城市计算是当今学术界的一个重要研究方向，且在环境、教育、能源和城市规划等领域有着广泛的应用，它旨在通过不断获取、整合和分析城市中多种异构大数据来解决城市所面临的挑战。本文主要针对空气质量预报、边缘检测及其应用、空气质量检测和水管预防性维护四个城市计算中急需解决的问题进行阐述。针对行业内传统方法老旧、准确性不高、成本高昂、计算复杂等问题，本文应用了特征学习的技术以更充分地利用特征信息来解决城市计算相关的问题。具体来说，本文提出了基于特征映射指派通道的集成学习的空气质量预报方法，提出了基于集成随机森林的边缘检测算法及研究了其在城市现代化教育中的对光学字符的分割及识别中的应用，提出了基于图像特征的空气质量检测算法和提出了基于档案特征的互激励点过程方法的城市管网预防性维护方法。实验结果表明，本文的方法很好地解决了城市计算中相关的问题，实验结果在多项性能指标上都有了显著提升。本文的主要工作和贡献如下： 1）空气质量预报即给定气象条件预报未来一段时间的空气质量，现有的方法存在准确性较低、计算成本高等问题。针对上述问题，本文提出了基于特征映射指派通道的集成学习的空气质量预报方法。具体来说，首先研究了直接使用特征进行统计预报，并提出了基于有监督的多通道的集成学习模型，在每个通道中都有一个预报器，可以在训练过程中根据样本的预报错误率动态的迭代更新这些预报器。其次，提出了基于深度玻尔兹曼机的充分统计量的特征映射的指派算法，用来有监督地选择最终用哪个通道进行预测。相比原始特征，该算法能额外获得模型的分布信息，以及隐含变量所传达的信息。最后，研究了用于城市空气质量预报的网络服务。实验结果表明，本文的预报准确性明显优于统计方法和数值方法。此外本章的特征映射也将用于后面做边缘检测。 2）边缘检测即提取自然图像中物体的轮廓信息，现有的方法存在准确率低、检测时间长等问题。针对上述问题，本文在分析了边缘检测的多尺度特征基础上，通过引入特征映射来获得更丰富的特征，提出了基于集成随机森林的边缘检测算法，并用于光学字符分割和空气污染照片分割。本文选取的多尺度特征，主要包括像素级别的特征和线段级别的特征。其中，像素级别的特征又包括多尺度点特征和多尺度全局特征，多尺度点特征又包括图像梯度特征、纹理抑制特征、亮度梯度特征、颜色梯度特征、轮盘算子特征。特征映射探寻包含了更多信息的隐变量，可以推导出更多附加的信息。相比于广泛使用的卷积神经网络，本文所使用的特征映射算法有更少的模型参数，相对于深度学习仅需要更少的样本。实验结果表明，本文的方法有检测速度快、准确性高等诸多优势。此外边缘检测结果也将用于分割空气质量图片和为管道预防性维护提供视觉纹理特征。 3）空气质量监测即通过某种传感器或者技术手段来获得当前所在环境的空气质量，现有的方法存在成本高昂、覆盖面小、携带不便等问题。针对上述问题，本文提出了基于图像特征的空气质量检测方法。首先，提出了基于图像字典特征表示的方法来检测空气质量，在基于传统方法来获取图像特征的基础上，引进了l_21范数来对目标函数进行约束。其次，提出了基于单通道深度卷积网络的方法来自动的提取图像特征来检测空气质量，具体来讲，提出了一个新的灵活负ReLU激活函数，新的激活函数具有灵活性高、训练高效等特点并可以有效地克服梯度消散问题。再次，提出了采用有序分类器来替换Softamax来对空气污染进行估计，序分类器则考虑了排序的问题。然后在单通道的基础上，为了获得更稳定的结果，提出了基于多通道集成深度卷积网络的方法来检测空气质量。最后，提出了基于物理基站、图像字典和深度卷积网络三种方法组合的城市空气质量混合检测网络服务。实验结果表明，本文的方法有成本低廉、覆盖面广、携带方便、准确率高等诸多优势。此外本章的空气质量检测结果也将为管道预防性维护提供环境特征信息。 4）城市管网预防性维护即通过使用材料型号、管道直径、纹理特征、环境信息和过去的失败事件等信息来评估每根管子失败的风险概率，现有的方法存在准确率低下的问题。针对上述问题，本文提出了基于档案特征的互激励过程方法的城市管网预防性维护算法。首先，研究了对城市管网预防性维护问题的数学建模。其次，提出了基于档案特征对基础强度函数建模，模型的复杂性跟样本的数量无关，仅依赖于特征的数目，其中档案特征包括管道固有的如材质、长度等属性，也包括边缘检测给出的视觉纹理特征，还包括管道所处的环境的空气质量和温度、湿度。最后提出了互激励点过程算法，考虑了不同失败类型的影响及其相互触发效果，并提出了优化算法对该问题进行求解。实验结果表明，相比于传统方法，本文方法的准确性最高。

Urban computing is an important research direction in the current research field and has a wide range of applications in environment, education, energy and city planning, etc. It is a process of acquisition, integration, and analysis of big and heterogeneous data generated by a diversity of sources in urban spaces to tackle the major issues. This paper selects four urgent cases including air pollution forecasting, contour detection and its application, measuring air quality and preventative urban pipe maintenance. Traditional industry based methods have some problems like old algorithm, low accuracy, high cost and complicated computation. All methods in this paper apply feature learning method to utilize feature information more efficiently to solve urban computing related problems. More specifically, this paper comes up with air pollution forecasting based ensemble learning with feature mapping to select channel, contour detection based stacking random forest learning and its application in segmentation and recognition to optical character in urban modern education, air quality estimation based photo feature and preventative urban pipe maintenance based a mutual-exciting point process with profile-specific base intensity modeling to describe this problem. The final experimental results show the methods in this paper solve urban computing problems very well with better results and better performance. The following is the work of this article: 1) Air forecasting means forecasting future air quality given weather condition. Current methods are low accuracy and high computational expense. To solve the above problems, this paper introduces air pollution forecasting based ensemble learning with feature mapping to select channel. Specifically speaking, firstly study using feature to forecast air pollution directly. Next introduce a supervised multi-channel ensemble learning model and in each channel there is a forecasting model, whose training is iteratively updated based on the metric of the prediction error on training samples. Then propose a channel assignment algorithm based sufficient statistics feature mapping over Deep Boltzmann Machine and this model is used in the testing stage to decide which channel the coming new sample shall be assigned to perform prediction. This feature mapping could get the training data distribution information and the information associated with hidden units. At last study a web service for air pollution forecasting. The experimental results show comparing with statistic methods and numeric methods, this paper’s method achieves best accuracy. Moreover, the feature mapping in this chapter is also used to detect contour later. 2) Contour detection means extract object outlines in natural images. The current methods are low accuracy and highly time consuming. In view of the above question, on the basis of analyzing multi-scale image features for contour detection, by introducing feature mapping for more abundant feature information, this paper proposes contour detection based stacking random forest learning to segment optical character and air pollution images. The multi-scale image features include pixel-level features and segment-level features. And pixel-level features include magnitude of gradient, direction of gradient, inhibition term, brightness gradient, color gradient and compass operator. The feature mapping exploits the hidden variable which encodes additional information. The proposed feature mapping model involves less model parameters and thus requires less training data. The experimental results show method in this paper is much faster and more accuracy. Besides the results of contour detection are also used to segment air quality images and offer visual texture features for preventative pipe maintenance. 3) Measure air quality means getting the local air quality by some sensors or technology. The current methods are expensive in the cost, small coverage area and inconvenient to carry. Aiming at these problems above, this paper proposes air quality estimation based photo feature. Firstly study air quality assessment by image dictionary learning which is based on traditional method to get features and add a l_21 norm to restraint the optimization formulation. Next design single channel convolutional neural network to extract features and estimate air's quality based on photos. Concretely speaking introduce a new flexible negative ReLU activation function and it is more flexible, less time-consuming and overcome the bias problem for the next layer. And adopt the ordinal classifier to replace Softmax to estimate the pollution level and it could capture the ranking information of the categories. Then on the basis of single channel, introduce an ensemble CNN for estimating air quality to get more stable results. At last propose a web service of hybrid measurement of air quality which based on the combination of physical station, image dictionary and convolution neural network. The experimental results show the method in this paper is low cost, large coverage area, convenient to carry and high accuracy. In addition the air quality measuring results in this chapter are also used to offer environmental feature information for preventative pipe maintenance. 4) Preventative urban pipe maintenance means using information like material type, pipe diameter, texture feature, environment information and past failure events to estimate pipe’s failure risk probability. The current methods are low accuracy. Focusing on these problems, this paper proposes preventative urban pipe maintenance based a mutual-exciting point process with profile-specific base intensity modeling. Firstly introduce the mathematical modeling for preventative urban pipe maintenance. Next devise a parametric model for the base intensity by the profile covariants and the model complexity is regardless of the instance size but depends on the number of covariants. The profile covariants include pipe’s inherent attributes, such as material, length, etc, visual texture features from contour detection, and environmental information such as air quality, temperature and humidity at the location of pipes. At last propose a mutual-exciting point process. It considers the impact of different failure types and models the triggering effect of event types. And devise an efficient optimization approach to solve this problem. The experimental results show comparing with other methods, the method in this paper achieves best performance.