Developing approaches in building classification and extraction with synergy of YOLOV8 and SAM models

The ability to extract meaningful information from visual material, such as photographs and videos, has significantly enhanced the potential for object recognition in various disciplines. However, challenges arise in the geospatial domain while features are extracted. Existing approaches primarily focus on remotely sensed images, emphasizing semantic segmentation tasks. This study, in contrast, prioritizes the extraction of buildings as well as the classification of the structures into residential and non-residential types using instance segmentation. The proposed model pipeline combines the YOLOV8 detection with the Segment Anything Model algorithm to achieve these objectives. The approach outlined in this research produces segmentation outcomes that align with evaluation metrics, comparable to those achieved by earlier instance segmentation methods and the segmentation strategies utilized for assessing building extraction performance. Additionally, the segmentation results are georeferenced using extracted geospatial information, and vector images of the identified building rooftops are generated. The approach demonstrates robustness in effectively segmenting target objects, regardless of diverse characteristics like shape, size, or orientation. The model pipeline exhibits superior precision (0.929), recall (0.838), and mean average precision (0.899) values. Moreover, the model produces results approximately 50% faster in terms of inference time compared to other instance segmentation models. The proposed model pipeline holds significant applicability in valorous fields, including urban planning, transportation planning, urban development.