Global Land Cover Mapping with Weak Supervision
The 2020 Data Fusion Contest, organized by the Image Analysis and Data Fusion Technical Committee (IADF TC) of the IEEE Geoscience and Remote Sensing Society (GRSS) and the Technical University of Munich, was opened on December 13, 2019. The final phase was opened on March 1, 2020, and closed on March 20. There were 159 registrations on the Codalab website and 33 teams entered the final phase. The final ranking is available at https://competitions.codalab.org/competitions/22289.
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 their approaches and metrics over all the undisclosed ground truth samples.
Track 1: Land cover classification with low-resolution labels
1st Place
Author: Caleb Robinson, Kolya Malkin, Lucas Hu, Bistra Dilkina, Nebojsa Jojic
Affiliation: Georgia Institute of Technology, Yale University, University of Southern California, Microsoft Research
Codalab account: calebrob6
Approach: A combination of iterative clustering and epitome representations, followed by deep image prior post-processing. No use of SEN12MS.
Metric (AA): 0.5749
2nd Place
Author: Yu Xia, Yue Liao, Hongyan Zhang, Guangyi Yang
Affiliation: Wuhan University
Codalab account: WHU_YuXia
Approach: Multi-branch fusion of unsupervised multi-resolution segmentation, random forest classification of remote sensing indexes, and convolutional neural network predictions with post-processing based on expert priors. Used SEN12MS.
Metric (AA): 0.5696
3rd Place
Author: Daniele Cerra, Nina Merkle, Corentin Henry, Kevin Alonso, Pablo d’Angelo, Stefan Auer, Reza Bahmanyar, Xiangtian Yuan, Ksenia Bittner, Maximilian Langheinrich, Guichen Zhang, Miguel Pato, Jiaojiao Tian, Peter Reinartz
Affiliation: German Aerospace Center (DLR)
Codalab account: Pineapples
Approach: Automated label pre-processing, a Gaussian Naive Bayes classifier trained on cluster centroids, and classes obtained by k-means clustering and random forests with bag of words features, followed by classification refinement designed for specific classes. Used SEN12MS.
Metric (AA): 0.5688
4th Place
Author: Huijun Chen, Changlin Xiao, Wei Liu, Rongjun Qin
Affiliation: The Ohio State University
Codalab account: Antonia
Approach: Automated label pre-processing, random forests, followed by classification refinement based on prior knowledge on class confusion. No use of SEN12MS.
Metric (AA): 0.5676
Track 2: Land cover classification with low- and high-resolution labels
1st Place
Author: Huijun Chen, Changlin Xiao, Wei Liu, Rongjun Qin
Affiliation: The Ohio State University
Codalab account: Antonia
Approach: An ensemble of random forests trained on refined labels. No use of SEN12MS.
Metric (AA): 0.6142
2nd Place
Author: Daniele Cerra, Nina Merkle, Corentin Henry, Kevin Alonso, Pablo d’Angelo, Stefan Auer, Reza Bahmanyar, Xiangtian Yuan, Ksenia Bittner, Maximilian Langheinrich, Guichen Zhang, Miguel Pato, Jiaojiao Tian, Peter Reinartz
Affiliation: German Aerospace Center (DLR)
Codalab account: Pineapples
Approach: As Track 1 third, but random forests trained on high-resolution labels for validation data, and no use of topic vectors and bag of words features.
Metric (AA): 0.6136
3rd Place
Author: Shuting Yin, Dafan Chen, Chengconghui Ma, Yanchao Lian
Affiliation: Xidian University
Codalab account: dfchen
Approach: A combination of random forests, k-means, and DeepLabv3++ with postprocessing and retraining. Used SEN12MS.
Metric (AA): 0.6095
Each team will present their methods in a special session at the forthcoming IGARSS conference in Hawai, US. Their papers will be included in the IGARSS proceedings.