The implementation of global initiatives to mitigate climate change, promote sustainable development, and safeguard ecosystems and biodiversity is closely linked to United Nations Sustainable Development Goal 13 (Climate action). The term "land cover classification," as defined by the United Nations System of Environmental-Economic Accounting (UN-SEEA), refers to the "physical and biological characteristics of the Earth's surface, encompassing natural vegetation and non-living surfaces (U.N., 2019). This is particularly evident in the creation of maps showing changes in land use and cover (LULC) at the global, national, and regional levels (Acar & Zengi̇N, 2023). Land use is the planned application of land management strategies by human agents to utilize land cover. It represents human activities such as industrial zones, residential zones, agricultural fields, grazing, logging, and mining, among others ((Teshager & Abeje, 2021); (Fatima & Javed, 2021)). In contrast, land cover refers to the characteristics of the earth's surface that are represented in the distribution of vegetation, water, ice, desert, and topography, as well as the immediate subsurface, which includes biota, soil, topography, surface, and groundwater (Nedd et al., 2021).

Furthermore, baseline data on planning, management, and sustainable resource use are derived from our understanding of temporal variations in LULC ((Gebeyehu et al., 2019); (Paudel et al., 2016); (Wulansari, 2017)). As a result, to analyze environmental processes and issues and to maintain or improve living standards, land-use data is required. It is particularly crucial to monitor LULC changes frequently in rapidly expanding regions, as unplanned, irregular urban population growth can alter urban climates (Dhakal et al., 2022). Thus, it is critical to understand the patterns and trends in LULC change at the regional, local, and global scales. Preserving the environment while boosting economic and social advantages is the greatest challenge (Accuracy Assessment of Land Use Change Analysis Using Google Earth in Sadar Watershed Mojokerto Regency, 2022). Land classification research is necessary to manage natural resources and environmental issues and to assess the current state (Aslanov et al., 2023). The natural science community has extensively used satellite images to assess changes in land cover and terrestrial land use at both qualitative and quantitative levels (Salman et al., 2020). Novel insights into the field of large-scale LULC mapping have been made possible by the worldwide consolidation and development of cloud computing platforms, artificial intelligence, machine learning, deep learning, and deep transfer learning; moreover, time series-based techniques and remotely sensed data ((Kwan et al., 2020); (Sefrin et al., 2020)). Seasonal and phenological properties of different LULC classes can be captured by integrating a range of features and spectral-temporal metrics derived from satellite image time series analysis (Htitiou et al., 2021). Classification accuracy will increase when LULC classes are mapped using these characteristics and metrics (Luo et al., 2022).

Geographic information system (GIS) and Remote sensing (RS) communities have long been interested in accurate and current LULC mapping, primarily because it provides important information for understanding human-environment interactions (Nasiri et al., 2022). Compared with traditional surveys, using RS to monitor LULC offers several benefits, including the ability to quickly and accurately create an inventory of broad regions ((Deb & Nathr, 2012); (Juliev et al., 2019)). On a global scale, researchers used GIS and RS to analyze, classify, and assess LULC change dynamics in Sentinel-2 imagery. For example, (Accuracy Assessment of Land Use Change Analysis Using Google Earth in Sadar Watershed Mojokerto Regency, 2022) assessed the accuracy of LULC change analysis using Google Earth in the Sadar Watershed in Indonesia. (Zaabar et al., 2023) compared Sentinel-2 and Landsat images for LULC classification in an object-based image analysis (OBIA) framework using the Random Forest (RF) and Support Vector Machine (SVM) methods in the Allala watershed of Algeria. The results showed that Sentinel-2 images processed with the RF method achieved higher accuracy than Landsat satellite images. (S. et al., 2022) studied the E-Beheira governorate in Egypt using principal component analysis and supervised classification of Sentinel-2 images to enhance LULC classification accuracy. (Teshager & Abeje, 2021) conducted a LULC change-detection analysis of the Kility Watershed in Ethiopia using Landsat and Sentinel-2 images from 1986 to 2019, employing the maximum likelihood algorithm for supervised classification. Additionally, the accuracy assessment and confusion matrix analysis were conducted to assess the reliability of the LULC analysis. Aimed to evaluate the potential of Sentinel-2 and Landsat-8 images' spectral-temporal metrics by using the Google Earth Engine (GEE) cloud computing to improve the accuracy of the LULC maps, (Pande, 2022) created a new machine learning algorithm in JavaScript for GEE to classify the LULC map and change detection using Sentinel-2 and Landsat-5 images with a 5-year time difference. The researchers noted that the GEE's ability to perform other analyses on the GEE cloud computing platform is very high.

Remarkable strides have been made in RS technology in recent years to adapt to surface changes on Earth. Sentinel-2 multispectral products from the European Space Agency (ESA) and the European Union (EU) are a major contribution to the Copernicus Program advances, which are used to monitor changes in the Earth's surface (Kumari & Karthikeyan, 2023). Sentinel-2 products are suitable sources for time-series feature extraction because they offer high temporal resolution, short revisit times, and a rich spectral configuration among medium-resolution satellite imagery (Liu et al., 2020). Two Sentinel-2 images (10 m spatial resolution) from 2017 and 2022 were used for LULC mapping and to analyze and detect degraded areas. The region of change between two images of the same scene taken at different times can be detected using change detection. LULC change detection is essential for assessing transitions between land classes. Due to its free and open-access approach, Sentinel-2 data have drawn significant interest from low-income countries, where funding for remotely sensed data acquisition is limited.

Based on the review of existing studies ((Gebeyehu et al., 2019); (Juliev et al., 2019); (Pande et al., 2021); (Accuracy Assessment of Land Use Change Analysis Using Google Earth in Sadar Watershed Mojokerto Regency, 2022)), most LULC analyses have relied on conventional supervised classification approaches, particularly maximum likelihood and object-based methods, applied to relatively small study areas with limited training samples. While these methods are widely used, their reliance on sparse, locally constrained samples often limits classification robustness and generalizability, especially in heterogeneous and degraded landscapes. Limited sample sizes tend to underrepresent spectral variability, leading to uncertainty in accurately delineating complex classes such as degraded bare land. In contrast, recent advances in AI-based classification suggest that large-scale, sample-rich models can substantially improve LULC mapping accuracy by better capturing spatiotemporal dynamics. However, such approaches remain underexplored in arid and semi-arid regions using dense Sentinel-2 time-series data, highlighting a critical methodological gap that this study aims to address.

Therefore, this study focused on a large-scale, sample-based AI land classification model to map the spatial distribution of degraded bare lands using Sentinel-2 imagery. Additionally, the current studies will help monitor changes in land classification over relevant periods. The specific objectives of this study are i) to map the LULC changes from 2017 to 2022, using Sentinel-2 LULC time series (Living Atlas) data, and ii) to use the statistical analysis to determine the accuracy of the research.

Using AI-based classification algorithms for Sentinel-2 imagery, the study provides a thorough evaluation of land-use and land-cover (LULC) changes across the Shimbay district from 2017 to 2022. Sentinel-2 high-resolution 10x10 m (Delwart, 2015) LULC maps and quantitative summaries for all major land cover types, including built-up areas not previously mapped in the literature, were produced district-wide. The overall accuracy of the AI-based classification method exceeded 85% (Karra et al., 2021), demonstrating the efficacy of automated classification models in generating dependable and geographically consistent LULC datasets. When applying high-resolution LULC products to region-specific studies, careful map selection and local validation are crucial. (Xu et al., 2024) conducted a thorough comparative validation of the 10 most recent 10 m global land cover maps, revealing significant variability in class-wise accuracy and regional performance. (Li et al., 2024) demonstrated enhanced generalization and robustness across multiple datasets by introducing a novel learning paradigm for foundation-model-based remote sensing change detection. This highlights the growing potential of large-scale AI models to capture intricate spatiotemporal land-surface dynamics relevant to this study. Also, (Aziz et al., 2024) showed that machine learning-based classification of remote sensing data can efficiently map forest cover with high accuracy. This supports the applicability of the AI-based methods used in this study and highlights the suitability of data-driven approaches for detailed land-cover characterization. Using remote sensing data, (Zafar et al., 2024) assessed the effectiveness of several machine learning algorithms for LULC mapping. They found that algorithm selection significantly affects classification accuracy, underscoring the need for careful model selection and validation in AI-based LULC mapping studies such as the current work. Another researcher, (Peng et al., 2024), conducted a comparative analysis of pre-trained CNN-based deep learning models for crop monitoring using remote-sensing LULC and land-change data. Also, (He et al., 2024) demonstrated the efficacy of a deep learning-based temporal semantic segmentation framework for time-series land cover change detection in capturing intricate spatiotemporal dynamics. Another study, (Tejasree & Agilandeeswari, 2024), highlighted the benefits of sequence-based deep learning models for capturing intricate land-surface patterns pertinent to time-series remote sensing analyses, demonstrating that deep LSTM networks efficiently model temporal and spectral dependencies for LULC classification of hyperspectral imagery.

The current study offers numerous methodological improvements over the work of (Bekanov et al., 2020), who used Landsat-8 images (30×30 m) (Vaughn Ihlen, 2019) and a traditional unsupervised pixel-based methodology to map only agricultural areas within the Shimbay region.  First, Sentinel-2 imagery enabled more accurate land-cover classification owing to its enhanced spectral sensitivity and finer spatial resolution.  Second, the current study identified patterns of land alteration and degradation that are not apparent in single-year evaluations, using a multi-temporal analysis covering 2017–2022. This temporal depth provides important information on ongoing changes in vegetation decline, agricultural land expansion, and potential early indicators of desertification. Third, the use of AI-based categorization eliminated the need for manual photo downloads and GIS-based reclassification, which are typical of conventional approaches, by automating the production of training data.  This invention increased accuracy and reproducibility while reducing analytical time.  However, some restrictions still exist.  Seasonal land-use differences may be missed when using annual composite images. For example, croplands harvested in early summer may show as bare soil for the rest of the year.  The paucity of pre-2017 imagery in the Sentinel-2 archive also prevented the analysis from extending back to earlier decades, limiting the long-term historical perspective on LULC change.

The results of this study are consistent with those of Bekanov, who also demonstrated the usefulness of GIS and remote sensing for LULC mapping in the Shimbay district.  By leveraging AI-based time-series analysis and achieving greater geographical precision, the current study builds on their contribution. Moreover, the integration of artificial intelligence techniques into land cover classification reflects an emerging consensus in recent research. According to studies like those by (Maxwell et al., 2018) and (Pelletier et al., 2019) Machine learning and deep learning algorithms improve classification accuracy, automate analysis, and reduce reliance on manually collected field data. In dry and semi-arid regions such as the Aral Sea region, where regular monitoring is essential for understanding environmental degradation and informing policy responses, these methods are particularly useful.

The results highlight the potential of AI-enhanced Sentinel-2 time-series analysis for accurate detection of land-cover changes, despite certain data constraints.  Regional planning and sustainable land management can be substantially aided by the Impact Observatory Global LULC dataset when combined with high-resolution imagery.  These findings highlight the need for governments to focus on protecting agricultural land and mitigating degradation.  Seasonal composites and predictive models should be used in future studies to estimate potential LULC changes in the larger Aral Sea basin.

The research was conducted in the western part of Uzbekistan, in the Republic of Karakalpakstan, which is surrounded by the Kyzylkum Desert and the Ustyurt Plateau. This research used the Impact Observatory's deep-learning AI-based classification model on Sentinel-2 data. During the research, some positive and negative outcomes of the Impact Observatory's deep-learning AI land classification model were identified. The advantage of this model is that it helps clarify land types in the research area that were previously unfamiliar and facilitates analysis of the dynamics of land-type change over the years. Additionally, it requires less time for analysis and mapping than other types of the LULC classification process. Because Sentinel-2 10m Land Use and Land Cover Time Series data is included ready classified land classes data on images. On the ArcGIS.com website, information on the LULC classification model accuracy is 85%. In our research work, the average overall accuracy was 76% in two 2017-2022 years. The Kappa statistics results of the classified images belonged to the "Substantial" class. Moreover, user accuracy and producer accuracy also help to improve the precision of the classified LULC images of the Impact Observatory's deep-learning AI model.

These findings are vital for policymakers and environmental managers, providing a basis for creating informed conservation strategies. The study highlights the importance of sustainable development practices to reduce negative effects on these important natural resources. The AI models utilized in this research can act as a blueprint for similar studies in other areas, improving our ability to manage and safeguard natural landscapes in the face of increasing environmental challenges. In conclusion, the result showed that the Sentinel-2 Land Use and Land Cover Time Series data can be usable for mapping and analyzing LULC changes. In addition, the Overall Accuracy and Kappa Coefficients results also showed the acceptability of the classified images. We suggest that in the case of the Republic of Uzbekistan and Karakalpakstan, need to improve the quantity and quality of the LULC classification research works to understand the change of the land cover by the impact of global climate change and the impact of the Aral Sea dust storms.

This research was conducted at the “Joint Research Center for Geoin-formatics” at the National Research University “Tashkent Institute of Irrigation and Agricultural Mecha-nization Engineers.” Also, thanks to the ArcGIS company developers for providing us with ready satellite image data to conduct research on Land Use and Land Cover Classifi-cation.

Author Contributions

Conceptualization: Jumaniyazov, I.; methodology: Jumaniyazov, I.; inves-tigation: Jumaniyazov, I., Juliev, M., Reimov, M., Oymatov, R.,; writ-ing—original draft preparation: Jumaniyazov, I.,; writing—review and editing: Jumaniyazov, I.,; visual-ization: Jumaniyazov, I.,. All au-thors have read and agreed to the published version of the manuscript.

The author declares that he has no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data is available upon Request.

This research received no external funding.