The 2017 IEEE GRSS Data Fusion Contest, organized by the IADF TC in collaboration with WUDAPT and GeoWiki, was opened on January 9, 2017. The test phase (and the evaluation server) was opened on March 13 and closed on April 2nd. We received over 800 submissions!
We would like to thank all the participants for their submissions. Evaluation and ranking were conducted by the Organizers, who carefully checked all valid submissions. The winners are reported below along with a brief overview of their methods and their Overall Accuracy (OA) and Kappa statistic over all the undisclosed ground truth pixels of the four test cities.

 

1st Place

Team Name: WXYZ
Authors: Naoto Yokoya, Pedram Ghamisi, Junshi Xia (University of Tokyo, DLR, and TU München, Japan – Germany)
Approach: Ensemble classifier (including Canonical Correlation Forests and Rotation Forests) over Landsat8 images and OpenStreetMap data.
Metrics: OA = 74.94 kappa = 0.71

 

2nd Place

Team Name: AGTDA
Authors: Sergey Sukhanov, Roel Heremans, Ivan Tankoyeu, Jérôme Louradour, Darya Trofimova, Christian Debes (AGT International, Switzerland)
Approach: Mixture of models (random forests, boosting and deep learning) regularized by conditional random fields over all the available data (Landsat8, Sentinel2 and OpenStreetMap). Also used extra open data (Landsat8 and OpenStreetMap).
Metrics: OA = 72.63 kappa = 0.68

 

3rd Place

Team Name: Camilasa
Authors: Camila Souza dos Anjos Lacerda, Marielcio Gonçalves Lacerda, Leidiane do Livramento Andrade, Roberto Neves Salles (Institute of Advanced Studies – Brazilian Air Force, Brazil),
Approach: Decision trees and random forests using spectral indices and PCA/MNF features. The classification is performed at the object level, with objects obtained by image segmentation.
Metrics: OA = 72.38 kappa = 0.68

 

4th Place

Team Name: Nanjingxyy
Authors: Yong Xu, Fan Ma, Deyu Meng, Chao Ren, Yee Leung (Chinese University of Hong Kong and Xi’an Jiaotong University, China)
Approach: Extreme gradient boosting, self-paced co-training and support-vector machines, using Sentinel2, Landsat8 and OpenStreetMap data in separate feature extraction streams.
Metrics: OA = 69.89 kappa = 0.65

Each team will present their method in a special session at the forthcoming IGARSS conference in Fort Worth (http://www.igarss2017.org/) that will take place on Tuesday July 25 (session TU4.L6: IEEE GRSS Data Fusion Contest). Their papers will be included in the IGARSS proceedings.