基于遥感和机器学习的内陆水体水深反演技术

文章主要根据机器学习算法（随机森林算法和极端梯度提升算法）和遥感水深反演的原理,利用Sentinel_2多光谱卫星数据和无人船实测水深数据,对内陆水体——梅州水库建立了随机森林（RF）、极端梯度提升（XGBoost）和支持向量机（SVM）水深反演模型,并对反演结果进行对比分析。结果表明：1）RF的训练精度为97%,测试精度为0.80;XGBoost模型的训练精度为97%,测试精度为0.79;SVM的训练精度为90%,测试精度为0.78。说明了在水深预测方面RF模型和XGBoost模型比SVM模型表现更好,对各个区段的水深值较为敏感。2）根据运行时间考察各个模型的效率,其中RF模型从读取数据至输出结果耗时3.92 s;XGBoost模型4.26 s;SVM模型6.66 s。因此,在反演精度和效率上RF模型优于XGBoost模型优于SVM模型,且RF模型的预测结果图细节更加丰富,轮廓更加分明;XGBoost模型次之,但总体效果也较好;SVM模型表现最差。由此可知,机器学习水深反演模型获得的水深结果精度明显提高,解决了传统水深反演模型精度不高的问题。     

All pixels are useful,but some are more useful: Efficient in situ data collection for crop-type mapping using sequential exploration methods

A precise knowledge of the crop distribution in the landscape is crucial for the agricultural sector to inform better management and logistics. Crop-type maps are often derived by the supervised classification of satellite imagery using machine learning models. The choice of data sampled during the data collection phase of building a classification model has a tremendous impact on a model's performance, and is usually collected via roadside surveys throughout the area of interest. However, the large spatial extent, and the varying accessibility to fields, often makes the acquisition of appropriate training data sets difficult. As such, in situ data are often collected on a best-effort basis, leading to inefficiencies, sub-optimal accuracies, and unnecessarily large sample sizes. This highlights the need for new more efficient tools to guide data collection. Here, we address three tasks that one commonly faces when planning to collect in situ data: which survey route to select among a set logistically feasible routes; which fields are the most relevant to collect along the chosen survey route; and how to best augment existing in situ data sets with additional observations. Our findings show that the normalised Moran's I index is a useful indicator for choosing the survey route, and that sequential exploration methods can identify the most important fields to survey on that route. The provided recommendations are flexible, overcome the main logistical constraints associated with in situ data collection, yield accurate results, and could be incorporated in a mobile application to assist data collection in real-time.     

Monitoring early stage invasion of exotic Spartina alterniflora using deep-learning super-resolution techniques based on multisource high-resolution satellite imagery: A case study in the Yellow River Delta,China

Over the past decades, Spartina alterniflora, one of the top exotic invasive plants in China, has expanded throughout coastal China. In the Yellow River Delta (YRD), the rapid expansion of S. alterniflora has caused serious negative ecological effects. Current studies have concentrated primarily on mapping the distribution of S. alterniflora with medium-resolution satellite imagery at the regional or landscape scale, which have a limited capability in early detection and monitoring of the invasive process at the patch scale. In this study, we proposed a framework for monitoring the early stage invasion of S. alterniflora patches in the YRD using multiyear multisource high-spatial-resolution satellite imagery with various ground sampling distances (WorldView-2, SPOT-6, GaoFen-1, GaoFen-2, and GaoFen-6 from 2012 to 2019). First, we proposed to use deep-learning-based image super-resolution models to enhance all images to submeter (0.5 m) resolution. Then, we adopted stepwise evolution analysis-based image segmentation and object-based classification rules to detect and delineate S. alterniflora patches from the super-resolved imagery. By investigating Super-Resolution Convolutional Neural Networks (SRCNN) and Fast Super-Resolution Convolutional Neural Networks (FSRCNN) and comparing these methods with the conventional bicubic interpolation method for image resolution enhancement, we concluded that FSRCNN was superior in constructing spectral and structural details from the 1 m/1.5 m/2 m resolution images to 0.5 m resolution. FSRCNN, in particular, was more effective and efficient in discerning and estimating the size of small S. alterniflora patches (<50 m²). Using our method, 76 of 83 field-measured small patches were accurately detected and the delineated S. alterniflora patch perimeters agreed well with the field-measured patch perimeters (root mean square error [RMSE] = 8.29 m, mean absolute percentage error [MAPE] = 23.46 %). The invasion process showed fast expansion from 2012 to 2015 and slow growth from 2016 to 2019. We observed that the landward limits of S. alterniflora patches were influenced by elevation and vicinity to tidal creeks.     

Mapping hurricane damage: A comparative analysis of satellite monitoring methods

Wetlands are the second-most valuable natural resource on Earth but have declined by approximately 70 % since 1900. Restoration and conservation efforts have succeeded in some areas through establishment of refuges where anthropogenic impacts are minimized. However, these areas are still prone to wetland damage caused by natural disasters. Severe storms such as Hurricane Irma, which made landfall as a Category 3 hurricane in southwest Florida (USA) on September 11, 2017, can cause the destruction of mangroves and other wetland habitat. Multispectral images from commercial satellites provide a means to assess the extent of the damage to different wetland habitat types with high spatial resolution (2 m pixels or finer) over large areas. Using such images presents a number of challenges, including deriving consistent and accurate classification of wetland and non-wetland vegetation. Machine learning methods have demonstrated high-accuracy mapping capabilities on small spatial scales, but require a large amount of robust training data. Meanwhile, ambitious efforts to map larger areas at finer resolutions may use hundreds of thousands of images, and therefore encounter Big-Data processing challenges. Large-scale efforts face the dilemma of adopting traditional mapping methods that may lend themselves to Big Data analytics but may result in accuracies that are inferior to new methods, or move to machine learning methods, which require robust training data. Given these considerations, we describe a version of the traditional Decision Tree (DT) approach and compare two common machine learning methods to derive land cover classes using a WorldView-2 image collected on November 12, 2018 to include one growing season after Hurricane Irma affected this area. Specifically, we compared the Support Vector Machine [SVM] and Neural Network [NN] methods, trained and validated with separate ground-truth datasets collected during a robust field campaign. Overall accuracies were only marginally different (85 % NN vs 83 % each DT and SVM), but healthy mangroves were more accurately identified with the DT (91 % vs 88 % NN and 86 % SVM), and degraded mangroves were more accurately identified with NN (62 % vs 57 % NN and 38 % DT). These results, combined with their respective training requirements, have implications for the direction with which large-scale high-resolution mapping of coastal habitats proceeds.     

结合数据增广和迁移学习的高分辨率遥感影像场景分类

深度学习在计算机视觉领域取得了显著的成果,如图像分类、人脸识别、图像检索等。对于遥感领域而言,获取用于训练CNN的有标签数据集通常是一个重大挑战。本文研究了如何将CNN用于高分辨率遥感影像的场景分类,为了克服缺乏大量有标签遥感影像数据集的问题,结合CNN采用了两种技术：数据增广和迁移学习。在UC Merced Land Use数据集上,验证了VGG16、VGG19、ResNet50、InceptionV3、DenseNet121等5种网络的性能,分别达到了98.10%、96.19%、99.05%、97.62%、99.52%的分类准确率。     

激光点云在地铁盾构隧道病害诊断中的应用

移动式三维激光扫描技术在地铁盾构隧道安全监测工作中应用较为成熟。本文以地铁盾构隧道监测点云数据为基础进行研究,实现了地铁盾构隧道病害智能诊断。首先通过激光点云生成灰度图像;在此基础上运用卷积神经网络CNN,对地铁盾构隧道中的渗漏水和裂缝的识别技术进行了深入研究;最终生成隧道病害智能诊断系统,为地铁安全运营提供了智能监测方法,有效提高了我国地铁运营监测的技术水准。     

基于机器学习的脉动源格林函数预报初探

自由面格林函数是边界元法求解海洋工程水动力学问题的基础,如何精确而快速地计算格林函数及其导数是水动力求解的难题。对无因次化表达的脉动点源格林函数计算建立的数据库,采用深度学习函数库Keras,对数据库进行学习,建立了神经网络预报模型,探讨了全局和局部学习及预报精度,研究了模型预报效率。研究表明机器学习模型预报的格林函数能够保证较高的精度,计算效率高于数值积分计算,低于解析函数为主的多项式逼近,为提高水动力问题求解效率,解决传统计算难题提供了新的思路。     

Evaluation of linear discriminant and support vector machine classifiers for classification of nitrogen status in mature oil palm from SPOT-6 satellite images: analysis of raw spectral bands and spectral indices

Nitrogen (N) management is important in sustaining oil palm production. Remote sensing-based approaches via spectral index have promise in assessing the N nutrition content. The objectives of this study are; (i) to examine the N classification capability of three spectral indices (SI) such as visible (Vis), near infrared (NIR) and a combination of visible and NIR (Vis + NIR) from the SPOT-6 satellite, and (ii) to compare the performance of linear discriminant analysis (LDA) and support vector machine (SVM) in discriminating foliar N content of mature oil palms. Nitrogen treatments varied from 0 to 2 kg per palm. The N-sensitive SIs tested in this study were age-dependent. The Vis index (BGRI₁) (CVA = 79.55%) and Vis + NIR index (NDVI, NG, IPVI and GNDVI) (CVA = 81.82%) were the best indices to assess N status of young and prime mature palms through the SVM classifier.     

联合空-谱信息的高光谱影像深度三维卷积网络分类

针对高光谱影像分类高维和小样本的特点,提出一种基于深度三维卷积神经网络的高光谱影像分类方法。首先,该方法直接以高光谱数据立方体为输入,利用三维卷积操作提取高光谱数据立方体的三维空-谱特征。然后,利用残差学习构建深层网络,提取更高层次的特征表达,以提高分类精度。最后,采用Dropout正则化方法防止过拟合。利用Pavia大学、Indian Pines和Salinas 3组高光谱数据进行试验验证,结果表明,与支持向量机和现有的基于深度学习的高光谱影像分类方法相比,该方法能有效提高高光谱影像的地物分类精度。     

级联卷积神经网络的遥感影像飞机目标检测

传统遥感影像飞机目标检测算法依赖于人工设计特征,对大范围复杂场景和多尺度的飞机目标稳健性较差,基于深层卷积神经网络的目标检测算法通常难以有效应对大幅影像的目标搜索和弱小目标检测问题,针对上述问题,本文提出了一种基于级联卷积神经网络的遥感影像飞机目标检测算法。首先根据全卷积神经网络能够支持输入任意大小图像的特点,采用小尺度浅层全卷积神经网络对整幅影像进行遍历和搜索,快速获取疑似飞机目标作为兴趣区域,然后利用较深层的卷积神经网络对兴趣区域进行更精确的目标分类与定位。为提高卷积神经网络对地物目标的辨识能力,在卷积层中引入多层感知器,并在训练过程中采取多任务学习与离线难分样本挖掘的策略;在测试阶段,建立影像金字塔进行多级搜索,并结合非极大值抑制消除冗余窗口,从而实现由粗到精的飞机目标检测与识别。对多个数据集下多种复杂场景的遥感影像进行测试,结果表明,本文方法具有较高的准确性和较强的稳健性,可为大幅遥感影像的飞机目标检测问题提供一个快速高效的解决方案。