基于Mask R-CNN的遥感影像土地分割与轮廓提取

传统的模式识别难以对土地遥感影像一次性精确统计分析.在精确分割出土地种类的前提下,本文提出了一种基于骨干网络为ResNet-101-RPN的Mask R-CNN的遥感影像土地分割与轮廓提取方法.该方法包括以下步骤：数据获取、图像去雾、遥感影像土地统计分析、土地分割和轮廓获取.在一个具有挑战性的卫星地图瓦片数据集上对所提出的方法进行训练和测试.实验结果表明,该方法以0.907的均值平均精度（mAP）和31.33像素的均值平均距离误差（mADE）获得了令人满意的不同种类土地分割和轮廓提取结果.

Land segmentation and contour extraction of remote sensing image based on Mask R-CNN

Urban spatial planning mainly relies on multiband remote sensing images.However,it is difficult for traditional pattern recognition methods to accurately analyze land covers from remote sensing images.On the premise of accurate land type classification,we propose a method of remote sensing image segmentation and contour extraction based on Mask R-CNN with ResNet-101-RPN as backbone network.The method includes the following steps:data acquisition,image defogging,statistical analysis of land covers on remote sensing image,land segmentation and contour acquisition.The proposed method is trained and tested on a challenging satellite tile dataset.Experimental results show that the method can obtain satisfactory land segmentation and contour extraction results by 0.907 of mean average precision (mAP) and 31.33 pixel of mean average distance error (mADE).