Geospatial Object Detection in High-Resolution Satellite Images Using an Improved Mayfly Optimization Algorithm Based on Long Short-Term Memory

Geospatial object detection in high-resolution satellite images has several applications, including environmental monitoring, military surveillance, marine monitoring, and urban planning. High Spatial and Spectral Resolution (HSSR) satellite imagery provides detailed ground-level information; however, research on this type of remote-sensing data remains limited, while object detection using such imagery remains a challenging task. In an attempt to provide effective object detection classification, this study utilizes an Improved Mayfly Optimization Algorithm based on Long Short-Term Memory (IMO-LSTM) framework trained and tested on the Dataset for Object Detection in Aerial images (DOTA). For feature extraction, Histogram of Oriented Gradients (HOG), Local Gabor Binary Pattern Histogram Sequence (LGBPHS), and Harris Corner Detection (HCD) were utilized, followed by IMO-based feature selection to retain the most relevant features. Lastly, the LSTM network is applied to classify objects in the selected images. When compared with existing methods such as Critical Feature-Capturing Network (CFC-Net), Self-adaptive Aspect Ratio Anchor (SARA), Cropping Region Proposal Network-based Scale Folding Network (CRPN-SFNet), and Ground Sample Distance (GSD), the proposed method achieved the best performance, attaining a mean Average Precision (mAP) of 89.53%.