Christobaltobbin/Segment_Anything_Model-SAM_vs_Traditional_Machine_Learning

This script will be comparing the traditional ML classification models and also classification using data derived from the Segment Anything Model within Kaza region in Southern Africa # Abstract The Kavango-Zambezi (KAZA) region which is apart of the Transfrontier Conservation Areas (TFCAs) aims at promoting biodiversity conservation and socio-economic development across borders. However, there has been a rapid increase in land transformation within the region through agricultural expansion, which has led to increased human-wildlife conficts (HWC) and threatening the objectives of these conservation areas.This thesis seeks to map and monitor such agricultural transformation with focus on the Binga District. The study leverages traditional Machine Learning (ML) classifcation and modern deep learning approaches, specifcally the Segment Anything Model (SAM) developed by Meta AI. Sentinel-2 and high-resolution PlanetScope satellite imagery were employed to classify and delineate agricultural felds within Binga District. AutoSklearn, an automated ML framework was used to classify and differentiate between agriculture and non-agricultural areas, highlighting challenges in binary classifcation and the high computational runtime required to optimize model performance and results. The results indicate that Sentinel-2’s multi-spectral capabilities contributed to a more accurate land cover classifcation. SAM was used to delineate agricultural felds using both satellite imagery, demonstrating the importance of spatial resolution in feld segmentation, with PlanetScope’s higher spatial resolution yielding better segmentation results. The study also explores the use of harmonic composites to potentially enhance the quality of feld segmentations. Overall, the integration of traditional ML and modern deep learning approaches in this study contributes to a deeper understanding of land-use change and illustrates how landcover classifcation can be enhanced by using multi-sensor and multi-algorithm approaches. Future studies can explore the integration of additional models and the temporal dynamics of land transformation to further refne the classifcation and se …