Outcome of the Paper Contest
The Paper Contest of the 2014 Data Fusion Contest, organized by the Image Analysis and Data Fusion (IADF) Technical Committee of the IEEE Geoscience and Remote Sensing Society was opened on February 19, 2014. The submission deadline was June 15, 2014. Participants submitted open-topic manuscripts using the data released for the Contest.
Evaluation and ranking were conducted by the Award Committee. The top-3 results are summarized below along with the abstracts of the submitted papers:
Title: Fusion of thermal infrared hyperspectral and VIS RGB data using guided filtering and supervised fusion graph
Authors: Wenzhi Liao (1), Frieke Van Coillie (2), Sidharta Gautama (1), Aleksandra Pizurica (1), and Wilfried Philips (1)
Affiliation: (1) Ghent University-TELIN-IPI-iMinds, Ghent, Belgium: (2) FORSIT Research Unit, Ghent University, Ghent, Belgium
Abstract. Nowadays, advanced technology in remote sensing allows us to acquire multi-sensor and multi-resolution data from the same geographic region. Fusion of these data sources for classification purposes however remains challenging. We propose a novel framework for fusion of low spatial resolution Thermal Infrared (TI) hyperspectral (HS) and high spatial resolution RGB data. First, we perform image fusion on the TIHS image by using visible RGB image and guided filtering in PCA (Principal Component Analysis) domain. Then, we couple feature extraction and data fusion of spectral features (from TI HS data) and spatial features (morphological features generated on RGB image) through a supervised fusion graph. Finally, the fused features are used by a SVM (Support Vector Machine) classifier to generate the final classification map. Experimental results on the classification of fusing real TI HS and RGB images demonstrate the effectiveness of the proposed method both visually and quantitatively.
Title: Decision-level fusion method based on sub-pixel mapping for multiresolution imagery classification
Authors: Shengwei Zhong, Ye Zhang, and Yidan Teng
Affiliation: Department of Information Engineering, Harbin Institute of Technology, China
Abstract. This paper describes a decision-level fusion classification framework of long-wave infrared (LWIR) hyperspectral images and fine resolution visible images. The proposed methodology consists of three sections. First, visible images are classified aided by spatial information. Second, a soft classification procedure is applied to obtain land cover fractions, followed by a subpixel mapping of these fractions based on the rule of weight majority voting. While the LWIR images contain abundant spectral information, visible images are also paid much attention to because of their extremely high resolution, which means abundant spatial information. The major novelty of this decision-level fusion approach is the inclusion of spatial information, obtained from the visible image. Experiments are conducted on the data sets of an urban area near Thetford Mines in Quebec, Canada, which are provided in the 2014 Data Fusion Contest. When compared to classification of the visible image by SVM and the spatial information aided strategy on visible image, the proposed algorithm shows better results. The advantage of the proposed approach is clearly demonstrated.
Title: Spectral-spatial fusion of multiresolution and multisensor images for classification
Authors: Sicong Liu (1) and Alim Samat (2)
Affiliation: (1) Department of Information Engineering and Computer Science, University of Trento, Trento, Italy; (2) Department of Geographical Information Science, Nanjing University, Nanjing, China
Abstract. This paper focuses on a challenging classification task by using the multiresolution and multisensor data, which include two data sets at different spectral ranges and spatial resolutions: i) a coarser-resolution long-wave infrared (LWIR, thermal infrared) hyperspectral (HS) data set; and ii) a very high resolution (VHR) data set acquired in the visible wavelength range. Both the data sets are provided by the Telops Inc. in IEEE GRSS 2014 Data fusion contest. The main purpose of this paper is to develop a proper way that fully utilizes two heterogeneous data sets and takes advantages of their characteristics thus benefiting the classification task. To this end, a classification approach based on the spectral-spatial feature analysis is proposed, which mainly consists of the following steps: 1) spatial-spectral feature extraction based on segmentation; 2) multi-level feature representation based on morphological profiles; 3) multiple features integration; 4) classification of the multi-feature set by using the rotation forest classification method. In particular, a graph-based segmentation is used for extracting the homogenous spatial features (objects) from the VHR data, whereas the principal component analysis is adopted for generating the spectral feature set from the HS data. Objects are found of their correspondences on both data sets and are described in a multi-level representation. Experimental results obtained on the considered multiresolution and multisensor data confirm the effectiveness of the proposed method, especially with respect to the use of limited training samples and to a low computational cost.