Nzuza, PRamoelo, AbelOdindi, JMwenge Kahinda, Jean-MarcMadonsela, Sabelo2020-11-062020-11-062020-09Nzuza, P., Ramoelo, A., Odindi, J., et al. 2020. Predicting land degradation using Sentinel-2 and environmental variables in the Lepellane catchment of the Greater Sekhukhune District, South Africa. Physics and Chemistry of the Earth, Parts A/B/C, pp. 1-121474-70651873-5193https://doi.org/10.1016/j.pce.2020.102931https://www.sciencedirect.com/science/article/pii/S1474706520303776http://hdl.handle.net/10204/11665Copyright: 2020, Elsevier. Due to copyright restrictions, the attached PDF file contains the abstract of the full-text item. For access to the full-text item, please consult the publisher's website.Land degradation is defined as the reduction of biological and economic productivity, which impedes the capacity of the land to provide ecosystem services. There is a need to move towards near real-time monitoring of land degradation using new sensors to detect degraded landscapes. Recently launched Sentinel-2 sensor presents the opportunity to collect high-resolution data regularly. Multi-temporal datasets provide crucial information to isolate evidence of land degradation from temporal changes in vegetation cover incurred as a result of climatic and phenological variability. This study applied an integrated approach involving multi-seasonal Sentinel-2 data with environmental variables (i.e. soil moisture, rainfall, slope, evapotranspiration, elevation, soil temperature, rainfall, soil temperature, aspect and albedo). A stratified random sampling approach based on dominant land cover types were used to assess land degradation. Field plots of 20 m × 20 m were setup with three 50 cm × 50 cm quadrants inside. In each quadrant, the percentage estimation of vegetation cover and Leaf Area Index was measured. The model training and validation was implemented using the Random Forest algorithm based on default parameters. The pooled model represents the dry and wet seasons combined. Results showed that pooled model had higher accuracies for Photosynthetic vegetation (PV) (R2 of 0.89, RMSE-11.46%, relRMSE-8.7%), Non-Photosynthetic Vegetation (NPV) (R2 of 0.93, RMSE-5.64%, relRMSE% 17.72) and Bare Soil (BS) (R2 of 0.92, RMSE 8.7% and relRMSE 11.46%). The pooled environmental model achieved accuracy of PV (R2 of 0.42, RMSE-20.67%, relRMSE 4.83%), NPV (R2 of 0.90, RMSE 6.50% and relRMSE 15%) and (BS R2 of 0.85, RMSE 8.64% and relRMSE% 11.5) in estimating vegetation cover.enLand degradation risksLepellane catchmentMachine learning algorithmsRandom forestSentinel-2 MSIArticleNzuza, P., Ramoelo, A., Odindi, J., Mwenge Kahinda, J., & Madonsela, S. (2020). Predicting land degradation using Sentinel-2 and environmental variables in the Lepellane catchment of the Greater Sekhukhune District, South Africa. http://hdl.handle.net/10204/11665Nzuza, P, A Ramoelo, J Odindi, Jean-marc Mwenge Kahinda, and Sabelo Madonsela "Predicting land degradation using Sentinel-2 and environmental variables in the Lepellane catchment of the Greater Sekhukhune District, South Africa." (2020) http://hdl.handle.net/10204/11665Nzuza P, Ramoelo A, Odindi J, Mwenge Kahinda J, Madonsela S. Predicting land degradation using Sentinel-2 and environmental variables in the Lepellane catchment of the Greater Sekhukhune District, South Africa. 2020; http://hdl.handle.net/10204/11665.TY - Article AU - Nzuza, P AU - Ramoelo, A AU - Odindi, J AU - Mwenge Kahinda, Jean-marc AU - Madonsela, Sabelo AB - Land degradation is defined as the reduction of biological and economic productivity, which impedes the capacity of the land to provide ecosystem services. There is a need to move towards near real-time monitoring of land degradation using new sensors to detect degraded landscapes. Recently launched Sentinel-2 sensor presents the opportunity to collect high-resolution data regularly. Multi-temporal datasets provide crucial information to isolate evidence of land degradation from temporal changes in vegetation cover incurred as a result of climatic and phenological variability. This study applied an integrated approach involving multi-seasonal Sentinel-2 data with environmental variables (i.e. soil moisture, rainfall, slope, evapotranspiration, elevation, soil temperature, rainfall, soil temperature, aspect and albedo). A stratified random sampling approach based on dominant land cover types were used to assess land degradation. Field plots of 20 m × 20 m were setup with three 50 cm × 50 cm quadrants inside. In each quadrant, the percentage estimation of vegetation cover and Leaf Area Index was measured. The model training and validation was implemented using the Random Forest algorithm based on default parameters. The pooled model represents the dry and wet seasons combined. Results showed that pooled model had higher accuracies for Photosynthetic vegetation (PV) (R2 of 0.89, RMSE-11.46%, relRMSE-8.7%), Non-Photosynthetic Vegetation (NPV) (R2 of 0.93, RMSE-5.64%, relRMSE% 17.72) and Bare Soil (BS) (R2 of 0.92, RMSE 8.7% and relRMSE 11.46%). The pooled environmental model achieved accuracy of PV (R2 of 0.42, RMSE-20.67%, relRMSE 4.83%), NPV (R2 of 0.90, RMSE 6.50% and relRMSE 15%) and (BS R2 of 0.85, RMSE 8.64% and relRMSE% 11.5) in estimating vegetation cover. DA - 2020-09 DB - ResearchSpace DP - CSIR KW - Land degradation risks KW - Lepellane catchment KW - Machine learning algorithms KW - Random forest KW - Sentinel-2 MSI LK - https://researchspace.csir.co.za PY - 2020 SM - 1474-7065 SM - 1873-5193 T1 - Predicting land degradation using Sentinel-2 and environmental variables in the Lepellane catchment of the Greater Sekhukhune District, South Africa TI - Predicting land degradation using Sentinel-2 and environmental variables in the Lepellane catchment of the Greater Sekhukhune District, South Africa UR - http://hdl.handle.net/10204/11665 ER -