DETECTION AND RECOGNITION OF OBJECTS IN REMOTE SENSING IMAGES BASED ON YOLOV11

Abstract

This paper investigates modern methods for object detection and recognition in Earth remote sensing (RS) images, focusing on the challenges posed by large data volumes and complex, large-scale scenes where traditional algorithms prove inefficient. The study centers on the use of deep neural networks, with a primary focus on the YOLOv11 model, the latest architecture in the YOLO family known for balancing high detection accuracy and real-time processing capabilities. The research aims to enhance the efficiency of object detection by utilizing YOLOv11 and developing recommendations for its adaptation to high-resolution satellite and aerial imagery. A comprehensive comparative analysis is conducted, evaluating YOLOv11's architectural features—such as its improved modular backbone-neck-head, adaptive loss function, and enhanced multi-scale feature extraction—against its predecessors, YOLOv8 and YOLOv10, and competing models like DETR and Faster R-CNN. The study utilizes standard RS datasets such as DOTA, DIOR, and FAIR1M, with a specific focus on the persistent challenge of detecting small and oriented objects. The results obtained demonstrate that YOLOv11 achieves a significant improvement in mean Average Precision (mAP over the 0.5 to 0.95 threshold) of up to 5.3 percent compared to YOLOv8, while simultaneously reducing the number of parameters by 12 percent. Experimental comparisons also show that while transformer-based models like RF-DETR may achieve slightly higher accuracy, YOLOv11 is 3.5 times faster, processing up to 80 FPS, which is critical for real-time applications. The conducted studies and visual experiments confirm the effectiveness and universality of the YOLOv11 model for practical tasks, including area monitoring, environmental analysis, infrastructure object classification, and automated mapping.