Abstract ID: 3.12295 | Accepted as Talk | Talk/Oral | TBA | TBA

Land use and land cover (LULC) changes significantly influence evapotranspiration (ET) in hydrology, necessitating accurate classification of LULC in high-elevation regions for reliable hydrological assessments. The use of outdated LULC maps in hydrological applications is problematic due to ongoing changes in catchments related to human and natural activities. Advancements in earth observation technologies, semi-automated classification techniques, and cloud computing platforms like Google Earth Engine have enabled rapid and accurate mapping of LULC types. The study aims to enhance seasonal LULC mapping accuracy in a South African montane fire-climax grassland by integrating Sentinel-1 and Sentinel-2 data with spectral indices, texture, and terrain features using machine learning classifiers (Random Forest (RF), Gradient Boosting (GB) and the Support Vector Machine (SVM)). This study presents an innovative method that explores LULC classification of South African montane fire-climax grasslands through multi-source data fusion and machine learning. This approach takes into consideration biennial management burns in autumn that are employed to preserve pristine conditions and mitigate encroachment, a critical factor affecting LULC classification. We conducted an analysis of LULC in four distinct seasons: Summer (DJF), Autumn (MAM), Winter (JJA), and Spring (SON). Multi-source data substantially enhanced the results of LULC classification including vegetation types, bare rock, burnt areas, water bodies, grasslands, and shrublands. The highest overall classification accuracies of 95% was achieved in Summer, 92% in Autumn, 85% in Winter, and 90% in Spring with the RF classifier. The inclusion of SAR data, elevation, and NDVI during non-burn periods as inputs into the RF classifier, as well as the addition of spectral index dBNR during biannual burn periods, consistently emerged as highly influential in the determination of LULC classes, as demonstrated by the variable importance analysis. These results are in line with previous research that has shown the advantages of combining optical and SAR data with sophisticated classifiers. This integration significantly improves the accuracy of LULC classification, particularly in difficult environments, and introduces a new dimension to accurately and efficiently classify LULC in montane fire-climax grasslands.