Land-cover classification in Addo Elephant National Park: Analyzing the impact of variables, classifiers, and object-based approach

A comprehensive understanding of land-use and land-cover (LULC) dynamics is vital in steering effective conservation and management efforts, especially in ecologically rich regions like Addo Elephant National Park (AENP). Despite its importance, up-to-date LULC maps of AENP remain scarce, needing an in-depth investigation to aid conservation planning. Using Landsat time-series data, this study produced 30-m resolution LULC maps for the years 2002, 2014, and 2022, and examined changes within six LULC categories. Object-based classification was compared with a pixel-based approach, revealing the superior performance of the pixel-based approach. Two machine learning (ML) techniques, Random Forest (RF) and Support Vector Machines (SVM), were compared with a deep learning (DL) technique, UNet++. The land-cover classification process using ML algorithms involved experimentation with various predictor variables, including spectral bands, spectral indices, time-series data, and textural information. Spectral mixture analysis was performed, and the resulting fraction layers were used as independent variables in the models. The study identified RF as the preferred classification algorithm using the optimal combination of these variables, achieving high accuracy of 89.1 %, 91.2 %, and 91.9 % for the years 2002, 2014, and 2022, respectively. Variable importance analysis highlighted the consistent significance of elevation, slope, and the time-series of normalized difference indices. The final land-cover maps revealed grass as the predominant class both inside and outside AENP, followed by thicket within the park, and agriculture outside it. Land-cover change analysis indicated small changes (<3 %), primarily involving transitions between thicket and grass classes inside the park, and grass and agriculture outside. • Land-cover classification in Addo Elephant National Park using Landsat data. • Impact of spectral indices, textural metrics, and SMA on model was analyzed. • Pixel-based classification outperformed the object-based classification approach. • Random Forest achieved higher overall accuracy compared to SVM and UNet++.