Zhou Guo

On combining multiscale deep learning features for the classification of hyperspectral remote sensing imagery

In recent years, satellite imagery has greatly improved in both spatial and spectral resolution. One of the major unsolved problems in highly developed remote sensing imagery is the manual selection and combination of appropriate features according to spectral and spatial properties. Deep learning framework can learn global and robust features from the training data set automatically, and it has achieved state-of-the-art classification accuracies over different image classification tasks. In this study, a technique is proposed which attempts to classify hyperspectral imagery by incorporating deep learning features. Firstly, deep learning features are extracted by multiscale convolutional auto-encoder. Then, based on the learned deep learning features, a logistic regression classifier is trained for classification. Finally, parameters of deep learning framework are analysed and the potential development is introduced. Experiments are conducted on the well-known Pavia data set which is acquired by the reflective optics system imaging spectrometer sensor. It is found that the deep learning-based method provides a more accurate classification result than the traditional ones.

Mining parameter information for building extraction and change detection with very high-resolution imagery and GIS data

Although much efforts have been made to develop automatic methods for building extraction from very high-resolution (VHR) imagery during the past 30 years; the methods with high performance are still unavailable due to the three issues: uncertainty of segmentation scales, selection of effective features, and sample selection. In this study, by introducing GIS data, a parameter mining approach is proposed to (1) mine parameter information for building extraction, and (2) detect changes of buildings between VHR imagery and GIS data. For the first target, the learning mechanism is proposed for identifying optimal segmentation scales, feature subsets, and samples. For the second target, the discovered information (i.e., optimal segmentation scales, feature subsets, and selected samples) is applied to classify the VHR imagery with a multilevel random forest (RF) classifier. The proposed approach is validated on two datasets: Dataset 1 and Dataset 2. The knowledge of building extraction is first learned from Dataset 1 and then used to classify both datasets, and change detection is conducted on Dataset 1. Results of change detection in Dataset 1 indicate that the false alarm ratio and omission error of increased buildings are 20.1% and 8.4%, while the false alarm ratio and omission error of destroyed buildings are 19.1% and 11.3%, respectively. Results of building extraction in Dataset 2 revealed scores of 81.50% and 81.09% at pixel- and object-based evaluation levels. Accordingly, our proposed method is successful in building extraction and change detection.

A Direction-Guided Ant Colony Optimization Method for Extraction of Urban Road Information From Very-High-Resolution Images

Typical object-based classification methods only take image object properties as criteria to classify roads, leaving the associated edge information unused. These methods often lead to fragmented road areas and inconsistent road widths and smoothness. Meanwhile, very-high-resolution (VHR) images contain a large amount of edge information and different types of geographic objects, thus, it is challenging to extract roads by typical edge-based extraction or grouping methods. In this study, a globally optimized method is developed to integrate both object and edge features to extract urban road information from VHR images. This novel method extends ant colony optimization (ACO) through deploying and moving ants (artificial agents) along roads with the guidance of comprehensive object and edge information. As ants spread pheromone along their paths, roads are recognized based on aggregated pheromone levels. A set of experiments on VHR images showed that our method significantly outperforms object-based classification methods with not only improved road extraction quality but also enhanced stability when applied to large and complex images.

A comparative study of the segmentation of weighted aggregation and multiresolution segmentation

Multiresolution segmentation (MRS) algorithm has been widely used to handle very-high-resolution (VHR) remote sensing images in the past decades. Unfortunately, segmentation quality is limited by the dependency of parameter selection on users’ experience and diverse images. Contrarily, the segmentation by weighted aggregation (SWA) can partly overcome the above limitations and produce an optimal segmentation for maximizing the homogeneity within segments and the heterogeneity across segments. However, SWA is solely tested and justified with digital photos in computer vision field instead of VHR images. This study aims at evaluating SWA performance on VHR imagery. First, multiscale spectral, shape, and texture features are defined to measure homogeneity of image objects for segmentation. Second, SWA is implemented to handle QuickBird, unmanned aerial vehicle (UAV), and GF-1 VHR images and further compared with MRS in eCognition software to demonstrate the applicability of SWA to diverse images in building, vegetation and water, forest stands, farmland, and mountain areas. Third, the results are fully evaluated with quantitative measurements on segmented objects and classification-based accuracy assessment on geographic information system vector data. The results indicate that SWA can produce higher quality segmentations, need fewer parameters and manual interventions, create fewer segmentation levels, incorporate more features, and obtain larger classification accuracy than MRS.

Exploring GIS knowledge to improve building extraction and change detection from VHR imagery in urban areas

Existing studies for building extraction in very high resolution (VHR) images consider little prior knowledge, and thus they have limited accuracies and can only apply to a small proportion of buildings. Combining VHR and GIS data to extract buildings is a feasible way to overcome the drawbacks above. However, the inaccurate positions of GIS data and time changes between the two data make it difficult to fuse them. This study aims at presenting the methods to resolve the position and time inconsistencies between the two data for extracting buildings and finding the changes. The methods begin with line extraction from VHR images. For extracting buildings, a two-level graph with four steps was presented to fuse GIS contours and the extracted line segments from VHR images, including choosing initial building edges, clustering the chosen edges, generating the building hypotheses and scoring building hypotheses. The time inconsistencies (i.e. the changes between the buildings in two data sources) were first identified by a scoring function. Then, the detected inconsistent buildings were further verified by the z-significance test from a statistical perspective. Experimental assessments performed on 10 typical urban regions indicated that the proposed methods are highly robust and convincing.

Polygonal Clustering Analysis Using Multilevel Graph-Partition

Existing methods of spatial data clustering have focused on point data, whose similarity can be easily defined. Due to the complex shapes and alignments of polygons, the similarity between non-overlapping polygons is important to cluster polygons. This study attempts to present an efficient method to discover clustering patterns of polygons by incorporating spatial cognition principles and multilevel graph partition. Based on spatial cognition on spatial similarity of polygons, four new similarity criteria (i.e. the distance, connectivity, size and shape) are developed to measure the similarity between polygons, and used to visually distinguish those polygons belonging to the same clusters from those to different clusters. The clustering method with multilevel graph-partition first coarsens the graph of polygons at multiple levels, using the four defined similarities to find clusters with maximum similarity among polygons in the same clusters, then refines the obtained clusters by keeping minimum similarity between different clusters. The presented method is a general algorithm for discovering clustering patterns of polygons and can satisfy various demands by changing the weights of distance, connectivity, size and shape in spatial similarity. The presented method is tested by clustering residential areas and buildings, and the results demonstrate its usefulness and universality.

An Extended Random Walker Approach for Object Extraction by Integrating VGI Data and VHR Image

Automatic extraction of objects in urban areas from very-high-resolution (VHR) images is of great significance to many applications. Existing approaches consider little information on spatial relationships, backgrounds, and prior knowledge of target objects, leading to that they did not perform well in object extraction. Now free and fast growing volunteered geographic information (VGI) can be accessed easily; thus, they can be used as prior information to improve the performance. This study develops an extended random walker (RW) approach to form a bottom-up and top-down mechanism for extracting target objects by combining VHR images and VGI data. Novel aspects of our approach include: 1) both the shape and spectral prior terms are incorporated into the extended RW algorithm; 2) an end-to-end framework is proposed to automatically select both foreground and background seeds with the assistance of VGI data; and 3) the shape prior of VGI data provides top-down information to select background and foreground seeds and help fuse bottom-up image information (i.e., foreground and background seeds and spatial relationships) to extract target objects. The extended RW approach was validated on building and lake datasets, and its performance is evaluated on both pixel and object levels. Quantitative comparisons with the original RW and random forest (RF) algorithm indicate that the proposed approach achieves significant better performance. Besides, it can successfully extract the partly occluded buildings.

Using random walker for knowledge transfer in classifying multi-temporal VHR images

Acquiring land cover types from very high resolution (VHR) images is of great significance to many applications and has been intensively studied for many years. The difficulties in image classification and the high frequencies of remote sensing image acquisition make it urgent to develop efficient knowledge transfer approaches for understanding multi-temporal VHR images. This letter proposed a knowledge transfer approach that uses the label information of the existing VHR images to classify multi-temporal images. The approach was implemented in three steps: object-based change detection, knowledge transfer of label information, and random walker (RW) classification. The proposed approach was tested by two datasets with each having two temporal images acquired on the same geographical areas. The experimental results showed that the proposed approach outperformed the support vector machine (SVM) algorithm in classifying multi-temporal images and can reduce the influence of spectral confusions on image classification.

A graph-based approach for the co-registration refinement of very-high-resolution imagery and digital line graphic data

ABSTRACT Co-registration refinement of very-high-resolution (VHR) imagery and digital-line-graphic (DLG) data is an important procedure before data fusion and analysis. However, existing approaches either make little consideration of topological relations between features or have to extract complete objects, which is very challenging. In this study, to overcome the drawbacks mentioned above, a graph-based approach is presented for the co-registering of VHR imagery and DLG data. Our proposed method uses a graph to represent the topological relations between buildings in both data sources, which helps match buildings in the two data sources and compute the affine transformation parameters. The proposed method is validated on three diverse VHR images, and two objective evaluation metrics (correctness and quality rate) are computed to evaluate its performance. It is shown that correctness and quality rate are averagely improved by 37.3% and 46.7%, respectively, after co-registration. These results indicate that our proposed method is effective in the co-registration of VHR imagery and DLG data.

High resolution urban image classification combining edge statistical features

Classification with very high resolution (VHR) urban images is challenging because of the great variations of spectrums of pixels inside objects. Plenty of structural information can be obtained over edge statistics. A methodology for incorporating image edge statistical information into conventional classification algorithms is described. The technique is built on the statistical information of edges which are generated by edge statistical model. This method has been tested on a selected site of Worldview-II data which covers north-west part of Beijing, China. Nine land-cover types have been classified to evaluate the effectiveness of edge-based features for urban image classification. The overall classification accuracy is 82.7% and 89.3% for pixel-based and object-based method for incorporating edge statistical features, respectively.