Comparison of residual and dense neural network approaches for building extraction from high-resolution aerial images

Applications including change detection, disaster management, and urban planning require precise building information, and therefore automatic building extraction has become a significant research topic. With the improvements in sensor and satellite technologies, more data has become available, and with the increased computational power, deep learning methods have emerged as successful tools. In this study, U-Net and FPN architectures using four different backbones (ResNet-50, ResNeXt-50, SE-ResNext-50, and DenseNet-121), and an Attention Residual U-Net approach were used for building extraction from high-resolution aerial images. Two publicly available datasets, Inria Aerial Image Labeling Dataset and Massachusetts Buildings Dataset were used to train and test the models. According to the results, Attention Residual U-Net model has the highest F1 score with 0.8154, IoU score with 0.7102, and test accuracy with 94.51% on the Inria dataset. On the Massachusetts dataset, FPN Dense-Net-121 model has the highest F1 score with 0.7565 and IoU score with 0.6188, and Attention Residual U-Net model has the highest test accuracy with 92.43%. It has been observed that, FPN with DenseNet backbone can be a better choice when working with small size datasets. On the other hand, Attention Residual U-Net approach achieved higher success when a sufficiently large dataset is provided.     

Object-based rice mapping using time-series and phenological data

Remote sensing techniques are often used in mapping rice, but high quality time-series remote sensing data are difficult to obtain due to the cloudy weather of rice growing areas and long satellite revisit interval. As such, rice mapping is usually based on mono-temporal Landsat TM/ETM+ data, which have large uncertainties due to the spectral similarity of different vegetation types. Moreover, conventional pixel-based classification method is unable to meet the required accuracy for rice mapping. Therefore, this study proposes a new strategy for mapping rice in cloud-prone areas using fused data of Landsat-8 OLI time-series and phenological parameters, based on the object-based method. We determine the critical growth stages of paddy rice from observed phenological data and MODIS-NDVI time-series data. The spatial and temporal adaptive reflectance fusion model (STARFM) is used to blend the MODIS and Landsat data to obtain a multi-temporal Landsat-like dataset for classification. Finally, an object-oriented algorithm is used to extract rice paddies from the Landsat-like, time-series dataset. The validation experiments show that the proposed method can provide high accuracy rice mapping, with an overall accuracy of 92.38% and a kappa coefficient of 0.85.     

基于迁移学习的暴恐图像自动识别

利用人工智能和深度学习技术自动化地分析互联网海量图片,快速、准确地识别有害的暴恐图像并及时处置是反恐工作的重要手段之一。研究了利用深度学习和迁移学习技术对暴恐图像进行分类识别。首先,定义了暴恐图像的主要概念特征,并针对性地构建数据集;其次,针对暴恐图像正样本较少的问题,设计深度神经网络模型和迁移学习方式;最后,基于构建的训练数据集进行模型训练和测试。结果显示:所提方法可以快速、准确地对互联网图片进行分类识别,平均分类准确率达到96.7%,从而有效降低人工检测的劳动强度,为反恐预警工作提供决策支持。      

Estimation of sediment settling velocity in estuarine and coastal waters using optical remote sensing data

A robust method has been developed for estimating sediment settling velocity (w_s) from high resolution optical remote sensing data in estuarine, coastal and harbor waters. This method estimates settling velocity as a function of the drag coefficient (C_d), Reynolds number (Re), grain size (D₅₀), specific gravity (Δ_SG) and grain shape (in terms of the Corey Shape Factor – CSF). These parameters were derived from the particulate inherent optical properties such as backscattering (b_bp), beam attenuation (c_p), suspended sediment concentration and turbidity using Landsat 8 OLI and HICO data. Preliminary results for the Gulf of Cambay in the eastern Arabian Sea and Yangtze river estuary in the East China Sea, showed that satellite-retrieved settling velocities (m?s^?1) varied from very low values in clear oceanic waters, intermediate values in coastal waters, to very high values in river plumes and sediment-laden coastal waters. The remote sensing retrievals of sediment properties and their settling velocities were generally consistent with the field and laboratory results, which indicate that the proposed methodology will have important implications in various coastal engineering, environmental and management studies.     

基于对抗和迁移学习的灾害天气卫星云图分类

针对卫星云图中的灾害天气数据存在严重不平衡问题,提出一个结合生成对抗学习（GAN）和迁移学习（TL）的卷积神经网络（CNN）框架以解决上述问题进而提高基于卫星云图的灾害天气分类精度。该框架主要包含基于GAN的数据均衡化模块和基于迁移学习的CNN分类模块。上述2个模块分别从数据和算法层面解决数据的类间不平衡问题,分别得到一个相对均衡的数据集和一个可在不同类别数据上提取相对均衡特征的分类模型,最终实现对卫星云图的分类,提高其中灾害天气的卫星云图类别分类准确率。与此同时所提方法在自建的大规模卫星云图数据上进行了测试,消融性和综合实验结果证明了所提数据均衡方法和迁移学习方法是有效的,且所提框架模型对各个灾害天气类别的分类精度都有显著提升。      

基于卫星遥感的岛礁影像多类目标智能化提取

卫星遥感影像具有背景复杂、目标尺度不一、观测方向各异、纹理不清晰等特点,主流的深度学习目标检测算法不能直接适用于卫星遥感影像的目标检测。改进了RetinaNet,使其适用于卫星遥感影像。首先设计了一种新的特征融合方式,融合ResNet50输出的特征图,使得融合后的特征图同时具有高层语义信息和低层纹理细节信息。为了减弱遥感影像复杂背景对目标特征的影响,设计了特征感知模块,在减弱噪声对特征图影响的同时增强有用特征。挑选DOTA数据集中船只、飞机和存储罐图像进行训练和测试。改进的算法与RetinaNet相比,飞机、船只和存储罐的平均精度分别提高了41%、25%、24%。基于高分二号卫星（GF 2）真实影像数据的试验结果表明,提出的算法能够用于卫星遥感岛礁影像的多类目标智能化提取。     

基于超像素与多尺度残差U-Net相结合的遥感图像飞机检测方法

遥感图像中存在飞机很小、角度和位置不确定且背景复杂等问题,从遥感图像中检测飞机是一项重要且具有挑战性的任务,因此,提出一种基于超像素与多尺度残差U-Net(Multi-scale Residual U-Net,MSRU-Net)相结合的遥感图像飞机检测方法。首先对遥感图像进行超像素预分割,将位置相邻且像素特征相似的像素点组成若干个超像素,保持图像进一步分割的有效特征;然后构建多尺度残差U-Net,学习其多尺度判别特征。与传统的飞机检测方法相比,该方法用少量的超像素代替大量像素表达图像特征,降低了图像分割的复杂度,再利用MSRU-Net分割遥感超像素图像,有效检测不同尺度的飞机图像。在公共飞机遥感图像数据集上实验,结果表明,该方法能够有效的检测遥感图像不同尺度的飞机图像,检测精确率达到91.2 %。     

基于U-net的紫外极光观测极光卵形态提取

极光卵形态提取是极光研究的重要手段.如何提高强干扰背景下的紫外极光图像极光卵形态提取精度,目前仍是一个难题.本文提出一种基于深度学习语义分割模型U-net的方法,实现了对极光卵形态的高精度提取.在Polar卫星紫外极光观测数据的实验结果表明,该方法相比于已有算法精度更高,对完整型极光卵和缺口型极光卵图像均能得到更加精确的提取结果,特别是针对强日辉干扰、灰度不均匀和对比度低情况下的紫外极光图像时,该方法显示了明显优势.      

太阳活动区EUV图像的生成式模型耀斑分级与预报

近年来,不断发射的空基观测台持续传送回海量日面图像及日地间气象数据,为采用人工智能技术对太阳活动进行预报预警提供了数据基础。但是,极端天气爆发少,样本量较少;中等程度爆发稍多,样本量较多;常规无爆发天气常见,样本较为集中,样本不均衡状况严重影响机器学习方法在空间天气领域的广泛应用。本文面向多源多通道多尺度日面图像信息,构建了来自SOHO和SDO的1996-2015年日面活动区图像数据集;针对数据分布的不平衡,对太阳活动区图像作耀斑分级与预报。在对比分析元学习算法的基础上,设计了结合分类头设计和卷积核初始化的生成式模型;在使网络轻量化的基础上,能够将M和X级耀斑预报的检测率指标相较于普通的深度学习模型和无监督度量式模型分别提升10%和7%。     

结合多层特征及空间信息蒸馏的医学影像分割

U-Net在医学影像分割领域是目前应用最广泛的分割模型，其“编码-解码”结构也成为了构建医学影像分割模型最常用的结构。尽管U-Net在许多领域实现了非常高的分割准确度，但是存在着计算复杂度高、推理速度慢、运行消耗内存大等问题，导致其难以在移动应用平台部署。为解决这一问题，提出了一种结合多层特征及空间信息蒸馏的医学影像分割方法TinyUnet。该方法使用轻量化的U-Net作为学生网络。考虑到小模型没有足够的学习能力，通过选择合适的蒸馏位置，对多层教师特征图进行蒸馏; 同时加强教师网络深层特征图的边缘，并构建边缘关键点图结构，采用图卷积网络对学生网络进行空间信息蒸馏，从而补充重要的边缘信息和空间信息。实验表明:在3个医学影像数据集上，TinyUnet能够达到U-Net 98.3%~99.7%的分割准确度，但是将U-Net的参数量平均降低了99.6%，运算速度提高了约110倍; 同时，与其他轻量化医学影像分割模型相比，TinyUnet不仅具有较高的分割准确度，而且占用内存更少，运行速度更快。      

Propulsionless planar phasing of multiple satellites using deep reinforcement learning

This work creates a framework for solving highly non-linear satellite formation control problems by using model-free policy optimisation deep reinforcement learning (DRL) methods. This work considers, believed to be for the first time, DRL methods, such as advantage actor-critic method (A2C) and proximal policy optimisation (PPO), to solve the example satellite formation problem of propellantless planar phasing of multiple satellites. Three degree-of-freedom simulations, including a novel surrogate propagation model, are used to train the deep reinforcement learning agents. During training, the agents actuated their motion through cross-sectional area changes which altered the environmental accelerations acting on them. The DRL framework designed in this work successfully coordinated three spacecraft to achieve a propellantless planar phasing manoeuvre. This work has created a DRL framework that can be used to solve complex satellite formation flying problems, such as planar phasing of multiple satellites and in doing so provides key insights into achieving optimal and robust formation control using reinforcement learning.     

Physical assessment of coastal vulnerability under enhanced land subsidence in Semarang,Indonesia, using multi-sensor satellite data

Research has been conducted in Semarang, Indonesia, to assess coastal vulnerability under enhanced land subsidence using multi-sensor satellite data, including the Advanced Land Observing Satellite (ALOS) Phased Array type L-band SAR (PALSAR), Landsat TM, IKONOS, and TOPEX/Poseidon. A coastal vulnerability index (CVI) was constructed to estimate the level of vulnerability of a coastline approximately 48.68?km in length using seven physical variables, namely, land subsidence, relative sea level change, coastal geomorphology, coastal slope, shoreline change, mean tidal range, and significant wave height. A comparison was also performed between a CVI calculated using seven parameters and a CVI using six parameters, the latter of which excludes the land subsidence parameter, to determine the effects of land subsidence during the coastal vulnerability assessment. This study showed that the accuracy of coastal vulnerability was increased 40% by adding the land subsidence factor (i.e., CVI 6 parameters?=?53%, CVI 7 parameters?=?93%). Moreover, Kappa coefficient indicated very good agreement (0.90) for CVI 7 parameters and fair agreement (0.3) for CVI 6 parameters. The results indicate that the area of very high vulnerability increased by 7% when land subsidence was added. Hence, using the CVI calculation including land subsidence parameters, the very high vulnerability area is determined to be 20% of the total coastline or 9.7?km of the total 48.7?km of coastline. This study proved that land subsidence has significant influence on coastal vulnerability in Semarang.     

A strategy for mineral and energy resource independence

Data acquired by Landsats 1, 2, and 3, are beginning to provide the information on which an improved mineral and energy resource exploration strategy can be based. Landsat 4 is expected to augment this capability with its higher resolution (30 m) and additional spectral bands in the Thematic Mapper (TM) designed specifically to discriminate clay minerals associated with mineral alteration. In addition, a new global magnetic anomaly map, derived from the recent Magsat mission, has recently been compiled by the National Aeronautics and Space Administration (NASA), the U.S. Geological Survey (USGS), and others. Preliminary, extremely small-scale renditions of this map indicate that global coverage is nearly complete and that the map will improve upon a previous one derived from Polar Orbiting Geophysical Observatory (POGO) data. Digital processing of the Landsat image data and Magsat geophysical data can be used to create three-dimensional stereoscopic models for which Landsat images provide surface reference to deep structural anomalies.Comparative studies of national Landsat lineament maps, Magsat stereoscopic models, and metallogenic information derived from the Computerized Resources Information Bank (CRIB) inventory of U.S. mineral resources, provide a way of identifying and selecting exploration areas that have mineral resource potential. Landsat images and computer-compatible tapes can provide new and better mosaics and also provide the capability for a closer look at promising sites.     

融合深度特征的输电线路金具缺陷因果分类方法

针对输电线路金具缺陷样本不足和缺陷目标形态多样化，仅仅利用深度学习模型导致金具缺陷分类准确率较低的问题，提出了一种结合深度网络和逻辑回归模型的因果分类方法。首先，通过样本扩充方法获得数量丰富化和角度多样化的数据集；然后，基于微调后的VGG16模型提取深度特征并进行特征处理，以构建符合因果关系学习的输入特征集；最后，通过全局混杂平衡进行金具缺陷特征与标签之间的因果关系学习，构建符合金具特点的因果逻辑回归模型，完成金具缺陷分类。为了证明所提方法的有效性，利用无人机实际采集的4类金具缺陷图片分别进行了实验，所使用的训练样本和测试样本数量较原始数据集提升了5倍左右。实验结果表明:所提方法可以实现对输电线路金具缺陷的精准分类，其中，防震锤相交和变形分类准确率分别达到了0.929 9和0.911 8，屏蔽环锈蚀和均压环损坏分类准确率分别达到了0.956 7和0.966 9。      

Automatic extraction of lineaments based on wavelet edge detection and aided tracking by hillshade

Lineaments refer to the linear or curvilinear textures on remote sensing image, whose general spatial distribution characteristics are often the response of deep geological sturcture at the surface. Firstly, we use wavelet modulus maximum transformation to detect the edges with 4 scales on Landsat – 8 OLI B5 image and analyze their multi-scale characteristics. As the result, it is determined that the optimal scale of edge detection is 4, and the outline that consist of the edge pixels is roughly corresponding to the geological structure of mine area. Thus the incomplete lineaments have been extracted by using the 2D otsu algorithm. Secondly, the hillshade map generated based on DEM is processed to generate binarized linear shadow. Finally, the linear shadow is superimposed on the lineaments preliminarily extracted to obtain the optimized lineaments. Experiment results show that, based on the method, there are some deformation and displacement between the lineaments extracted and the actual geological structure, and it fail to effectively extract Qilinchang Fault, but lineaments are in good correspondence with Kuangshanchang Fault, Dongtou Fault and Niulan River Fault, which are basically in accord with the geological structure framework of the mine area.     

真实场景水下语义分割方法及数据集

随着水下生物抓取技术的不断发展,高精度的水下物体识别与分割成为了挑战。已有的水下目标检测技术仅能给出物体的大体位置,无法提供物体轮廓等更加细致的信息,严重影响了抓取效率。为了解决这一问题,标注并建立了真实场景水下语义分割数据集DUT-USEG,该数据集包含6 617张图像,其中1 487张具有语义分割和实例分割标注,剩余5 130张图像具有目标检测框标注。基于该数据集,提出了一个关注边界的半监督水下语义分割网络(US-Net),该网络通过设计伪标签生成器和边界检测子网络,实现了对水下物体与背景之间边界的精细学习,提升了边界区域的分割效果。实验表明:所提方法在DUT-USEG数据集的海参、海胆和海星3个类别上相较于对比方法提升了6.7%,达到了目前最好的分割精度。      

Landsat 8地表温度产品验证与分析

Landsat系列卫星热波段具有60~120m的空间分辨率，对各种环境监测起到了重要的作用。随着Landsat系列卫星在全球范围内地表温度（land surface temperature，LST）产品的发布，其验证工作也随之展开，然而对于长时间序列的精度验证工作仍然缺乏。以黑河流域中游为研究区，利用研究区内湿地站（SD）、戈壁站（GB）和大满超级站（CJZ）三个气象站的地面测量数据对2013-2016年清晰无云的31景Landsat 8地表温度产品进行了验证与分析，并将Landsat 8地表温度产品与广泛使用的普适性单通道算法（JMS）反演结果进行了对比。结果表明，Landsat 8地表温度产品与普适性单通道算法反演结果精度均较高，在各个站点处R2均优于0.949。基于所有站点分析，Landsat 8地表温度产品精度稍高于普适性单通道算法反演结果。     

基于深度增强学习的卫星姿态控制方法

针对卫星在执行丢弃载荷或捕获目标等复杂任务时遭遇的姿态突然发生变化的问题,采用深度增强学习方法对卫星姿态进行控制,使卫星恢复稳定状态。具体来说,首先搭建飞行器的姿态动力学环境,并将连续的控制力矩输出离散化,然后采用Deep Q Network算法进行卫星自主姿态控制训练,以姿态角速度趋于稳定作为奖励获得离散行为的最优智能输出。仿真试验表明,面向空间卫星姿态控制的深度增强学习算法能够在卫星受到突发随机扰动后稳定卫星姿态,并能有效解决传统PD控制器依赖被控对象质量参数的难题。所提出的方法采用自主学习的方式对卫星姿态进行控制,具有很强的智能性和一定的普适性,在未来卫星执行复杂空间任务中的智能控制方面有着很好的应用潜力。     

Main results and perspectives of some Chilean experiences developed with low cost and accurate spatial remote sensing technology

This paper intends to summarize the main results and perspectives of several Chilean Programs developed by using low cost but accurate remote sensing techniques. Due to paper restrictions, three main applications will be shown : use of satellite data collection systems to measure meteorological data in Ant arctic Peninsula; study of geothermal resources in Los Andes Range by using multispectral and multitemporal Landsat images; and snowmelt runoff forecasting for Andean watersheds by using Landsat data. All these applications have allowed to obtain important and useful results and low cost, reliable and accurate methodologies have been obtained for these studies.     

Comparison of Landsat OLI,ASTER, and Sentinel 2A data in lithological mapping : A Case study of Rich area (Central High Atlas,Morocco)

The eastern part of the Rich area consists of the massive Paleozoic and Meso-Cenozoic cover formations that present the geodynamic development of the study area, where is characterized by various carbonate facies of Jurassic age. The geographical characteristic of the study area leaves the zone difficult to map by conventional methods. The objective of this work focuses on the mapping of the constituent lithological units of the study area using multispectral data of Landsat OLI, ASTER, and Sentinel 2A MSI. The processing of these data is based on a precise methodology that distinguishs and highlights the limits of the different lithological units that have an approximate similarity of spectral signature. Three techniques were used to enhance the image including Principal Component Analysis (PCA), Minimum Noise Fraction (MNF), and Independent Component Analysis (ICA). Lithological mapping was performed using two types of supervised classification : Maximum likelihood classifier (MLC) and Support Vector Machine (SVM).The results of processing data show the effectiveness of Sentinel 2A data in mapping of lithological units than the ASTER and Landsat OLI data. The classification evaluation of two methods of the Sentinel 2A MSI image showed that the SVM method give a better classification with an overall accuracy of 93,93% and a Kappa coefficient of 0.93, while the MLC method present an overall accuracy of 82,86% and a Kappa coefficient of 0.80. The results of mapping obtained show a good correlation with the geological map of the study area as well as the efficiency of remote sensing in identification of different lithological units in the Central High Atlas.