Characterising social-ecological drivers of landuse/cover change in a complex transboundary basin using singular or ensemble machine learning

• Published in: Remote sensing applications 2022-08, Vol.27, p.100773, Article 100773
• Main Authors: Kavhu, Blessing, Eric Mashimbye, Zama, Luvuno, Linda
• Format: Article
• Language: English
• Subjects: Landscape change; Machine learning; Remote sensing; Spatial modelling; Water resources
• ISSN: 2352-9385; 2352-9385
• DOI: 10.1016/j.rsase.2022.100773

Studies have focused on understanding land use/cover (LULC) change through regression techniques. However, machine learning (ML) techniques and their ensembles may provide more accurate results. Accurate determination of drivers of LULC can guide reliable land use planning and effective natural resource management. In this study, we tested the utility of ML techniques and ensemble modelling to explain the social-ecological drivers of LULC in the Okavango basin. The Deep Neural Network (DNN) coupled with climate-based regionalization of the study area were used for LULC classification of the years 2002, 2013 and 2020. Centroids of 22 LULC transitions of the same period were used to separately calibrate five (5) machine algorithms (namely Random Forests (RF), Gradient Boost Models (GBM) and Maximum Entropy (MaxEnt), Classification Tree Analysis (CTA) and their ensemble. Model performance was evaluated using the Receiver Operating Characteristic (ROC) and the True skill statistic (TSS). Variable importance was used to assess the contribution of social-ecological variables to each LULC transition. Variables that were determined to be driving LULC were then used to predict future LULC using the Artificial Neural Network and Cellular Automata (ANN-CA). Analysis results show that average LULC classification accuracy for the study period (2002, 2013 and 2020) was (Bsh Koppen zone; OA = 95.03, Kappa = 0.94), (Cwa Koppen zone; OA = 95.29, Kappa = 0.94), Cwb Koppen zone; OA = 96.04, Kappa = 0.95). The ML ensemble performed better (ROC> 95, TSS >87) than singular ML models based on two separate model evaluation metrics. The Random Forest classifier outperformed other singular ML (ROC = 90.41, TSS = 84.2). Based on the top-performing ensemble model, distance from rivers, population density, annual average temperature, drought severity, fire frequency and distance from towns influence the conversion of natural to anthropogenic LULC classes (importance >0.5). On the other hand, distance from rivers, soil organic carbon, precipitation, GDP, elevation, population density and annual average temperature importantly influenced conversion from one natural LULC class to another natural LULC class. The study revealed that natural classes (wetland, shrubland, water and woodlands) will gradually decrease at the expense of anthropogenic classes (built-up and cultivated) in future (2040). Despite proposing the necessity of a basin-wide land-use plan to minimise pressure on resources