【作者】 姚远；

【导师】 李林宜；

【作者基本信息】 武汉大学 ， 模式识别与智能系统， 2022， 硕士

【摘要】 内河洲滩是河湖水系中存在的重要陆地区域,对河湖水系生态环境建设具有重要意义。利用遥感技术进行内河洲滩及其建筑物的智能识别,是当前水利遥感及相关研究领域所需解决的重要课题,具有重要的理论与实际应用价值。然而,由于遥感影像中内河洲滩及其地物的复杂性与多样性,内河洲滩及其建筑物的智能识别仍存在较大困难。本论文从遥感影像中智能识别内河洲滩及其建筑物,主要研究内容及相关成果如下:(1)针对遥感影像内河洲滩识别难度大及精度不高的问题,本文利用遥感影像中内河洲滩的形态特征及光谱特征,提出了一种联合遥感特征指数与增强型XGBoost的内河洲滩遥感智能识别方法。遥感特征指数MNDWI和NDVI对于水体及植被目标具有高敏感性,有利于洲滩识别;极端梯度提升算法(XGBoost)计算速度快、分类精度高,本文在普通型XGBoost的基础之上引入了交叉验证及网格搜索参数优化方法,形成了增强型XGBoost,实现了内河洲滩遥感智能识别。以Landsat8、GF-1卫星影像为实验数据,与决策树(Decision Tree)、随机森林(Random Forest)、普通型XGBoost等多种方法进行了对比分析,验证了本文方法的有效性。以Landsat8 OLI长江研究区实验数据为例,内河洲滩识别的总体精度与kappa系数分别达到了96.0%与0.920,比普通型XGBoost分别高出了2%和0.041,本文方法具有更高的内河洲滩识别精度,识别结果完整、准确。(2)针对内河洲滩建筑物特点,本文利用遥感影像建筑物空间形态特征,对传统U-net结构加以改进,融合了残差结构、密集连接及空洞卷积结构,提出了适用于内河洲滩建筑物识别的改进型U-net网络。以FCN16S、FCN8S、Segnet、U-net为对比方法,利用Massachusetts及WHU建筑物数据集,进行建筑物识别实验,验证了本文方法的有效性。本文还利用GF-2遥感影像,进行了算法迁移实验。本文方法在Massachusetts及WHU数据集上的识别结果的总体精度分别为92.37%、97.00%,m Io U分别为78.07%、88.85%,优于传统方法。在基于GF-2影像的迁移实验中,本文方法也取得了很好的建筑物识别结果,识别精度高,建筑物边缘识别准确,对小型建筑物敏感度高,且具有良好的可迁移性。更多还原

【Abstract】 Inland river bottomland is the important land area in the river and lake water system,and therefore the construction of the ecological environment of the river and lake water system is of great significance.The application of remote sensing technology for the intelligent recognition of the inland river bottomland and its buildings has important theoretical and practical application value,which is an important topic in the current water conservancy remote sensing and related research field.However,due to the complexity and diversity of inland river bottomland in remote sensing imagery,the intelligent recognition of inland river bottomland and its buildings is still difficult.This paper focuses on the intelligent recognition of the inland river bottomland and its buildings from remote sensing imagery,and the research content and related achievements of the paper mainly include as follows:(1)In view of the difficulty and low accuracy of inland river bottomland recognition in remote sensing imagery,this paper proposes a remote sensing intelligent recognition method of inland river bottomland by combining remote sensing characteristic indexes and enhanced XGBoost.The remote sensing characteristic indexes MNDWI and NDVI are highly sensitive to water and vegetation targets,which are conducive to bottomland recognition.Extreme gradient boosting(XGBoost)has fast calculation speed and high classification accuracy,based on which,this paper introduces the parameter optimization method of cross validation and grid search to form enhanced XGBoost.Taking landsat8 and GF-1 satellite images as experimental data,the effectiveness of this method is verified by comparing with many methods such as decision tree,random forest and ordinary XGBoost.Taking the experimental data of Landsat8 OLI Yangtze River research area as an example,the overall accuracy and kappa coefficient of inland river bottomland recognition are96.0% and 0.920 respectively.Compared with ordinary XGBoost,it is 2% higher and0.041 higher respectively.Therefore,the proposed method has higher inland river bottomland recognition accuracy,and the recognition results are complete and accurate.(2)In view of the characteristics of inland river bottomland buildings,this paper proposes an improved U-net network suitable for building recognition in inland river bottomland,which makes full use of the spatial morphological characteristics of buildings in remote sensing imagery and improves the basic U-net structure and integrates residual structures,dense connections and dilated convolution structures.Taking FCN16 s,FCN8s,Segnet and U-net as comparison methods,building recognition experiments are carried out based on Massachusetts and WHU building datasets to verify the effectiveness of the proposed method.In addition,this paper performs migration experiments using GF-2 remote sensing imageries.The overall accuracy of the recognition results of proposed method on Massachusetts and WHU datasets are 92.37% and 97.00% respectively,and m Io U are 78.07% and 88.85%respectively,which is better than the traditional method.In the GF-2 remote sensing imagery-based migration experiment,the proposed method also achieved good building recognition results,which has high recognition accuracy,accurate building edge recognition results,high sensitivity to small buildings and good portability.更多还原