AI-Based Garbage Segmentation from Aerial Drone Images Using YOLOV8

Abstract

The YOLOv8n-seg model has been used for AI garbage segmentation from drone images. The focus is on urban waste monitoring. The model achieved strong results for detection and segment performance with a box precision of 0.821, a recall of 0.656 and a mask mAP50 of 0.581 on the valid set. This metric reveals the capability of segmenting and localizing the trash correctly in the complex visual environment. The confusion matrix showed that the model performed very well in the classification of the background. However, there was one case where the background was incorrectly categorized as "garbage. This means that we need to further fine-tune the model so that both these classes are well-separated. Qualitative assessments demonstrated that in different aerial images, the garbage and other solid wastes were able to be detected with a confident prediction, although there were some false alarms for vegetation and textured surfaces. The loss, accuracy and recall measure have increased continually, indicating that the model continues to learn consistently as it is trained further.  Looking at the recall/precision curves, it can also be seen that detection accuracy and completeness were not scalable when the confidence threshold is set to a higher value, as the highest F1 score value is 0.73 when the threshold is the lowest.  Although the model has not yet been used on actual UAVs, it also showed the potential for real-time segmentation of video simulations, potentially offering integration in intelligent waste management systems and drone logistics.  Future work is recommended that will reduce false positives and help distinguish background. As a result, the research supports lightweight YOLOv8 segmentation models as viable contenders for intelligent automated rubbish monitoring activities.