Outcome of the Classification Contest

The Classification 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 March 5, 2014, i.e., participants had two weeks to generate and submit their classification maps by using 1-m resolution thermal hyperspectral and 20-cm resolution color data.

42 teams worldwide participated to the Classification Contest. Ranking was based on the kappa statistics computed, for each submitted classification map, with respect to undisclosed test samples. The top-10 results are summarized below along with short abstracts, which describe the key ideas of the various algorithms and were submitted by the participants together with the corresponding classification maps. The color legend for the classes in the maps is as follows:

unclassified road trees red roof grey roof concrete roof vegetation bare soil

The test samples will be disclosed after the end of the Paper Contest.

1st Place

Authors:Huang Xin, Liu Hui, and Zhu Tingting

Affiliation: State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, P. R. China

E-mail: lhui@whu.edu.cn

A hierarchical classification framework is designed for this contest. First, PCA is applied to LWIR data and the first 5 bands are kept. Then all data are resampled to 0.5 m. Various textural features and two indices, vegetation index and morphological building index, were extracted. Second, trees and vegetation are first detected as a whole and then discriminated. Then road, soil and roofs are identified successively with different combinations of features. Finally, an object-based classification map is obtained by fusing with the mean shift segmentation result at 0.2 m.

Kappa statistics:

2nd Place

Authors: Chunlei Weng and Shiyun Guo

Affiliation: State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, P. R. China

E-mail: wengchunlei@126.com

The multi-scale segmentation map (three levels) is created vertically on the visible image. Trees and Vegetation are firstly extracted using multi-features (3 visible bands and 4 GLCM textural features) at scale 1. Roads, concrete-roofs and the other three classes (red roof, grey roof and bare soil) are subsequently identified at scale 2 and scale 3 by multi-features (visible with 4 PCA bands extracted from the LWIR image) with a shape-based post-processing (e.g., length-width ratio). Finally the objects classified as roofs are checked with shape feature (e.g., length>550,  geometrical index<0.3), and redefined by non-roof classes.

Kappa statistics:

3rd Place

Authors: Jiayi Li, Min Guo, Wei He, Tao Shuai, Hongyan Zhang, and Liangpei Zhang

Affiliation: State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, P. R. China


The classification algorithm description is given as follows:
1. Pre-processing: Noise removal on the LWIR image.
2. Road extraction: 1) Coarse road extraction by combining the LWIR image classification and the RGB image segmentation; 2) Dubious pixels removal with pseudo-vegetable index (P-VI); 3) Meticulous road extraction via RGB wavelet feature based classification result.
3. Classification: Utilize the spatial and attribute features of the high-spatial image in its RGB, Ycbcr and HSV space and the P-VI to construct 9 classification maps for the remaining classes.
4. Post-processing: Majority vote the 9 classification maps and obtain the final object-oriented classification result for the whole image.

Kappa statistics:

4th Place

Authors: Harini Sridharan and Anil Cheriyadat

Affiliation: Geographic Information Science and Technology Group, Oak Ridge National Lab, USA

E-mail: sridharanh@ornl.gov

Preprocessing: Thermal correction for the LWIR data was done using the In-Scene Atmospheric Compensation algorithm. Noise reduction was performed using kernel-MNF transformation and band reduction was performed using piecewise constant approximation. Spectral (Greenness Index) and spatial (Gabor, Pixel Shape Index) derivatives were obtained from the visible data.
Multi-Level schema: The 7 classes were grouped into 5 levels of hierarchical schema and a separate SVM classifier was developed for each level using the most suitable subset of features for each level. Each SVM classifier was modified with a spatial kernel to include contextual information during classification and avoid expensive step of post-processing. The context for each pixel was obtained from image segments using multi-resolution segmentation. The probability maps for each class were obtained and merged to get the final results.

Kappa statistics:

5th Place

Authors: Xuehua Guan, Qikai Lu, and Dawei Wen

Affiliation: State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, P.R. China

E-mail: qikai_lu@hotmail.com

The visible image (0.2-m) is firstly segmented to delineate the boundaries. Afterwards, the first three principal components of the LWIR image with GLCM and 3D-DWT textures extracted from the down-sampled VIS image (1-m) are used as spectral-spatial features. The visible, three PCs, eight-dimensional GLCM and eight-dimensional 3D-DWT features (1-m) are stacked and classified by an SVM. MRF is then utilized to refine the initial result. The classification result (1-m) is projected into the 0.2-m boundary map. Finally, the roofs that were wrongly identified as roads are further rectified by considering the properties of area and length-width ratio.

Kappa statistics:

6th Place

Authors: Yanfei Zhong, Ji Zhao, Bei Zhao, and Liangpei Zhang

Affiliation: State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, P. R. China

E-mail: zhongyanfei@whu.edu.cn

The spectral and spatial information is integrated by using the principal component analysis (PCA), gray-level co-occurrence matrix (GLCM), mean shift segmentation, support vector machine (SVM) and conditional random field (CRF) methods. The hyperspectral image is first upsampled to the 0.2m spatial resolution, and then the PCA approach is utilized to reduce the dimension of spectral features. Moreover, the color image, the texture feature obtained by GLCM, the spatial feature is also used. The SVM and CRF methods are integrated to classify the image based on the obtained features.

Kappa statistics:

7th Place

Authors: Xudong Kang, Shutao Li, and Leyuan Fang

Affiliation: Vision and Image Processing Laboratory, Hunan University

E-mail: xudong_kang@hnu.edu.cn

First, the visible image is down-sampled to equal the size of the visible and hyperspectral images. Then, the spectral dimension of the hyperspectral image is reduced with PCA. The first 20 PCs and the down-sampled visible image are combined for classification.Using a two stage scheme, the features are classified with SVM. The obtained SVM probabilities are optimized with extend random walkers. The two-stage scheme aims at learning more training samples. The optimization step considers the spatial information in the visible image. Finally, the resulting probabilities are up-sampled and refined with guided filtering to obtain the classification map.

Kappa statistics:

8th Place

Authors: Lily Lee and Geoffrey Brown

Affiliation: MIT Lincoln Lab, USA

E-mail: leel@ll.mit.edu and geoffrey.brown@ll.mit.edu

We first apply an edge-based multi-modal image registration to align the LWIR data with each swath of the visible data then train a random forest classification model on a combination of the visible image and the four largest principal components of the LWIR data.  To correct for temperature differences between the swaths, we normalize the LWIR data over average intensity of each swath then classify using our random forest model.  We then re-normalize the LWIR data swaths using only road intensities, then re-classify. A segmentation of visible image is applied to the pixel classification to reduce noise.

Kappa statistics:

9th Place

Authors: Lauren Stachowiak, Jessica Moehl, Sarah Lewis-Gonzles, Yan Li, Yuan Liu, Ephraim Love, Joseph Roberts, Erik Schmidt, Kelly Sims, Vi Tran, and Nicholas Nagle

Affiliation: University of Tennessee, USA

E-mail: nnagle@utk.edu

We chose a support vector machine (SVM) approach to classify hyperspectral thermal and multispectral visible range data.  Features used included the first three bands from the MAF-MNF filtered LWIR data, the red-green-blue spectral signatures, GLCM statistics calculated from a panchromatic image, log-Gabor filters calculated at 4 different scales, and the Structural Feature Set.  All input data was resampled to the spatial resolution of the visible raster data. The models were calibrated using bootstrapped training data and each class was derived in stages based on highest class probability (red – gray, vegetation – tree, bare soil – road, concrete). Intermediate label outputs from the SVM were then post-processed using contextual, cell-based statistics.

Kappa statistics:

10th Place

Authors: Wenzhi Liao, Frieke Van Coillie, Sidharta Gautama, Rik Bellens, Aleksandra Pizurica, and Wilfried Philips

Affiliation: W. Liao, S. Gautama, R. Bellens, A. Pizurica, and W. Philips are with iMinds-Telin-IPI, Faculty of Engineering, Ghent University. F. Van Coillie is with FORSIT Research Unit, Faculty of Bioscience Engineering, Ghent University

E-mail: wenzhi.liao@telin.ugent.be

Nowadays, advanced technology in remote sensing allows to retrieve multi-sensor and multi-resolution data from the same geographic region. Fusion of these data sources for land cover classification remains challenging. We propose a novel fusion framework for low spatial-resolution TIHS and high spatial-resolution RGB data. First, we perform super-resolution on the TIHS data using PCA and the RGB data. Then, we couple feature extraction and data fusion of spectral features (from TIHS data) and spatial features (morphological features generated on RGB data) through graph embedding. Finally, the fused features are used by SVM classifiers to generate the final classification map.

Kappa statistics: