Landscape Structure and Fragmentation Indices for Site Selection in Aerial Object Detection Training

This study investigates an optimized site selection process for training object detection models by leveraging segmentation images. The approach ensures the preservation of landscape morphology variability while minimizing redundancy caused by similar traits across dataset regions. To achieve this, we integrate landscape structure and fragmentation indices for each image and employ Principal Component Analysis (PCA) to identify and select areas exhibiting the highest geographic variability. We test our hypothesis by training and comparing two YOLOv8 models: one utilizing data from all dataset areas and another utilizing training data exclusively from the PCA-selected high-variability areas and validation data from the low-variability areas. The findings indicate promising avenues for future research that integrate Landscape Structure and Fragmentation Indices with Remote Sensing techniques.