Abstract: Greenhouse serve as critical infrastructure in modern agriculture,where precise monitoring holds significant implications for agricultural modernization and food security.However,traditional segmentation and vectorization methods exhibit substantial limitations in high-precision boundary extraction due to complex optical characteristics and gradient transition properties of greenhouse edges.This study proposes a novel boundary extraction framework integrating gradient feature learning and geographic active contour modeling.Initially,a Vision Transformer based pretrained encoder extracts high-dimensional image features,while a multi-task segmentation decoder concurrently generates mask,edge,and gradient representations.Subsequently,a gradient field construction model guides the vectorization process,coupled with geographic active contour-based postprocessing to significantly enhance boundary smoothness and vectorization accuracy.Experimental results demonstrate superior performance over conventional vectorization methods in metrics including intersection over union (IoU)and maximum angular error,particularly excelling in complex geographic contour extraction tasks.This framework provides an innovative solution for remote sensing monitoring in facility agriculture.

Key words: gradient features, deep learning, geographic feature segmentation, vectorization