Land use land cover change detection using machine learning in google earth engine of Morna Watershed

Introduction

Land use and land cover (LULC) dynamics are critical indicators of environmental change, reflecting complex interactions between natural processes and human activities (Osman et al., 2023) [14] . Rapid population growth, urban expansion, agricultural intensification, and climate variability have accelerated landscape transformation, significantly impacting ecosystem services, hydrology, and biodiversity (Sun and Zhang, 2025) [21] . Consequently, accurate and timely monitoring of LULC is essential for sustainable land management and effective environmental planning (Tahmi et al., 2026) [22] . Remote sensing has emerged as a powerful tool for LULC mapping due to its synoptic coverage and temporal consistency (Loukika et al., 2021) [11] . The availability of multi-temporal satellite data such as Landsat and Sentinel, combined with advancements in cloud-based platforms like Google Earth Engine (GEE), has greatly improved large-scale environmental monitoring capabilities (Pande et al., 2023) [16] . GEE enables efficient processing of vast geospatial datasets by utilizing parallel processing and Map Reduce architecture, facilitating advanced analyses without the need for high-end local computing resources (Ahmed et al., 2024) [2] . The integration of machine learning (ML) algorithms, including Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Trees (CART), has further enhanced classification accuracy (Basheer et al., 2022) [4] . These models effectively handle complex, nonlinear data and have demonstrated high performance in diverse landscapes (Belay et al., 2024; Tesfaye et al., 2024) [5,23] . Studies indicate that RF frequently outperforms other classifiers in terms of overall accuracy and robustness, particularly in agriculturally dominant regions (Pande et al., 2024) [17] . However, challenges remain in selecting optimal training data, integrating auxiliary variables, and ensuring consistent performance across regions ( [3,11] . In India, particularly in Maharashtra, watershed systems are increasingly affected by land transformation driven by rainfed agriculture and limited water management (Pande, 2022) [15] . The Morna watershed exemplifies such vulnerability, characterized by steep slopes and high erosion potential (Ingle, 2022) [10] . Despite its hydro-ecological importance, previous studies have largely focused on morphometry rather than advanced LULC analysis (Chavhan, 2015, Shinde et al., 2024) [7,19] . Furthermore, a specific challenge in this region is the spectral confusion between urban impervious surfaces and the region's characteristic black cotton soils, which necessitates sophisticated feature-engineering. In this context, the present study aims to address these gaps by employing machine learning algorithms within the GEE platform for LULC mapping and change detection in the Morna watershed. The specific objectives are, a) To generate accurate LULC maps using multi-temporal satellite imagery (Landsat 8/9 and Sentinel-2). b) To evaluate the performance of selected machine learning classifiers (RF, SVM, and CART). c) To analyze spatiotemporal LULC changes and land cover transitions within the watershed.

Materials and Methods

Study area

Morna is tributary of Koyna River, which is one of the major tributaries of river Krishna in Maharashtra. As shown in Fig. 1, this catchment lies in Satara district of Maharashtra state in western India, between 17º 24' N to 17º 50' N latitude and 73º 46' E to 74º 0' E longitudes. The total area of watershed is 13,285.13 ha. Length of Morna river is 24.25 km. The average minimum and maximum temperatures are 7.50C to 38.50C, respectively and May is hottest month. The average annual rainfall is 2012 mm. The soil is laterite having dark reddish to yellowish red color. The textural class is clay-to clay loam and has pH value between 4 and 5.8. Cropping pattern of the watershed is dominated by cereals, paddy and millets which are major Kharif crops (Bowlekar et al. 2018, Shinde and Nandgude, 2024) [6,18] .

Data Types and Sources

Multi-temporal satellite imagery was acquired from the Landsat series (Landsat 5 TM, Landsat 7 ETM+, and Landsat 8 OLI/TIRS) through the Google Earth Engine (GEE) platform. Landsat data were selected due to their long-term availability, moderate spatial resolution (30 m), and suitability for LULC classification. Satellite images corresponding to selected time periods (e.g., 1985, 1995, 2005, 2015, and 2025) were used to capture temporal changes. Images were selected based on minimal cloud cover and seasonal consistency, typically during the post-monsoon or early winter period to ensure optimal spectral separability of land cover classes. [24,23] . Surface reflectance products (Landsat Collection 2 Level 2 and Sentinel-2 Level-2A) were prioritized to minimize the influence of atmospheric interference on spectral values (Belay et al., 2024) [5] . Images were precisely clipped to the Morna watershed boundary using predefined vector data to isolate the specific area of interest (Sultana et al., 2024; Moharir et al., 2024) [20,12] .

Temporal compositing, specifically employing median filtering, allowed for the generation of cloud-free, representative imagery for each study year by reducing variability in pixel values (Nasiri et al., 2022) [13] . These comprehensive steps ensured the integrity of spectral information for accurate machine learningbased LULC classification (Osman et al., 2023) [14] .

Variables and Spectral Indices

To enhance class separability and improve classification accuracy, several spectral indices were derived from the satellite imagery. These includes, Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Built-up Index (NDBI), Bare Soil Index (BSI). Additionally, topographic variables such as elevation and slope derived from the SRTM DEM were incorporated as auxiliary predictors, as these factors significantly influence land cover distribution and classification performance.

Training and Validation Data

Training and validation datasets were generated using a combination of field knowledge, high-resolution imagery, and visual interpretation. Representative samples for each LULC class were collected using a stratified random sampling approach to ensure adequate representation of all land cover types (Basheer et al., 2022) [4] . The dataset was divided into training and validation subsets, typically following a 70:30 split.

The training dataset was used to develop classification models, while the validation dataset was used to assess model performance (Feizizadeh et al., 2023; Tesfaye et al., 2024) [23] .

Classification and Modelling Approach

Supervised machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Trees (CART), were implemented within the GEE platform. These algorithms were selected due to their robustness, efficiency, and proven performance in LULC classification (Hasoba et al., 2025; Ahmad et al., 2026) [9,1] [8,25] .

Accuracy Assessment

The accuracy of the classified LULC maps was evaluated using a confusion matrix approach. Performance metrics including Overall Accuracy (OA), Kappa Coefficient (K), Producer's Accuracy (PA), and User's Accuracy (UA) were computed.

Producers Accuracy

Where, C ii is the number of correctly classified pixels for class is the total number of actual pixels in class I (The sum of the row in the confusion matrix corresponding to class i)

User's Accuracy

Where, C ii is the number of correctly classified pixels for class i.

is the total number of pixels classified as class i (the sum of the column in the confusion matrix corresponding to class i)

Kappa Coefficient

Where, P o is the observed agreement (i.e., overall accuracy) P e is the expected agreement by chance Post-classification comparison was employed to analyze LULC changes over time. Classified maps from different years were compared to quantify transitions between land cover classes. Change matrices were generated to identify dominant trends such as agricultural expansion, deforestation, or urban growth.

Results and Discussion

Accuracy Assessment

The classification performance of Random Forest (RF), CART, and SVM models was evaluated using Accuracy (OA) and the Kappa coefficient for the years 1995, 2005, 2015, and 2025as shown in Table 1. The results show that RF consistently outperformed the other models across all years. The accuracy of RF ranged from 97.32% (1995) to 98.56% (2015), which was higher than CART (96.06%-97.24%) and SVM (92.96%-97.77%). Similarly, the Kappa coefficient values for RF (0.948-0.973) were also higher compared to CART (0.927-0.949) and SVM (0.863-0.958), indicating stronger agreement and reliability. Finally, Random Forest proved to be the most accurate and stable classifier, making it the best model for LULC classification in this study.

LULC Change Detection (1995-2005)

The LULC change analysis between 1995 and 2005 shows that a large portion of the landscape remained stable as shown in

Conclusion

The LULC change analysis from 1995 to 2025 reveals that forest and barren land are the dominant and relatively stable land cover classes in the study area. However, significant transitions such as forest to agriculture and barren to agriculture indicate increasing anthropogenic pressure and agricultural expansion. Additionally, the growth of built-up areas highlights ongoing urbanization. These changes suggest a gradual transformation of natural landscapes into human-dominated land uses, emphasizing the need for sustainable land management strategies.

References

1. Ahmad M, Alam K, Ahmad M, Khan K, Zeb B, Ditta A. Integrating Google Earth Engine and machine learning for urban land use and land cover dynamics analysis. Environmental Monitoring and Assessment. 2026;198(2):105. https://doi.org/10.1007/s10661-025-14960-0
2. Ahmed R, Zafor MA, Trachte K. Land-use and land-cover changes in Cottbus city and Spree-Neisse district, Germany, in the last two decades: a study using remote sensing data and Google Earth Engine. Remote Sensing. 2024;16(15):2773. https://doi.org/10.3390/rs16152773
3. Alshehri B, Zhang Z, Liu X. A review of Google Earth Engine for land use and land cover change analysis: trends, applications and challenges. ISPRS International Journal of Geo-Information. 2025;14:416. https://doi.org/10.3390/ijgi14110416
4. Basheer S, Wang X, Farooque AA, Nawaz RA, Liu K, Adekanmbi T, et al. Comparison of land use land cover classifiers using different satellite imagery and machine learning techniques. Remote Sensing. 2022;14(19):4978. https://doi.org/10.3390/rs14194978
5. Belay H, Melesse AM, Tegegne G. Scenario-based land use and land cover change detection and prediction using the cellular automata-Markov model in the Gumara watershed, Upper Blue Nile Basin, Ethiopia. Land. 2024;13(3):396. https://doi.org/10.3390/land13030396
6. Bowlekar A, Sawant H, Nandgude S, Mahale D. Morphometric analysis of Morna River catchment using geographic information system. Agricultural Engineering Today. 2019;43(1):54-61.
7. Chavhan SR. Palaeochannel mapping of Morna watershed, Akola district, Maharashtra, India. Science Park Research Journal. 2015;2(51).
8. Fariz TR, Suhardono S, Fadhilla S. Rapid mapping of land cover changes in tropical watershed in Indonesia using Google Earth Engine. IOP Conference Series: Earth and Environmental Science. 2024;1414(1):012077. https://doi.org/10.1088/1755-1315/1414/1/012077
9. Hasoba AMM, Yasin EHE, Osman MBO, Czimber K. Monitoring ecosystem dynamics using machine learning: random forest-based LULC analysis in Dinder Biosphere Reserve, Sudan. Engineering Proceedings. 2025;94(1):2. https://doi.org/10.3390/engproc2025094002
10. Ingle NB. Morphometric analysis of Morna River sub basin using geospatial approach. International Journal of Multidisciplinary Research and Development. 2022;9(1):71-78.
11. Loukika KN, Keesara VR, Sridhar V. Analysis of land use and land cover using machine learning algorithms on Google Earth Engine for Munneru River Basin, India. Sustainability. 2021;13(24):13758. https://doi.org/10.3390/su132413758
12. Moharir KN, Pande CB, Gautam VK, Elsahabi M. Estimation of land surface temperature and LULC changes impact on groundwater resources in the semi-arid region of Madhya Pradesh, India. Advances in Space Research. 2024;75(1):233-247. https://doi.org/10.1016/j.asr.2024.09.025
13. Nasiri V, Deljouei A, Moradi F, Sadeghi SMM, Borz SA. Land use and land cover mapping using Sentinel-2, Landsat-8 satellite images and Google Earth Engine: a comparison of two composition methods. Remote Sensing. 2022;14(9):1977. https://doi.org/10.3390/rs14091977
14. Osman MA, Abdel-Rahman EM, Onono JO, Olaka LA, Elhag MM, Adan M, et al. Mapping, intensities and future prediction of land use/land cover dynamics using Google Earth Engine and CA-artificial neural network model. PLoS ONE. 2023;18(7):e0288694. https://doi.org/10.1371/journal.pone.0288694
15. Pande CB. Land use/land cover and change detection mapping in Rahuri watershed area (MS), India using Google Earth Engine and machine learning approach. Geocarto International. 2022;37(26):13860-13880. https://doi.org/10.1080/10106049.2022.2086622
16. Pande C, Diwate P, Orimoloye I, Sidek L, Mishra A, Moharir K, et al. Impact of land use/land cover changes on evapotranspiration and model accuracy using Google Earth Engine and classification and regression tree modeling. Geomatics, Natural Hazards and Risk. 2023;15. https://doi.org/10.1080/19475705.2023.2290350
17. Pande C, Srivastava A, Moharir K, Radwan N, Sidek L, Alshehri F, et al. Characterizing land use/land cover change dynamics by an enhanced random forest machine learning model: a Google Earth Engine implementation. Environmental Sciences Europe. 2024;36:1-23. https://doi.org/10.1186/s12302-024-00901-0
18. Shinde SS, More DD, Shendkar DA, Dahiphale MS, Thange AB, Nandgude SB. Temporal and spatial dynamics of land use and land cover in Shirur Kasar Tehsil, Maharashtra using geospatial technology. International Journal of Agricultural Inventions. 2024;9(1):193-204. https://doi.org/10.46492/IJAI/2024.9.1.25
19. Shinde S, Nandgude S. Soil organic carbon stocks assessment under different land use land cover in Morna watershed, India using GIS technique. Mysore Journal of Agricultural Sciences. 2024;58(3).
20. Sultana MS, Jayalakshmi G, Likhitha CD. Time series analysis of surface water areas using Sentinel imagery on Google Earth Engine: a spatial approach. Scientific Reports. 2024. https://doi.org/10.1007/978-981-97-6489-1_25
21. Sun Y, Zhang J. Advances in research on the impact of land use/cover changes on watershed hydrological processes. Frontiers in Sustainable Development. 2025;5(3):193-196. https://doi.org/10.54691/5xyggb72
22. Tahmi S, Rui X, Guechi I, Hannache F, Wang X, Chebika H. Mapping four decades of change: deep learning and Google Earth Engine for annual LULC classification in semi-arid regions. Environmental Modelling and Software. 2026;200:106940. https://doi.org/10.1016/j.envsoft.2026.106940
23. Tesfaye W, Elias E, Warkineh B, Tekalign M, Abebe G. Modeling of land use and land cover changes using Google Earth Engine and machine learning approach: implications for landscape management. Environmental Systems Research. 2024;13(1):31. https://doi.org/10.1186/s40068024-00366-3
24. Vyas N, Sundar K, Sharma A, Adhikari PS. Leveraging Sentinel-2 multispectral data and machine learning algorithms for land use land cover mapping in semi-arid regions. In: Proceedings of the 3rd International Conference on Advancement in Computation and Computer Technologies; 2025; Gharuan, India. p. 439-443. https://doi.org/10.1109/InCACCT 6 5424.2025.11011472
25. Waindeshkar P, Umrikar B, Srujanika S, Shirke A. Impact of land use changes on groundwater recharge in Deccan basaltic aquifers from western India. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences. 2025;X-5/W2-2025:719-723. https://doi.org/10.5194/isprs-annals-X-5-W2-2025-7192025