基于增强生成对抗网络的滑坡影像集超分辨率重建

针对高分遥感影像用于滑坡智能识别缺少高清训练集的问题,在组建高清滑坡训练集过程中,为充分利用低分辨率滑坡影像,本文采用基于增强型生成对抗网络模型(ESRGAN)实现了低分辨率滑坡影像集超分辨率重建。ESRGAN模型在SRGAN模型基础上,通过移除批归一化层、加入多级残差网络与残差缩放系数,提升了生成器的特征提取性能与稳定性,并采用迁移学习方法,基于毕节滑坡影像集与云南南景高速公路滑坡影像集进行试验验证。试验结果表明,基于迁移学习的ESRGAN模型在峰值信噪比(PSNR)与结构相似性(SSIM)方面获得更高得分,超分辨率重建取得更优结果。本文研究结果为获取滑坡高分辨率遥感影像集提供了一种新的技术方法。     

An Automated Approach for Land Cover Classification Based on a Fuzzy Supervised Learning Framework

This paper proposes an automatic framework for land cover classification. In majority of published work by various researchers so far, most of the methods need manually mark the label of land cover types. In the proposed framework, all the information, like land cover types and their features, is defined as prior knowledge achieved from land use maps, topographic data, texture data, vegetation’s growth cycle and field data. The land cover classification is treated as an automatically supervised learning procedure, which can be divided into automatic sample selection and fuzzy supervised classification. Once a series of features were extracted from multi-source datasets, spectral matching method is used to determine the degrees of membership of auto-selected pixels, which indicates the probability of the pixel to be distinguished as a specific land cover type. In order to make full use of this probability, a fuzzy support vector machine (SVM) classification method is used to handle samples with membership degrees. This method is applied to Landsat Thematic Mapper (TM) data of two areas located in Northern China. The automatic classification results are compared with visual interpretation. Experimental results show that the proposed method classifies the remote sensing data with a competitive and stable accuracy, and demonstrate that an objective land cover classification result is achievable by combining several advanced machine learning methods.     

迁移学习支持下的高分影像山地滑坡灾害解译模型

无人机低空遥感是近年来新兴的一种快速获取灾情信息的手段,如何利用无人机高分影像构建滑坡灾害解译模型是实现快速自动解译滑坡的关键。针对该问题,对比了多种影像特征提取方法,将迁移学习(TL)特征和支持向量机(SVM)引入到构建滑坡灾害自动解译模型中,提出了一种TL支持下的高分影像滑坡灾害解译模型。选取5·12汶川地震及4·20芦山地震系列无人机影像构建了滑坡灾害样本库并进行了实验,TL特征方法整体分类准确度ACC为95%,ROC达到0.98,识别准确率达到97%。结果表明,所提方法可用于高分影像滑坡自动解译,同时可用于大面积高分影像中快速山地滑坡灾害定位及检测。     

融合半监督学习的无监督遥感影像场景分类EI北大核心CSCD

自监督学习可以不依赖样本标签对遥感影像进行特征提取,但是特征分类仍然依赖有监督方法。为了克服有监督特征分类过程的不足,实现遥感影像特征的无监督自动分类,本文提出一种融合半监督学习的无监督语义聚类方法。首先,使用自监督学习提取遥感影像特征,抽象出图像包含的高层语义信息;然后,基于特征相似度寻找每个样本最相似的近邻,使用在线聚类将相似样本聚为一类,训练一个线性分类器;最后,根据聚类结果为高置信度样本生成伪标签,构造标注样本集,使用半监督方法对模型微调。在4个公开遥感影像场景分类数据集EuroSAT、GID、AID和NWPU-RESISC45上进行验证,分类精度分别达到了94.84%、63.55%、76.42%和86.24%。本文方法结合了在线聚类和半监督学习的优点,缓解了已有方法存在的误差积累和样本利用不充分的问题,在完全不使用标注样本的情况下,充分利用自监督特征训练分类模型,对遥感影像进行场景分类,达到接近有监督学习的分类效果,具有良好的应用价值。     

高分辨率遥感影像解译中的机器学习范式

高分辨率遥感影像解译是遥感信息处理领域的研究热点之一,在遥感大数据知识挖掘与智能化分析中起着至关重要的作用,具有重要的民用和军事应用价值.传统的高分辨率遥感影像解译通常采用人工目视解译方式,费时费力且精度低.所以,如何自动、高效地实现高分辨率遥感影像解译是亟待解决的问题.近年来,随着人工智能技术的飞速发展,采用机器学习...     

一种用于遥感影像语义分割对抗域适应方法

为利用已有标注的影像数据集实现对未标注遥感影像的语义分割,提出一种对抗域适应的方法.首先在生成对抗网络的基础上,利用基于香农熵的不确定图,进行域间的对抗学习,实现已标注的影像(源域)与未标注的影像(目标域)之间的迁移学习;其次为进一步提升模型的无监督学习能力,使用基于伪标签提纯的自学习策略.为验证所提方法的有效性,使用...     

顾及样本敏感性的滑坡易发性评价

滑坡作为一种危害极大的自然地质现象,严重威胁着人民的生命财产安全。因此,科学、准确地评价滑坡体的易发性至关重要。随着机器学习的发展,基于机器学习的滑坡易发性评价逐渐成为研究热点。而在真实情况中,滑坡区域与非滑坡区域面积占比悬殊,这使得机器学习模型的应用存在较严重的样本不均衡问题。本文采用样本敏感性分析方法,综合多个机器学习模型在不同比例的正负滑坡样本集上的表现,以获取最均衡滑坡样本集;并在此样本集基础上采用深度随机森林模型,在示范研究区开展滑坡易发性评价。最终的评价结果接近真实分布,表明本文方法具有较好的有效性。     

生成式对抗网络的高光谱遥感图像分类方法研究

针对基于深度学习的分类模型在训练样本较少时所遭受的潜在过拟合问题,提出一种具备过拟合抑制的生成式对抗网络分类算法,并应用于高光谱图像分类.该算法在每次迭代时,首先,依据训练样本的标签信息使判别器网络拟合训练样本的数据分布;然后对训练样本的高维特征进行均值最小化,该过程会重新更新判别器网络参数,减小参数的值和方差,以抑制...     

基于局域自适应信息理论测度学习的高光谱目标探测方法

传统基于信号检测的目标探测方法需要依赖特定的统计假设,只有在符合条件的情况下才能取得较好的目标探测结果。为了克服这一缺陷,提出了一种基于局域自适应的信息理论测度学习方法。首先将信息理论测度学习方法作为目标主函数,然后加以局域自适应决策法则进行约束,自适应地减小相似样本对距离,增大不相似样本对距离,使得在考虑阈值的同时兼顾测度学习前后距离的改变情况来进行目标探测决策,从而更好地实现目标探测。实验证明,该方法与其他经典目标探测方法或测度学习方法相比,可以更好地实现目标与背景分离,能够更有效地对高光谱影像数据进行目标探测。     

结合对抗网络与辅助任务的遥感影像无监督域适应方法

使用机器学习进行遥感影像标注的一个重要前提是有足够的训练样本,而样本的标注是非常耗时的。本文采用了域适应的方法来解决遥感影像场景分类中小样本量的无监督学习问题,提出了结合对抗网络与辅助任务的遥感影像域适应方法。首先建立了基于深度卷积神经网络的遥感影像分类框架;其次,为了学习到域不变特征,在标签分类器的基础上增加域分类器,并使域损失函数在其反射传播时的梯度与标签损失的梯度相反,从而保证域分类器不能区分样本来自于哪个域;最后引入了辅助分类任务,扩充了样本的同时使网络更具泛化能力。试验结果表明,本文方法优于主流的无监督域适应方法,在小样本遥感影像无监督分类中得到了较好的效果。     

火烧迹地信息遥感快速提取方法研究

利用福建省福州地区ETM图像，在分析火烧迹地光谱特征的基础上，研究了火烧迹地信息遥感快速提取方法。文中设置了4种提取方案，即监督分类法、NDVI参与的监督分类法、NDVI差值法及谱间关系法，最后，对提取结果进行了精度评价。     

车辆轨迹数据的道路学习提取法

车辆轨迹数据的道路信息提取是地理信息领域的热点也是难点之一,深度学习的快速发展为该问题的解决提供了一种思路与方法。本文针对车辆轨迹数据的车行道级道路提取问题,引入深度学习领域的生成式对抗网络,利用残差网络构建深层网络和多尺度感受野感知轨迹数据不同细节特征,构建了基于条件生成式对抗网络的轨迹方向约束下车行道级道路提取模型。首先提出了朝向-颜色映射栅格化转换方法,实现轨迹朝向信息向HSV颜色空间的转换;然后利用样本数据学习模型参数;最后将训练模型应用到郑州、成都、南京3个试验区域提取车行道级道路数据。试验结果表明,本文方法能够有效地提取完整的车行道级道路数据。     

农作物叶片病害迁移学习分步识别方法

农作物病害的准确识别有助于提高农作物的产量与质量。针对作物病害图像训练样本数据量较少的情况,本文采用迁移学习方法结合分步识别模型对多种农作物病害种类进行了识别。将PlantVillage公开数据集中3种农作物的10类作物叶片图像作为训练样本,利用直接识别方法分别对VGG16和ResNet的原始模型和迁移学习模型进行训练,得到模型的分类结果;提出分步识别方法,将训练样本按作物种类和病害类型归类,分别进行模型训练并构建分步识别模型。试验结果表明：利用迁移学习方法能够在原始模型的基础上将识别精度提高20%以上;在其基础上引入分步识别方法并与直接识别方法对比,能够将VGG16和ResNet模型精度再分别提高14%和8%。本文提出的迁移学习分步识别方法能够实现在小样本训练数据情况下的作物病害准确识别,可为作物病害防治提供有效的技术支持。     

遥感影像居民地目标Bayesian网络识别算法研究

利用Naive Bayesian网络的学习和推理机制，提出一种在遥感影像上提取居民地目标的方法。该方法通过对所选取的正负样本进行学习，获取Bayesian网络的重要参数，即条件概率和概率分布密度。在此基础上，根据正负样本所构建的条件概率网，对未知类别信息的影像进行分类，从而获取居民地目标的信息。通过对实际全色SPOT5影像中居民地目标的提取，表明该方法具有较高的识别率。     

Landslide Mapping and Assessment by Integrating Landsat-8, PALSAR-2 and GIS Techniques: A Case Study from Kelantan State,Peninsular Malaysia

Integration of satellite remote sensing data and GIS techniques is an applicable approach for landslide mapping and assessment in highly vegetated regions with a tropical climate. In recent years, there have been many severe flooding and landslide events with significant damage to livestock, agricultural crop, homes, and businesses in the Kelantan river basin, Peninsular Malaysia. In this investigation, Landsat-8 and phased array type L-band synthetic aperture radar-2 (PALSAR-2) datasets and analytical hierarchy process (AHP) approach were used to map landslide in Kelantan river basin, Peninsular Malaysia. Landslides were determined by tracking changes in vegetation pixel data using Landsat-8 images that acquired before and after flooding. The PALSAR-2 data were used for comprehensive analysis of major geological structures and detailed characterizations of lineaments in the state of Kelantan. AHP approach was used for landslide susceptibility mapping. Several factors such as slope, aspect, soil, lithology, normalized difference vegetation index, land cover, distance to drainage, precipitation, distance to fault, and distance to the road were extracted from remotely sensed data and fieldwork to apply AHP approach. The excessive rainfall during the flood episode is a paramount factor for numerous landslide occurrences at various magnitudes, therefore, rainfall analysis was carried out based on daily precipitation before and during flood episode in the Kelantan state. The main triggering factors for landslides are mainly due to the extreme precipitation rate during the flooding period, apart from the favorable environmental factors such as removal of vegetation within slope areas, and also landscape development near slopes. Two main outputs of this study were landslide inventory occurrences map during 2014 flooding episode and landslide susceptibility map for entire Kelantan state. Modeled/predicted landslides with a susceptible map generated prior and post-flood episode, confirmed that intense rainfall throughout Kelantan has contributed to produce numerous landslides with various sizes. It is concluded that precipitation is the most influential factor for landslide event. According to the landslide susceptibility map, 65% of the river basin of Kelantan is found to be under the category of low landslide susceptibility zone, while 35% class in a high-altitude segment of the south and south-western part of the Kelantan state located within high susceptibility zone. Further actions and caution need to be remarked by the local related authority of the Kelantan state in very high susceptibility zone to avoid further wealth and people loss in the future. Geo-hazard mitigation programs must be conducted in the landslide recurrence regions for reducing natural catastrophes leading to loss of financial investments and death in the Kelantan river basin. This investigation indicates that integration of Landsat-8 and PALSAR-2 remotely sensed data and GIS techniques is an applicable tool for Landslide mapping and assessment in tropical environments.     

Effects of Landslide Sampling Strategies on the Prediction Skill of Landslide Susceptibility Modelings

In this study, landslide susceptibility assessments were achieved using logistic regression, in a 523 km² area around the Eastern Mediterranean region of Southern Turkey. In reliable landslide susceptibility modeling, among others, an appropriate landslide sampling technique is always essential. In susceptibility assessments, two different random selection methods, ranging 78–83% for the train and 17–22% validation set in landslide affected areas, were applied. For the first, the landslides were selected based on their identity numbers considering the whole polygon while in the second, random grid cells of equal size of the former one was selected in any part of the landslides. Three random selections for the landslide free grid cells of equal proportion were also applied for each of the landslide affected data set. Among the landslide preparatory factors; geology, landform classification, land use, elevation, slope, plan curvature, profile curvature, slope length factor, solar radiation, stream power index, slope second derivate, topographic wetness index, heat load index, mean slope, slope position, roughness, dissection, surface relief ratio, linear aspect, slope/aspect ratio have been considered. The results showed that the susceptibility maps produced using the random selections considering the entire landslide polygons have higher performances by means of success and prediction rates.     

基于自动搜索和光谱匹配技术的训练样本纯化算法

提出了一种基于局部自动搜索和光谱匹配技术的监督分类训练样本的纯化方法。该方法首先利用遥感影像中像元的灰度信息在图像上局部范围内自动搜索和选择最佳样区位置，然后利用光谱匹配的思想对寻找到的最佳样区在光谱空间上进一步纯化。实验结果证明，通过手工选择样区的辅助，该算法能够自动有效地搜寻到最佳样区的位置，并对最佳样区进行纯化处理。原始遥感图像经过本文的样区纯化算法处理后，无论是目视判读效果，还是分类后混淆矩阵的统计及分类精度，均优于纯化处理前的分类结果，具有一定的实用价值。     

An adaptive unimodal subclass decomposition (AUSD) learning system for land use and land cover classification using high-resolution remote sensing

Land use and land cover classification is an important application of remote-sensing images. The performances of most classification models are largely limited by the incompleteness of the calibration set and the complexity of spectral features. It is difficult for models to realize continuous learning when the study area is transferred or enlarged. This paper proposed an adaptive unimodal subclass decomposition (AUSD) learning system, which comprises two-level iterative learning controls: The inner loop separates each class into several unimodal Gaussian subclasses; the outer loop utilizes transfer learning to extend the model to adapt to supplementary calibration set collected from enlarged study areas. The proposed model can be efficiently adjusted according to the variability of spectral signatures caused by the increasingly high-resolution imagery. The classification result can be obtained using the Gaussian mixture model by Bayesian decision theory. This AUSD learning system was validated using simulated data with the Gaussian distribution and multi-area SPOT-5 high-resolution images with 2.5-m resolution. The experimental results on numerical data demonstrated the ability of continuous learning. The proposed method achieved an overall accuracy of over 90% in all the experiments, validating the effectiveness as well as its superiority over several widely used classification methods.     

中分辨率遥感图像土地利用与覆被分类的方法及精度评价

利用TM、SPOT及CBERS-1等中分辨率卫星图像,对土地覆被的专家系统分类方法、居民地的决策树提取方法以及水体的迭代混合提取方法进行了试验,其总体精度达到87.89%,与常用的监督分类方法相比精度可提高7.86%.专家系统分类的结果叠合居民地、水体等易于混分专题信息,可以形成精度较高的土地利用与覆被分类结果。     

一种基于迁移学习的栅格型图表自动分类方法

图表自动分类是实现图表内容解译及信息提取的前提.基于Inception V3模型,采用迁移学习的方法,对图表自动分类方法进行了研究,提出了对5类栅格图表进行自动分类的模型.通过与传统的图表自动分类方法的对比试验,测试了在不同学习率参数下模型的分类效果,给出了针对结构化特征的指数衰减学习率的数学表达式.结果表明:采用迁移...