrupted bodies of water, increased wetlands, and salt crust areas along the retreating shoreline and the prevalence of bare lands. Random Forest-based counterparts of the Aral Sea and Lake Urmia showed better ability to discriminate between water, saline surfaces, and transitional wet areas as compared to single-index threshold classification. Under such circumstances, Random Forest minimized spectral confusion and improved the identification of shoreline dynamics, especially when extreme water loss is involved. The Random Forest-generated estimates of the surface areas of Al-Razzaza Lake showed a steep decrease between the estimates of around 262.98 km² in 2022 and 57.08 km² in 2025, which is a rate of decline that is comparable with the rates of rapid shrinkage that have been recorded in Lake Urmia during severe drought years [38].

The match in the trends of NDWI and the spatial patterns of NDVI, as well as the Random Forest-based estimates of the surface areas, enhances the confidence in the findings and contributes to their similarity with the existing literature on the subject of lake-shrinkage. There was no need to use spectral indices only, as the incorporation of Random Forest classification allowed distinguishing water, wet soil, salt crust, vegetation, and bare land, just like in the case of more sophisticated methods used in other degraded lake systems. On the whole, the results show that the Al-Razzaza Lake is experiencing the same level of hydrological stress that is characterized by the same magnitude and spatial manifestations as other well-known lakes that are shrinking, which highlights the existence of regional significance in the observed changes and the ability of the integrated RS and machine learning techniques to monitor the lakes over long periods.