XGBoost-LSTM Hybrid Model for Predicting Vegetation Cover in Kenya\'s North Rift Region: Leveraging LULC and Environmental Variables

Abstract

In Kenya\'s North Rift region, rapid Land Use Land Cover (LULC) transformations have resulted from agricultural activities, population growth, and infrastructure development. While these changes support economic growth and food security, they impact terrestrial vegetation health and ecosystem stability. Accurate mapping and prediction of these changes is crucial for sustainable land use strategies. A key vegetation health metric is the Normalized Difference Vegetation Index (NDVI). However, predicting NDVI from LULC is challenging due to their complex non-linear relationships and environmental influences. Additionally, satellite-based NDVI predictions face limitations from atmospheric distortions and sensor noise complicating spatial and temporal trend analysis. To enhance NDVI prediction, this study explored a hybrid approach that integrates spatial analysis with temporal forecasting for NDVI prediction. It employed ensemble methods for spatial feature extraction and Long Short-Term Memory (LSTM) networks for capturing temporal dependencies. First, LULC changes and their relationship with NDVI were analyzed, followed by a comparison of ensemble methods to identify the better model for capturing non-linear interactions between LULC and environmental variables. The superior ensemble model was then integrated with LSTM to develop a hybrid model. Comparing the performance of ensemble models, the results showed that XGBoost performed superiorly. Therefore, based on XGBoost\'s superior performance, an XGBoost-LSTM hybrid model was developed, improving prediction accuracy with an RMSE of 0.085 and an R² score of 0.92. The XGBoost-LSTM hybrid model surpassed individual model performance, enhancing predictive accuracy, scalability, and robustness. This study provides a framework for large-scale vegetation health monitoring with practical applications in sustainable land management and climate adaptation.

Introduction

The text presents a study on Land Use and Land Cover (LULC) changes and their impact on vegetation health, measured using the Normalized Difference Vegetation Index (NDVI), in Kenya’s North Rift region. The study proposes a hybrid XGBoost-LSTM model to improve NDVI prediction by combining spatial and temporal analysis.

• Background: LULC changes caused by urbanization, agricultural expansion, deforestation, population growth, and climate variability significantly affect vegetation, biodiversity, and ecosystem stability. Monitoring these changes is important for sustainable land management.

• NDVI: NDVI is used to measure vegetation health from satellite imagery. High NDVI indicates healthy vegetation, while low or negative values generally indicate sparse vegetation, bare land, or water. NDVI changes according to land-use conditions, climate, soil moisture, and human activities.

• Study objective: The research aims to:

  1. Analyze LULC changes and their relationship with NDVI.
  2. Compare Random Forest (RF) and XGBoost for modeling complex environmental relationships.
  3. Combine the better-performing ensemble model with LSTM to capture both spatial and temporal NDVI patterns.

• Previous research: Traditional LULC mapping methods were labor-intensive, while remote sensing and machine learning have improved classification. CNNs, RNNs, and LSTMs provide stronger capabilities for complex spatial and temporal analysis, but deep-learning approaches often require large datasets and significant computational resources.

• Research gap: Existing hybrid models can be computationally expensive, data-intensive, and highly dependent on satellite imagery. Many also focus primarily on either spatial or temporal patterns rather than effectively integrating both. The proposed approach attempts to overcome these limitations using structured environmental variables such as temperature, precipitation, Bare Soil Index (BSI), Soil Moisture Index (SMI), and LULC.

• Study area: The research focuses on Kenya's North Rift region, covering Trans-Nzoia, Uasin Gishu, Elgeyo-Marakwet, Nandi, Samburu, Turkana, Baringo, and West Pokot. The region contains forests, grasslands, shrublands, wetlands, agricultural areas, and arid zones.

• Datasets: The study uses Landsat imagery from 1990–2021 and Sentinel-2 imagery for 2022, along with precipitation, temperature, BSI, and SMI data. These datasets support both LULC classification and NDVI prediction.

• Models:

  • Random Forest and XGBoost were used for LULC classification and spatial feature extraction.
  • LSTM was used to model temporal NDVI patterns.
  • The best ensemble model was then combined with LSTM to create a hybrid prediction system.

• LULC findings: Cropland increased substantially, particularly between 2016 and 2020, while natural forest areas showed a declining trend. Increasing bareland suggested possible environmental degradation caused by factors such as deforestation, erosion, and overgrazing.

• NDVI findings: Natural forests had the highest NDVI values, indicating healthy vegetation. Croplands had moderate NDVI, while grasslands showed lower values. Water bodies and bareland generally produced negative NDVI values.

• Important predictors: LULC was the most influential factor in predicting NDVI, followed by minimum temperature, precipitation, and soil moisture. Maximum temperature and BSI had comparatively lower influence.

• Classification performance: XGBoost achieved an overall 88% LULC classification accuracy, with strong performance for natural forests, bareland, and grasslands. Cropland was the most difficult category to classify, mainly because of its relatively low recall.

• Model comparison: Random Forest using only LULC performed poorly for NDVI prediction (RMSE = 0.6948, R² = 0.5281). Adding environmental variables improved it considerably (RMSE ≈ 0.47, R² ≈ 0.78). An advanced XGBoost model incorporating lagged NDVI and rolling environmental features achieved RMSE = 0.5008 and R² = 0.7553, although redundant features may have introduced noise or overfitting.

Conclusion

Model 3 (XGBoost with selected features) proved to be the most effective NDVI predictor, showing that a well-curated feature set outperforms complex models with excessive variables. While Random Forest performed moderately (RMSE = 0.47, R² = 0.78), it failed to capture intricate relationships as effectively as XGBoost. Model 2 (XGBoost with additional features) did not significantly improve accuracy, likely due to redundant features or overfitting. The hybrid LSTM-XGBoost model combined LSTM\'s ability to capture temporal trends with XGBoost\'s feature importance insights, achieving superior accuracy (RMSE = 0.085, R² = 0.92). This approach enhances NDVI prediction in diverse land-use scenarios. Despite its strengths, the hybrid model faces challenges such as computational complexity, potential overfitting, and data quality issues stemming from inconsistencies in satellite imagery. Nevertheless, it holds significant potential for real-world applications, including environmental monitoring, land-use planning, and climate adaptation strategies. Future research should focus on refining model architecture, improving the quality of input data, and incorporating socio-economic variables to enhance predictive accuracy and scalability across different ecological contexts.

References

[1] Acuña-Alonso, C., García-Ontiyuelo, M., Barba-Barragáns, D., Álvarez, X.: Development of a convolutional neural network to accurately detect land use and land cover. MethodsX 12, 102719 (2024). [2] Belgiu, M., Dr?gu?, L.: Random forest in remote sensing: A review of applications and future directions. ISPRS Journal of Photogrammetry and Remote Sensing 114, 24–31 (2016). [3] Chang, C.H., Lin, J., Chang, J.W., Huang, Y.S., Lai, M.H., Chang, Y.J.: Hybrid deep neural networks with multi-tasking for rice yield prediction using remote sensing data. Agriculture 14(4), 513 (2024). [4] Cheng, X., Sun, Y., Zhang, W., Wang, Y., Cao, X., Wang, Y.: Application of deep learning in multitemporal remote sensing image classification. Remote Sensing 15(15), 3859 (2023). [5] Kulithalai Shiyam, P., Deka, P.C.: Spatio-temporal classification and prediction of land use and land cover change for the Vembanad Lake system, Kerala. Environmental Science and Pollution Research 29(57), 86220–86236 (2022). [6] Li, Z., Chen, B., Wu, S., Su, M., Chen, J.M., Xu, B.: Deep learning for urban land use category classification: A review and experimental assessment. Remote Sensing of Environment 311, 114290 (2024). [7] Liping, C., Yujun, S., Saeed, S.: Monitoring and predicting land use and land cover changes using remote sensing and GIS techniques—a case study of a hilly area, Jiangle, China. PloS One 13(7), e0200493 (2018). [8] Martinuzzi, F., Mahecha, M.D., Camps-Valls, G., Montero, D., Williams, K.: Learning extreme vegetation response to climate drivers with recurrent neural networks. Nonlinear Processes in Geophysics 31(4), 535–557 (2024). [9] Mulinge, W., Gicheru, P., Murithi, F., Maingi, P., Kihiu, E., Kirui, O.K., Mirzabaev, A.: Economics of land degradation and improvement in Kenya. Economics of Land Degradation and Improvement–A Global Assessment for Sustainable Development, pp. 471–498 (2016). [10] Najmuddin, O., Li, Z., Khan, R., Zhuang, W.: Valuation of land-use/land-cover-based ecosystem services in Afghanistan - an assessment of the past and future. Land 11(11), 1906 (2022). https://doi.org/10.3390/land11111906 [11] Nyariki, D.M., Amwata, D.A.: The value of pastoralism in Kenya: Application of total economic value approach. Pastoralism 9(1), 1–13 (2019). [12] Rangel, A., Terven, J., Cordova-Esparza, D.M., Chavez-Urbiola, E.A.: Land cover image classification. arXiv preprint arXiv:2401.09607 (2024). [13] Sahoo, A.R., Chakraborty, P.: Hybrid CNN Bi-LSTM neural network for hyperspectral image classification. arXiv preprint arXiv:2402.10026 (2024). [14] Seifu, T.K., Woldesenbet, T.A., Alemayehu, T., Ayenew, T.: Spatio-temporal change of land use/land cover and vegetation using multi-MODIS satellite data, western Ethiopia. The Scientific World Journal 2023(1), 7454137 (2023). [15] Singh, S.K., Laari, P.B., Mustak, S., Srivastava, P.K., Szabó, S.: Modelling of land use land cover change using earth observation data-sets of Tons River Basin, Madhya Pradesh, India. Geocarto International 33(11), 1202–1222 (2018). [16] Solórzano, J.V., Mas, J.F., Gao, Y., Gallardo-Cruz, J.A.: Land use land cover classification with U-Net: Advantages of combining Sentinel-1 and Sentinel-2 imagery. Remote Sensing 13(18), 3600 (2021). [17] Talukdar, S., Singha, P., Mahato, S., Pal, S., Liou, Y.A., Rahman, A.: Land-use land-cover classification by machine learning classifiers for satellite observations—a review. Remote Sensing 12(7), 1135 (2020). [18] Taye, M.M.: Understanding of machine learning with deep learning: architectures, workflow, applications and future directions. Computers 12(5), 91 (2023). [19] Ullah, S., Qiao, X., Abbas, M.: Addressing the impact of land use land cover changes on land surface temperature using machine learning algorithms. Scientific Reports 14(1), 18746 (2024). [20] Wang, J., Bretz, M., Dewan, M.A.A., Delavar, M.A.: Machine learning in modelling land-use and land cover-change (LULCC): Current status, challenges and prospects. Science of The Total Environment 822, 153559 (2022). [21] Wiatkowska, B., S?odczyk, J., Stokowska, A.: Spatial-temporal land use and land cover changes in urban areas using remote sensing images and GIS analysis: The case study of Opole, Poland. Geosciences 11(8), 312 (2021). [22] Zhang, X., Zhou, Y., Luo, J.: Deep learning for processing and analysis of remote sensing big data: A technical review. Big Earth Data 6(4), 527–560 (2022). [23] Zhao, Y., Hou, P., Jiang, J., Zhao, J., Chen, Y., Zhai, J.: High-spatial-resolution NDVI reconstruction with GA-ANN. Sensors 23(4), 2040 (2023).

Copyright

Copyright © 2026 Eve Khatundi Wamasebu, Betty Mayeku, Jonathan Mutonyi, Nelly Masayi. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.