基于视觉提示词的耕地实例化分割方法

Instance Segmentation Method of Cultivated Land Based on Visual Prompts

耕地是国家粮食安全的根本保证,基于遥感影像的耕地提取为耕地保护,遏制耕地“非农化”,防止“非粮化”提供基本技术支撑。为实现基于高分遥感影像的耕地精细化提取,基于目标检测技术和神经网络大模型(segment anything model,SAM),实现简单标注样本下的耕地实例化分割,提出了一种基于视觉提示词的耕地实例化分割方法(instance segmentation method of farmland based on visual prompts,ISFVP)。在提出的方法中,首先利用耕地目标检测网络检测出耕地的大概范围(以矩形框表示),然后将耕地所处的范围作为提示词输入到网络大模型SAM中,实现耕地范围的精细化提取。实验证明,该方法可从高分遥感影像中自动化检测出耕地,为耕地的精细化提取提供了一种可行的技术方案。

Cultivated land is the fundamental guarantee of national food security.Cultivated land extraction based on remote sensing images provides basic technical support for the protection of cultivated land,to stop any attempt to use it for purposes other than agriculture and specifically grain production.In order to realize the refined extraction of cultivated land based on high-resolution remote sensing images,we used the target detection technology and the large neural network segment anything model(SAM)to realize the instance segmentation of cultivated land under simple labeled samples,and proposed an instance segmentation method of cultivated land based on visual prompts.In the proposed method,we used the cultivated land target detection network to detect the approximate range of cultivated land(represented by a rectangular box)at first.And then,we input the range of cultivated land into the large network model SAM as a prompt word to realize the refined extraction of cultivated land range.Experimental result proves that this method can automatically detect cultivated land from high-resolution remote sensing images,and provide a feasible technical solution for the refined extraction of cultivated land.