面向对象结合深度学习方法的矿区地物提取

为了快速准确获取煤炭矿区的地物信息，以达到辅助安排和部署矿区安全生产工作的目的，采用无人机低空遥感拍摄的方式获取了矿区内的高清影像数据，并提出一种基于面向对象和深度学习的矿区无人机影像地物提取方法。首先利用面向对象的分类方法配合人工校正，制作用于深度学习语义分割的标签，再采用 FCN － 32s，FCN －8s 和 U － Net 3 种深度学习语义分割模型提取图像特征，训练出 3 种不同的分类模型，并基于此提出多数投票和打分算法 2 种集成模型改进地物提取精度。实验结果表明，面向对象结合深度学习方法的地物提取准确率、Kappa 系数较传统面向对象方法均有明显提升。其中打分集成模型识别效果最好，在测试集上的整体准确率为 94． 55% ，高出面向对象方法 5． 96 百分点; Kappa 系数为 0． 819 1。

To quickly and accurately obtain ground feature information of coal mining areas and assist in arranging and deploying safe production work in the mining zones, we collected high-resolution image data within the target mining area via low-altitude unmanned aerial vehicle (UAV) remote sensing photography, and proposed a ground feature extraction method for mining-area UAV images based on object-oriented technology and deep learning. First, we generated labels for deep learning semantic segmentation using an object-oriented classification method combined with manual correction. Subsequently, three deep learning semantic segmentation models, namely FCN-32s, FCN-8s and U-Net, were employed to extract image features, and three separate classification models were trained. On this basis, two ensemble models, namely majority voting and scoring algorithm, were put forward to improve the accuracy of ground feature extraction. The experimental results show that both the ground feature extraction accuracy and Kappa coefficient of the method combining object-oriented technology and deep learning are significantly improved compared with the traditional object-oriented method. Among them, the scoring ensemble model achieved the best recognition performance, with an overall accuracy of 94.55% on the test set, which is 5.96 percentage points higher than that of the single object-oriented method, and a Kappa coefficient of 0.8191.

搜集汇总

背景与挑战

背景概述

该数据集采用无人机低空遥感拍摄获取矿区高清影像数据，结合面向对象和深度学习方法提取地物信息，旨在辅助矿区安全生产工作的安排和部署。实验结果表明，该方法的地物提取准确率和Kappa系数较传统方法有明显提升，其中打分集成模型在测试集上的整体准确率达到94.55%。