국토위성영상 이용한 YOLO-Seg 기반의 농경지 추출 모델 개발

Abstract

This study developed an agricultural parcel detection model based on the YOLO-Seg algorithm using high-resolution satellite imagery of CAS500-1(Compact Advanced Satellite 500-1). We analyzed the effects of grid overlap and confidence threshold settings on object detection accuracy. The mainresults showed that a 60% grid overlap achieved optimal performance with 70% detection accuracy and about 5% false detection rate within one minuteof processing time. By analyzing ratios of overlapping from 10% to 90%, the detection accuracy improved, but the processing time significantlyincreased, requiring a balance between accuracy and efficiency. The proposed method of spatial continuity correction using polygon centroids improveddetection accuracy by approximately 5%p while significantly reducing false detection rates. The system demonstrated 80% detection accuracy when theconfidence threshold was set between 0.5 and 0.7, which outperformed traditional detection methods. Additionally, our approach addressed the boundarydetection issues in split grid analysis through an optimized overlap strategy. This automated detection system can be effectively applied to agriculturalmonitoring, resource management, and policy development, providing a cost-effective solution for large-scale agricultural parcel monitoring.