Comparative performance of convolutional neural network,weighted and conventional support vector machine and random forest for classifying tree species using hyperspectral and photogrammetric data

ABSTRACT

The classification of tree species can significantly benefit from high spatial and spectral information acquired by unmanned aerial vehicles (UAVs) associated with advanced classification methods. This study investigated the following topics concerning the classification of 16 tree species in two subtropical forest fragments of Southern Brazil: i) the potential integration of UAV-borne hyperspectral images with 3D information derived from their photogrammetric point cloud (PPC); ii) the performance of two machine learning methods (support vector machine – SVM and random forest – RF) when employing different datasets at a pixel and individual tree crown (ITC) levels; iii) the potential of two methods for dealing with the imbalanced sample set problem: a new weighted SVM (wSVM) approach, which attributes different weights to each sample and class, and a deep learning classifier (convolutional neural network – CNN), associated with a previous step to balance the sample set; and finally, iv) the potential of this last classifier for tree species classification as compared to the above mentioned machine learning methods. Results showed that the inclusion of the PPC features to the hyperspectral data provided a great accuracy increase in tree species classification results when conventional machine learning methods were applied, between 13 and 17% depending on the classifier and the study area characteristics. When using the PPC features and the canopy height model (CHM), associated with the majority vote (MV) rule, the SVM, wSVM and RF classifiers reached accuracies similar to the CNN, which outperformed these classifiers for both areas when considering the pixel-based classifications (overall accuracy of 84.4% in Area 1, and 74.95% in Area 2). The CNN was between 22% and 26% more accurate than the SVM and RF when only the hyperspectral bands were employed. The wSVM provided a slight increase in accuracy not only for some lesser represented classes, but also some major classes in Area 2. While conventional machine learning methods are faster, they demonstrated to be less stable to changes in datasets, depending on prior segmentation and hand-engineered features to reach similar accuracies to those attained by the CNN. To date, CNNs have been barely explored for the classification of tree species, and CNN-based classifications in the literature have not dealt with hyperspectral data specifically focusing on tropical environments. This paper thus presents innovative strategies for classifying tree species in subtropical forest areas at a refined legend level, integrating UAV-borne 2D hyperspectral and 3D photogrammetric data and relying on both deep and conventional machine learning approaches.     

Multi-threaded compression of Earth observation time-series data

ABSTRACT

Earth observation data are typically compressed using general-purpose single-threaded compression algorithms that operate at a fraction of the bandwidth of modern storage and processing systems. We present evidence that recently developed multi-threaded compression codecs offer substantial benefits over widely used single-threaded codecs in terms of compression efficiency when applied to a selection of moderate resolution imaging spectroradiometer (MODIS) datasets stored in the HDF5 format. Compression codecs from the LZ77 and Rice families are shown to vary in efficacy when applied to different MODIS data products, highlighting the need for compression strategies to be tailored to different classes of data. We also introduce LPC-Rice, a new multi-threaded codec, that performs particularly well when applied to time-series data.     

Re-Freshing Cartography

Abstract

This paper presents the cartographic elements of a system for classifying and visualizing high-dimensional geographic datasets. The system has been developed as part of the Land Ocean Interactions in the Coastal Zone (LOICZ) project. The goal of the system is to develop regional and global typologies of coastal zones using large multi-variable datasets. Our implementation bring together statistical clustering algorithms with visualization capabilities to allow easy analysis and comprehension of the result. The two main tasks of the visualization are to allow for discrimination of multiple classes and to show relationships between those classes. These are accomplished in two different visual presentations. In both cases, the system selects colors appropriate to the purpose. In the latter case — showing relationships — the system uses a novel iterative refinement algorithm to select the colors. The result show that the system is successful at both generation the classes and visualizing the relationships between them.     

Assessing global Sentinel-2 coverage dynamics and data availability for operational Earth observation (EO) applications using the EO-Compass

ABSTRACT

Sentinel-2 scenes are increasingly being used in operational Earth observation (EO) applications at regional, continental and global scales, in near-real time applications, and with multi-temporal approaches. On a broader scale, they are therefore one of the most important facilitators of the Digital Earth. However, the data quality and availability are not spatially and temporally homogeneous due to effects related to cloudiness, the position on the Earth or the acquisition plan. The spatio-temporal inhomogeneity of the underlying data may therefore affect any big remote sensing analysis and is important to consider. This study presents an assessment of the metadata for all accessible Sentinel-2 Level-1C scenes acquired in 2017, enabling the spatio-temporal coverage and availability to be quantified, including scene availability and cloudiness. Spatial exploratory analysis of the global, multi-temporal metadata also reveals that higher acquisition frequencies do not necessarily yield more cloud-free scenes and exposes metadata quality issues, e.g. systematically incorrect cloud cover estimation in high, non-vegetated altitudes. The continuously updated datasets and analysis results are accessible as a Web application called EO-Compass. It contributes to a better understanding and selection of Sentinel-2 scenes, and improves the planning and interpretation of remote sensing analyses.     

Potential of Landsat-8 and Sentinel-2A composite for land use land cover analysis

Abstract

This study proposes the development of a multi-sensor, multi-spectral composite from Landsat-8 and Sentinel-2A imagery referred to as ‘LSC’ for land use land cover (LULC) characterisation and compared with respect to the hyperspectral imagery of the EO1: Hyperion sensor. A three-stage evaluation was implemented based on the similarity observed in the spectral response, supervised classification results and endmember abundance information obtained using linear spectral unmixing. The study was conducted for two areas located around Dhundi and Rohtak in Himachal Pradesh and Haryana, respectively. According to the analysis of the spectral reflectance curves, the spectral response of the LSC is capable of identifying major LULC classes. The kappa accuracy of 0.85 and 0.66 was observed for the classification results from LSC and Hyperion data for Dhundi and Rohtak datasets, respectively. The coefficient of determination was found to be above 0.9 for the LULC classes in both the datasets as compared to Hyperion, indicating a good agreement. Thus, these three-stage results indicated the significant potential of a composite derived from freely available multi-sensor multi-spectral imagery as an alternative to hyperspectral imagery for LULC studies.     

Opening the London Underground to visually handicapped tourists

Abstract

There are several practical rules for determining categories (class intervals) for maps representing statistical data, like arithmetic, geometric or equal steps etc. In this paper, however a coherent method is proposed to provide statistically separable Classes on a map with minimum redundancy in terms of information content.

The number of class intervals can be directly computed by means of appropriate statistical methods if the widths of classes are determined by t-test, i.e. when their difference is significant at a high level of confidence. A class narrower than this width would represent data in different categories due only to variance, however, the selection of wider classes leads to a certain loss of information.

The class intervals determined this way should be positioned on the statistical data-set so that each category contains approximately equal number of data providing maximum information content of the output map. At the final step the class intervals derived this way should be rounded, if necessary, to provide user-friendly maps.     

ColorBrewer.org: An Online Tool for Selecting Colour Schemes for Maps

Abstract

Choosing effective colour schemes for thematic maps is surprisingly difficult. ColorBrewer is an online tool designed to take some of the guesswork out of this process by helping users select appropriate colour schemes for their specific mapping needs by considering: the number of data classes; the nature of their data (matched with sequential, diverging and qualitative schemes); and the end-use environment for the map (e.g., CRT, LCD, printed, projected, photocopied). ColorBrewer contains 'learn more' tutorials to help guide users, prompts them to test-drive colour schemes as both map and legend, and provides output in five colour specification systems.     

FunnelCloud: a cloud-based system for exploring tornado events

ABSTRACT

Recent research has shown an increase in the number of extreme tornado outbreaks per year. The characterization of the spatio-temporal pattern of tornado events is therefore a critical task in the analysis of meteorological data. Currently, there are a large number of available meteorological datasets that can be used for such analysis. However, much of these data are distributed across multiple websites and are not accessible in a central location. This poses a significant challenge for a scientist who is interested in exploring meteorological patterns associated with tornado events. This paper presents a novel system which uses cloud-based technology for integrating, storing, exploring, analyzing, and visualizing meteorological data associated with tornado outbreaks. The system employs a novel NoSQL database schema and web services architecture for data integration and provides a user friendly interface that allows scientists to explore the spatio-temporal pattern of tornado events. Furthermore, scientists can use this interface to analyze the relationship between different meteorological variables and properties of tornado outbreaks using a number of spatio-temporal statistical and data mining methods. The efficacy of the system is demonstrated on a use case centered on the analysis of climatic indicators of large spatio-temporally clustered tornado outbreaks.     

Integrating spectral and non-spectral data to improve urban settlement mapping in a large Latin-American city

ABSTRACT

Information on urban settlements is crucial for sustainability planning and management. While remote sensing has been used to derive such information, its applicability can be compromised due to the complexity in the urban environment. In this study, we developed a remote sensing method to map land cover types in a large Latin-American city, which is well known for its mushrooming unplanned and informal settlements. After carefully considering the landscape complexity there, we designed a data fusion method combining multispectral imagery and non-spectral data for urban and land mapping. Specifically, we acquired a cloud-free Landsat-8 image and two non-spectral datasets, i.e., digital elevation models and road networks. Then, we implemented a set of experiments with different inputs to evaluate their merits in thematic mapping through a supervised protocol. We found that the map generated with the multispectral data alone had an overall accuracy of 73.3% but combining multispectral imagery and non-spectral data yielded a land cover map with 90.7% overall accuracy. Interestingly, the thermal infrared information helped substantially improve both the overall and categorical accuracies, particularly for the two urban classes. The two types of non-spectral data were critical in resolving several spectrally confused categories, thus considerably increasing the mapping accuracy. However, the panchromatic band with higher spatial resolution and its derived textural measurement only generated a marginal accuracy improvement. The novelties of our work are with the successful separation between the two major types of urban settlements in a complex environment using a carefully designed data fusion approach and the insight into the relative merits of the thermal infrared information and non-spectral data in helping resolve the issue of class ambiguity. These findings should be valuable in deriving accurate urban settlement information which can further advance the research on socio-ecological dynamics and urban sustainability.     

Fusion of Radar and Optical Data for Land Cover Mapping

Abstract

This study examined the complementarity of radar and optical data for feature identification. Spaceborne radar and Landsat Thematic Mapper (TM ) multispectral data sets were assessed independently and in combination to classify a site near Wad Medani, Sudan. Radar processing procedures included speckle reduction, texture extraction and post‐processing smoothing. Relative accuracy of the resultant classifications was established by comparison to ground truth information derived from field visitation. Neither speckle filtering nor post‐classification smoothing were improvements over the poor results obtained with the unfiltered, original radar data. Texture measures were significant improvements over the original data (20 percent overall accuracy increase) and several, but not all, individual classes had excellent results. Landsat TM had good overall results (80 percent correct) but considerable spectral confusion between urban and bare soil. Combination of radar with Landsat TM greatly improved results, achieving near perfect classification of all individual classes. The systematic strategy of this study, determination of the best individual method before introducing the next procedure, was effective in managing a complex set of analysis possibilities.     

Comparison and Integration of Radar and Optical Data for Land Use/Cover Mapping

Abstract

This study examined the complementarity of spaceborne radar and optical data for surface feature identification. RADARSAT data sets were assessed independently and in combination with Landsat Thematic Mapper (TM) multispectral data. The primary methodology was spectral signature extraction and the application of a statistical decision rule to classify the surface features for a site near Kericho, Kenya. Relative accuracy of the resultant classifications was established by digital integration and comparison to reference information derived from field visitation. Speckle filtering was a great improvement over the poor results achieved with the unfiltered, original radar data but still not adequate for accurate land cover classification. The extraction and use of Variance texture measures was found to be very advantageous. The overall results were not significant improvements over speckle removal (6% increase) but several individual classes, forest and urban, had excellent results with texture. Combinations of radar with Landsat TM greatly improved results, achieving near perfect classification of all individual classes. The highest overall accuracy was achieved with a merger that included the best individual texture image and six reflectance bands of the TM data. The systematic strategy of this study, determination of the best individual method before introducing the next procedure, was effective in managing a very complex, almost infinite set of analysis possibilities.     

A procedural footprint enhancement of global topographic surface with multiple levels of detail

ABSTRACT

Virtual globes are technologies for visual navigation through a three-dimensional, multi-resolution model of the entire planet. Data representations used in virtual globes, however, lack geometric flexibility at high-resolution levels of the planet-wide terrain surface. This is a problem especially if boundaries between individual geospatial features and the terrain are important. A novel integration of individual polygonal boundaries with a specific multi-resolution representation of the planet-wide terrain is developed in this article. In the preparation stage, the integration relies on an original simplification algorithm applied to the polygonal boundaries between geospatial features and the terrain. Its output is a multiple level-of-detail (LOD) geometry, which can be combined with a known multi-LOD representation of the terrain that uses run-time triangulation. This data representation is suitable for storage in existing database systems, avoids any data redundancy across LODs, and is even independent of the subdivision schema that partitions the planet's surface for the sake of dealing with LODs. At run-time, a novel reconstruction algorithm stitches geometric parts from different LODs together in a manner that augments the multi-LOD representation of the terrain. Within a certain proximity range from a given position, the method reconstructs a scene that preserves topological relations between the boundaries of geospatial features with the terrain. The method also guarantees that certain nearest proximity to the given position consists of the best geometries that correspond to the original datasets. Such properties of the method close up the gap between a mere exploratory visualization of static, pre-generated models and the models supporting geospatial analysis, which is deemed crucial for applications in Geographic Information Systems, Building Information Modelling and other software industries. A prototype implementation and experiment results that prove this method are also presented.     

Developments in digital map production and data integration within the British Geological Survey

Abstract

Digital techniques for cartographic data capture and high quality map production have been developed and applied over some 18 years to the mapping of the geoscientific datasets of the British Geological Survey, in particular to the geochemical dataset. Over this period, technological advances and developments in both vector and raster data processing techniques have facilitated high quality map production and the integration and display of multiple datasets. This paper reviews the developments in high quality map production and interpretation of survey information, with particular regard to the Regional Geochemical Atlas Series, through the application of image processing techniques on the display and analysis of multiple datasets.     

Applying VGI to collaborative research in the humanities: the case of ARTL@S

ABSTRACT

Using Artl@s as an example of a project that relies on volunteered geographic information (VGI), this article examines the specific challenges that exist, beyond those frequently discussed in general VGI systems (e.g., participants’ motivation and data quality control) in regard to sharing research data in humanities: (1) most data from the humanities is qualitative and collected from multiple data sources which are often inconsistent and unmappable; (2) data is usually interconnected with multiple relationships among different tables which creates challenges for both mapping and query functionality; (3) data is both geographical and historical. Consequently addresses that no longer exist have to be geolocated and visualized on historical basemaps and spaces must be represented diachronically; (4) the design of web map application needs to balance both sophisticated research requirements and a user-friendly interface; (5) finally contributors expect their data to be cited or acknowledged when used in other studies and users need metadata and citation information in order to reuse and repurpose datasets.

In this article, we discuss how Artl@s, a project which developed a georeferenced historical database of exhibition catalogues, addresses these challenges. Artl@s provides a case study for VGI adoption by digital humanities scholars for research data sharing, as it offers features, such as flexible batch data contribution, interrelated spatial query, automatic geolocalization of historical addresses, and data citation mechanisms.     

Validating the use of object-based image analysis to map commonly recognized landform features in the United States

ABSTRACT

The U.S. Geological Survey (USGS) National Geospatial Program (NGP) seeks to i) create semantically accessible terrain features from the pixel-based 3D Elevation Program (3DEP) data, and ii) enhance the usability of the USGS Geographic Names Information System (GNIS) by associating boundaries with GNIS features whose spatial representation is currently limited to 2D point locations. Geographic object-based image analysis (GEOBIA) was determined to be a promising method to approach both goals. An existing GEOBIA workflow was modified and the resulting segmented objects and terrain categories tested for a strategically chosen physiographic province in the mid-western US, the Ozark Plateaus. The chi-squared test of independence confirmed that there is significant overall spatial association between terrain categories of the GEOBIA and GNIS feature classes. Contingency table analysis also suggests strong category-specific associations between select GNIS and GEOBIA classes. However, 3D visual analysis revealed that GEOBIA objects resembled segmented regions more than they did individual landform objects, with their boundaries often failing to correspond to match what people would likely perceive as landforms. Still, objects derived through GEOBIA can provide initial baseline landscape divisions that can improve the efficiency of more specialized feature extraction methods.     

Feature selection for airborne LiDAR data filtering: a mutual information method with Parzon window optimization

ABSTRACT

Filtering is one of the key steps for Digital Elevation Model (DEM) generation from airborne Light Detection and Ranging (LiDAR) data. Machine-learning-based filters have emerged as a class of filtering algorithms in recent years. Most existing studies mainly focus on feature generation due to limited available features a point cloud possesses. More than 30 features have been described in the existing literature. But most generated features are based on geometric information of points. Several redundant and irrelevant features may not necessarily improve the filtering accuracy. Hence, this paper proposes a feature-selection method using minimal-Redundancy-Maximal-Relevance (mRMR) combined with Parzen window optimization to deal with both discrete and continuous features. An optimal/suboptimal feature subset is constructed for machine-learning filters in various landscapes. Experimental results based on AdaBoost show that height-related features, particularly height itself, are of the greatest significance in both urban and rural scenes. Moreover, different subsets can be selected from the datasets of the two landscapes by our feature-selection strategy, which increases the data relevance for describing each geographical landscape. This study provides guidelines for the selection of optimal/suboptimal features for point cloud filtering based on machine-learning algorithms.     

Resolution Control for Balancing Overview and Detail in Multivariate Spatial Analysis

Abstract

Parallel coordinates, re-orderable matrices, and dendrograms, widely used methods for visual exploration of multivariate data, are systematically integrated in a complementary manner for supporting multi-resolution visual data analysis with an enhanced overview + detail exploratory strategy. There are three main topics: (1) dynamic control across resolutions at which data are explored; (2) coordination and color mapping among the views; and (3) enhanced features of each view designed for the overview + detail exploratory tasks. A case study analysis is used to demonstrate the potential for boosting productivity for exploration tasks by coordinating the views through user-controlled resolutions within a highly interactive analysis environment. The case study is focused on a complex, geographically referenced dataset including public health, demographic and environmental components.     

New measures for analysis and comparison of shape distortion in world map projections

ABSTRACT

World maps can have quite different depictions of reality depending on the projection adopted, and this can influence our perception of the world. In this respect, shape is a significant property that needs to be considered, especially when representing large regions in general-purpose world maps. A map projection distorts most geometric properties (area, distance, direction/angle, shape, and specific curves) and usually preserves a single property or provides a compromise between different properties when transforming terrestrial features from globe to plane. The distortions are mainly classified based on area, distance and direction/angle and analyzed with Tissot’s theorem. However, this theorem offers a local (pointwise) solution, so the distortion assessment is valid at infinitesimal scale (i.e. for very small regions). For this reason, different approaches are required to analyze the distortions at finite scale (i.e. for larger regions). However, there are very few attempts at analyzing and comparing shape distortion of landmasses in world map projections owing to the fact that shape measurement is difficult and usually involves measuring different characteristics. Seeking to fill this gap, in this study, compactness and elongation distortion measures are introduced. In this regard, 16 world map projections are analyzed and compared with these distortion measures in a GIS environment, based on map datasets of continents and countries. An analysis of the effect of the levels of detail of the datasets is also presented.     

Stakeholder analysis of the governance framework of a national SDI dataset – whose needs are met in the buildings and address register of the Netherlands?

ABSTRACT

National spatial data infrastructures are key to achieving the Digital Earth vision. In many cases, national datasets are integrated from local datasets created and maintained by municipalities. Examples are address, building and topographic information. Integration of local datasets may result in a dataset satisfying the needs of users of national datasets, but is it productive for those who create and maintain the data? This article presents a stakeholder analysis of the Basisregistratie Adressen en Gebouwen (BAG), a collection of base information about addresses and buildings in the Netherlands. The information is captured and maintained by municipalities and integrated into a national base register by Kadaster, the Cadastre, Land Registry and Mapping Agency of the Netherlands. The stakeholder analysis identifies organisations involved in the BAG governance framework, describes their interests, rights, ownerships and responsibilities in the BAG, and maps the relationships between them. Analysis results indicate that Kadaster and the municipalities have the highest relative importance in the governance framework of the BAG. The study reveals challenges of setting up a governance framework that maintains the delicate balance between the interests of all stakeholders. The results provide guidance for SDI role players setting up governance frameworks for national or global datasets.