A-FPN算法及其在遥感图像船舶检测中的应用

光学遥感图像船舶检测主要面临两个挑战:光学遥感图像背景复杂,船舶检测易受海浪、云雾及陆地建筑等多方面干扰;遥感图像分辨率低,船舶目标小,对于其分类与定位带来很大困难;针对上述问题,在FPN的基础上,提出一种融入显著性特征的卷积神经网络模型A-FPN (Attention-Based Feature Pyramid Networks)。首先,利用卷积提取图像特征金字塔;然后,利用顶层金字塔逐级构建显著特征层,抑制背景信息,通过金字塔顶层的细粒度特征提高浅层特征的表达能力,构建自上而下的多级显著特征映射结构;最后利用Softmax分类器进行多层级船舶检测。A-FPN模型利用显著性机制引导不同感受下的特征进行融合,提高了模型的分辨能力,对遥感图像处理领域具有重要应用价值。实验阶段,利用公开的遥感目标检测数据集NWPU VHR-10中的船舶样本进行测试,准确率为92.8%,表明A-FPN模型适用于遥感图像船舶检测。     

经济地理学视角下创新网络与区域增长研究述评

创新是引领经济发展的第一动力，创新与区域增长之间的关系成为经济地理学者关注的重点领域。文章通过梳理近年来有关创新网络测度、创新网络与区域增长关系、创新网络作用于区域增长方式等方面的文献发现：网络资本可以弥补社会资本在解释企业创新结网经济价值方面的不足，区域增长呈现出网络化特征已成为学界共识，知识流与邻近性能较好地解释创新网络与区域增长的关系机理。然而，当前研究对社会资本如何促进区域增长，网络资本与区域增长关系模型如何构建，不同类型的邻近性与知识对区域增长的影响有何差异等问题尚不明确，建议重视网络资本对区域增长的作用并实证检验二者的关联，同时还应重视创新网络中企业家精神、创新个体心理行为特征等因素对区域增长的影响。     

Evaluating techniques for mapping island vegetation from unmanned aerial vehicle (UAV) images: Pixel classification,visual interpretation and machine learning approaches

We evaluate three approaches to mapping vegetation using images collected by an unmanned aerial vehicle (UAV) to monitor rehabilitation activities in the Five Islands Nature Reserve, Wollongong (Australia). Between April 2017 and July 2018, four aerial surveys of Big Island were undertaken to map changes to island vegetation following helicopter herbicide sprays to eradicate weeds, including the creeper Coastal Morning Glory (Ipomoea cairica) and Kikuyu Grass (Cenchrus clandestinus). The spraying was followed by a large scale planting campaign to introduce native plants, such as tussocks of Spiny-headed Mat-rush (Lomandra longifolia). Three approaches to mapping vegetation were evaluated, including: (i) a pixel-based image classification algorithm applied to the composite spectral wavebands of the images collected, (ii) manual digitisation of vegetation directly from images based on visual interpretation, and (iii) the application of a machine learning algorithm, LeNet, based on a deep learning convolutional neural network (CNN) for detecting planted Lomandra tussocks. The uncertainty of each approach was assessed via comparison against an independently collected field dataset. Each of the vegetation mapping approaches had a comparable accuracy; for a selected weed management and planting area, the overall accuracies were 82 %, 91 % and 85 % respectively for the pixel based image classification, the visual interpretation / digitisation and the CNN machine learning algorithm. At the scale of the whole island, statistically significant differences in the performance of the three approaches to mapping Lomandra plants were detected via ANOVA. The manual digitisation took a longer time to perform than others. The three approaches resulted in markedly different vegetation maps characterised by different digital data formats, which offered fundamentally different types of information on vegetation character. We draw attention to the need to consider how different digital map products will be used for vegetation management (e.g. monitoring the health individual species or a broader profile of the community). Where individual plants are to be monitored over time, a feature-based approach that represents plants as vector points is appropriate. The CNN approach emerged as a promising technique in this regard as it leveraged spatial information from the UAV images within the architecture of the learning framework by enforcing a local connectivity pattern between neurons of adjacent layers to incorporate the spatial relationships between features that comprised the shape of the Lomandra tussocks detected.     

SIMULATION AND PREDICTION OF DEBRIS FLOW USING ARTIFICIAL NEURAL NETWORK

Debris flow is one of the most destructive phenomena of natural hazards. Recently, major natural haz-ard, claiming human lives and assets, is due to debris flow in the world. Several practical methods for forecasting de-bris flow have been proposed, however, the accuracy of these methods is not high enough for practical use because of the stochastic and non-linear characteristics of debris flow. Artificial neural network has proven to be feasible and use-fill in developing models for nonlinear systems. On the other hand, predicting the future behavior based on a time se-ries of collected historical data is also an important tool in many scientific applications. In this study we present a three-layer feed-forward neural network model to forecast surge of debris flow according to the time series data collect-ed in the Jiangjia Ravine, situated in north part of Yunnan Province of China. The simulation and prediction of debris flow using the proposed approach shows this model is feasible, however, further studies are needed.     

Identification of physical processes inherent in artificial neural network rainfall runoff models

The emergence of artificial neural network (ANN) technology has provided many promising results in the field of hydrology and water resources simulation. However, one of the major criticisms of ANN hydrologic models is that they do not consider/explain the underlying physical processes in a watershed, resulting in them being labelled as black‐box models. This paper discusses a research study conducted in order to examine whether or not the physical processes in a watershed are inherent in a trained ANN rainfall‐runoff model. The investigation is based on analysing definite statistical measures of strength of relationship between the disintegrated hidden neuron responses of an ANN model and its input variables, as well as various deterministic components of a conceptual rainfall‐runoff model. The approach is illustrated by presenting a case study for the Kentucky River watershed. The results suggest that the distributed structure of the ANN is able to capture certain physical behaviour of the rainfall‐runoff process. The results demonstrate that the hidden neurons in the ANN rainfall‐runoff model approximate various components of the hydrologic system, such as infiltration, base flow, and delayed and quick surface flow, etc., and represent the rising limb and different portions of the falling limb of a flow hydrograph. Copyright © 2004 John Wiley & Sons, Ltd.     

Generalized strain probing of constitutive models

Advanced material constitutive models are used to describe complex soil behaviour. These models are often used in the solution of boundary value problems under general loading conditions. Users and developers of constitutive models need to methodically investigate the represented soil response under a wide range of loading conditions. This paper presents a systematic procedure for probing constitutive models. A general incremental strain probe, 6D hyperspherical strain probe (HSP), is introduced to examine rate‐independent model response under all possible strain loading conditions. Two special cases of HSP, the true triaxial strain probe (TTSP) and the plane‐strain strain probe (PSSP), are used to generate 3‐D objects that represent model stress response to probing. The TTSP, PSSP and general HSP procedures are demonstrated using elasto‐plastic models. The objects resulting from the probing procedure readily highlight important model characteristics including anisotropy, yielding, hardening, softening and failure. The PSSP procedure is applied to a Neural Network (NN) based constitutive model. It shows that this probing is especially useful in understanding NN constitutive models, which do not contain explicit functions for yield surface, hardening, or anisotropy. Copyright © 2004 John Wiley & Sons, Ltd.     

Automatic Determination of Stellar Atmospheric Parameters Using Neural Networks and Instance-Based Machine Learning

In this article we show how machine learning methods can beeffectively applied to the problem of automatically predictingstellar atmospheric parameters from spectral information, a veryimportant problem in stellar astronomy. We apply feedforwardneural networks, Kohonen's self-organizing maps andlocally-weighted regression to predict the stellar atmosphericparameters effective temperature, surface gravity and metallicityfrom spectral indices. Our experimental results show that thethree methods are capable of predicting the parameters with verygood accuracy. Locally weighted regression gives slightly betterresults than the other methods using the original dataset asinput, while self-organizing maps outperform the other methods when significant amounts of noise are added. We also implemented a heterogeneous ensemble of predictors, combining the results given by the three algorithms. This ensemble yields better results than any of the three algorithms alone, using both the original and the noisy data.     

水下滑翔器主体外形优化设计

对水下滑翔器主体外形的优化设计问题进行研究.首先,在考虑内部机械结构限制的条件下,建立了主体外形优化数学模型,并采用CFX计算了若干主体形状的绕流阻力.在此基础上建立了映射主体外形的尺寸参数与绕流阻力的BP神经网络.然后将建立的神经网络作为优化设计问题的目标函数,采用坐标轮换法对滑翔器主体外形进行了优化设计,确定了尺寸参数的最优解.     

Reliability assessment of serviceability performance of braced retaining walls using a neural network approach

In urban environments, one major concern with deep excavations in soft clay is the potentially large ground deformations in and around the excavation. Excessive movements can damage adjacent buildings and utilities. There are many uncertainties associated with the calculation of the ultimate or serviceability performance of a braced excavation system. These include the variabilities of the loadings, geotechnical soil properties, and engineering and geometrical properties of the wall. A risk‐based approach to serviceability performance failure is necessary to incorporate systematically the uncertainties associated with the various design parameters. This paper demonstrates the use of an integrated neural network–reliability method to assess the risk of serviceability failure through the calculation of the reliability index. By first performing a series of parametric studies using the finite element method and then approximating the non‐linear limit state surface (the boundary separating the safe and ‘failure’ domains) through a neural network model, the reliability index can be determined with the aid of a spreadsheet. Two illustrative examples are presented to show how the serviceability performance for braced excavation problems can be assessed using the reliability index. Copyright © 2005 John Wiley & Sons, Ltd.     

Comments on ‘A comparative study of ANN and neuro-fuzzy for the prediction of dynamic constant of rockmass’

Singh et al (2005) examined the potential of the ANN and neuro-fuzzy systems application for the prediction of dynamic constant of rockmass. However, the model proposed by them has some drawbacks according to fuzzy logic principles. This discussion will focus on the main fuzzy logic principles which authors and potential readers should take into consideration.