基于Siamese卷积神经网络的影像瓦片变化检测技术

针对地理信息变化较快而传统更新方式效率不高的问题,目前许多学者提出了各类变化检测的方法,但这些方法大都是基于影像数据进行试验,对影像预处理要求较高,且检测精度的稳定性较差,受数据源质量影响较大。而天地图、百度地图、谷歌地图等地图中均可免费下载各种级别的影像瓦片,因此本文提出利用天地图影像瓦片进行试验,采用Siamese卷积神经网络（SCNN）和深度学习技术,开发基于SCNN的高精度变化监测算法,以快速发现变化区域,实现地理信息变化信息检测。     

改进注意力机制的遥感地貌识别算法

对遥感地貌进行识别,近年来一直是遥感图像应用领域的研究热点。使用深度学习算法识别遥感影像具有比传统方法更高的准确率和稳健性。针对遥感影像中目标复杂度高、特征信息多等问题,本文提出了一种基于改进注意力机制的遥感图像识别算法,即将并联注意力机制（CS）和神经网络模型相结合,借助弱监督学习来辅助训练。同时采用双损失函数来缓解数据过拟合问题。试验结果表明,本文模型总精度为98.35%,Kappa系数达0.95,优于其他深度学习算法,能有效地识别出自然地貌。     

基于R-FCN_ResNet的深海冷泉区生物的识别与分布研究

针对深海冷泉生物数量多、密度大、特征低,人工识别它们的种类和数取它们的数量耗时费力且准确度低这一问题,利用残差学习块,改进了基于区域的全卷积网络算法,用以深海冷泉生物的自动识别与检测。首先,构建了一组包括5类目标生物的深海冷泉生物图像目标检测数据集;然后,在TensorFlow框架下对比了R-FCN_ResNet、Faster R-CNN和SSD_MobileNet三种典型的卷积神经网络算法。从训练耗损时间、平均检测速度、平均置信度三方面权衡,突出了R-FCN_ResNet模型的优势;最后,将测试集图片输入至该模型中检测深海冷泉生物的种类和数量,并且以此复现了生物的分布情况。实验证明,所提方法结合本文数据集进行的深海冷泉生物识别与分布研究,是有效可行的,且具有较高的目标分类和定位准确率。     

深度学习在高分辨率遥感影像冬油菜提取中的应用

近年来,深度学习在基于高分辨率遥感影像的农作物种植信息提取领域应用广泛。本文充分利用油菜在盛花期的光谱特征,提出了基于深度学习理论的单时相高分辨率遥感影像油菜分布提取方法。以2016年湖北省沙洋县作为研究区域,获取油菜盛花时期高分一号（GF-1）影像,并以沙洋县为基础影像,通过手工标记制作油菜训练样本。设计两种深度学习框架模型,一种以卷积神经网络（CNN）为框架,构建一维卷积神经网络（1D-CNN）模型,第二种以循环神经网络（RNN）为框架,组合门控循环单元（GRU）模型,训练标准样本模型,完成油菜分类提取。最后,与传统支持向量机（SVM）、随机森林（RF）方法进行了结果对比。试验结果表明,本文设计的基于深度学习CNN和RNN模型提取的冬油菜空间分布精度和面积精度皆优于其他两种方法,为进一步实现冬油菜提取自动化提供试验基础。     

Maintaining accurate,current, rural road network data: An extraction and updating routine using RapidEye,participatory GIS and deep learning

Accurate and current road network data is fundamental to land management and emergency response, yet challenging to produce for unpaved roads in rural and forested regions using traditional cartographic approaches. Automatic extraction of roads from satellite imagery using deep learning is a promising alternative gaining increasing attention, however most efforts have focused on urban paved roads and used very high spatial resolution imagery, which is less frequently available for rural regions. Additionally, road extraction routines still struggle to produce a fully-connected, vectorized road network. In this study covering a large forested area in Western Canada, we developed and evaluated a routine to automatically extract unpaved road pixels using a convolutional neural network (CNN), and then used the CNN outputs to update a pre-existing government road network and evaluate if and how it would change. To cover the large spatial extent mapped in this study, we trained the routine using moderately high-resolution satellite imagery from the RapidEye constellation and a ground-truth dataset collected with smartphones by organizations already operating and driving in the region. Performance of the road extraction was comparable to results achieved by others using very high-resolution imagery; recall accuracy was 89–97%, and precision was 85–91%. Using our approach to update the pre-existing road network would result in both removals and additions to the network, totalling over 1250 km, or about 20 % of the roads previously in the network. We discuss how road density estimates in the study area would change using this updated network, and situate these changes within the context of ongoing efforts to conserve grizzly bears, which are listed as a Threatened species in the region. This study demonstrates the potential of remote sensing to maintain current and accurate rural road networks in dynamic forest landscapes where new road construction is prevalent, yet roads are also frequently de-activated, reclaimed or otherwise not maintained.     

Evaluations and comparisons of rule-based and machine-learning-based methods to retrieve satellite-based vegetation phenology using MODIS and USA National Phenology Network data

Vegetation phenology is a sensitive indicator that reflects the vegetation–atmosphere interactions and vegetation processes under global atmospheric changes. Fast-developing remote sensing technologies that monitor the land surface at high spatial and temporal resolutions have been widely used in vegetation phenology retrieval and analysis at a large scale. While researchers have developed many phenology retrieving methods based on remote sensing data, the relationships and differences among the phenology retrieving methods are unclear, and there is a lack of evaluation and comparison with the field phenology recoding data. In this study, we evaluated and compared eight phenology retrieving methods using Moderate Resolution Imaging Spectroradiometer (MODIS) and the USA National Phenology Network data from across North America. The studied phenology retrieving methods included six commonly used rule-based methods (i.e., amplitude threshold, the first-order derivative, the second-order derivative, the third-order derivative, the relative change curvature, and the curvature change rate) and two newly developed machine learning methods (i.e., neural network and random forest). At the large scale, the start of the season (SOS) values, derived by all methods, had similar spatial distributions; however, the retrieved values had large uncertainties in each pixel, and the end of the season (EOS) inverted values were largely different among methods. At the site scale, the SOS and EOS values extracted by the rule-based methods all had significant positive correlations with the field phenology observations. Among the rule-based methods, the amplitude threshold method performed the best. The machine learning methods outperformed the rule-based methods in terms of retrieving the SOS when assessed using the field observations. Our study highlighted that there were large differences among the methods in retrieving the vegetation phenology from satellite data and that researchers must be cautious in selecting an appropriate method for analyzing the satellite-retrieved phenology. Our results also demonstrated the importance of field phenology observations and the usefulness of the machine learning methods in understanding the satellite-based land surface phenology. These findings provide a valuable reference for the future development of global and regional phenology products.     

Assessing the benefit of satellite-based Solar-Induced Chlorophyll Fluorescence in crop yield prediction

Large-scale crop yield prediction is critical for early warning of food insecurity, agricultural supply chain management, and economic market. Satellite-based Solar-Induced Chlorophyll Fluorescence (SIF) products have revealed hot spots of photosynthesis over global croplands, such as in the U.S. Midwest. However, to what extent these satellite-based SIF products can enhance the performance of crop yield prediction when benchmarking against other existing satellite data remains unclear. Here we assessed the benefits of using three satellite-based SIF products in yield prediction for maize and soybean in the U.S. Midwest: gap-filled SIF from Orbiting Carbon Observatory 2 (OCO-2), new SIF retrievals from the TROPOspheric Monitoring Instrument (TROPOMI), and the coarse-resolution SIF retrievals from the Global Ozone Monitoring Experiment-2 (GOME-2). The yield prediction performances of using SIF data were benchmarked with those using satellite-based vegetation indices (VIs), including normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), and near-infrared reflectance of vegetation (NIRv), and land surface temperature (LST). Five machine-learning algorithms were used to build yield prediction models with both remote-sensing-only and climate-remote-sensing-combined variables. We found that high-resolution SIF products from OCO-2 and TROPOMI outperformed coarse-resolution GOME-2 SIF product in crop yield prediction. Using high-resolution SIF products gave the best forward predictions for both maize and soybean yields in 2018, indicating the great potential of using satellite-based high-resolution SIF products for crop yield prediction. However, using currently available high-resolution SIF products did not guarantee consistently better yield prediction performances than using other satellite-based remote sensing variables in all the evaluated cases. The relative performances of using different remote sensing variables in yield prediction depended on crop types (maize or soybean), out-of-sample testing methods (five-fold-cross-validation or forward), and record length of training data. We also found that using NIRv could generally lead to better yield prediction performance than using NDVI, EVI, or LST, and using NIRv could achieve similar or even better yield prediction performance than using OCO-2 or TROPOMI SIF products. We concluded that satellite-based SIF products could be beneficial in crop yield prediction with more high-resolution and good-quality SIF products accumulated in the future.     

采用双注意力机制Deeplabv3+算法的遥感影像语义分割

针对DeepLabv3+网络在遥感影像上呈现出拟合速度慢,边缘目标分割不精确,大尺度目标分割类内不一致、存在孔洞等缺陷,提出在该网络中引入双注意力机制模块（Dual Attention Mechanism Module, DAMM）,设计并实现了将DAMM结构与ASPP（Atous Spatial Pyramid Pooling）层串联或并联的2种不同连接方式网络模型 ,串联连接方式中先将特征图送入DAMM后,再经过ASPP结构;并联连接方式中将双注意力机制层与ASPP层并行连接,网络并行处理主干网提取特征图,再融合两层处理特征信息。将改进的2种方法通过INRIA Aerial Image高分辨率遥感影像数据集验 证,结果表明,串联或并联方式2种网络都能有效改善Deeplabv3+的不足,并联方式网络性能更好,其对原网络缺陷改善效果更明显,并在测试数据集上mIoU达到85.44%,比Deeplabv3+提高了1.8%,而串联方式网络提高了1.12%。并联结构网络更符合本文需求,其形成了一种对DeepLabv3+网络上述问题进行统一改善的方案。     

A lightweight ensemble spatiotemporal interpolation model for geospatial data

ABSTRACT

Missing data is a common problem in the analysis of geospatial information. Existing methods introduce spatiotemporal dependencies to reduce imputing errors yet ignore ease of use in practice. Classical interpolation models are easy to build and apply; however, their imputation accuracy is limited due to their inability to capture spatiotemporal characteristics of geospatial data. Consequently, a lightweight ensemble model was constructed by modelling the spatiotemporal dependencies in a classical interpolation model. Temporally, the average correlation coefficients were introduced into a simple exponential smoothing model to automatically select the time window which ensured that the sample data had the strongest correlation to missing data. Spatially, the Gaussian equivalent and correlation distances were introduced in an inverse distance-weighting model, to assign weights to each spatial neighbor and sufficiently reflect changes in the spatiotemporal pattern. Finally, estimations of the missing values from temporal and spatial were aggregated into the final results with an extreme learning machine. Compared to existing models, the proposed model achieves higher imputation accuracy by lowering the mean absolute error by 10.93 to 52.48% in the road network dataset and by 23.35 to 72.18% in the air quality station dataset and exhibits robust performance in spatiotemporal mutations.     

基于编解码网络的航空影像像素级建筑物提取

建筑物提取在城市规划等土地利用分析中发挥着重要作用。用于提取建筑物的传统方法通常基于手工特征和分类器,导致精度较低。本文基于编解码结构的卷积神经网络CNN(Convolutional Neural Networks),自主学习多级的和具有区分度的特征来更好地辨识建筑物和背景,实现航空影像中的像素级建筑物提取。该网络由编码子网络和解码子网络两部分组成,编码子网络对输入图像进行空间分辨率压缩,完成特征提取;解码子网络从特征中提升空间分辨率,完成像素级的建筑物提取。此外,本文使用视野增强FoVE(Field-of-View Enhancement)方法减轻边缘现象(切片边缘附近的建筑物提取精度通常低于中心区域附近的精度)的影响,并分别在两个建筑物提取标准数据集上的实验表明,编解码卷积神经网络能有效实现像素级建筑物提取,FoVE能有效提高建筑物提取准确率;通过改变预测时切片大小和重叠度,分析其对建筑物提取结果的影响,揭示了FoVE的饱和性。