IEEE Geoscience and Remote Sensing Magazine - March 2020 - 154

TECHNICAL COMMITTEES
NAOTO YOKOYA, PEDRAM GHAMISI, RONNY HÄNSCH, AND MICHAEL SCHMITT

2020 IEEE GRSS Data Fusion Contest:
Global Land Cover Mapping With
Weak Supervision

he 2020 Data Fusion Contest,
organized by the IEEE Geoscience and Remote Sensing Society (GRSS) Image Analysis and
Data Fusion Technical Committee
(IADF TC) and the Technical University of Munich, aims to promote
research in automatic large-scale
land cover mapping from globally
available multimodal satellite data
with weak supervision (Figure 1).
As in previous years [1]-[3], the contest will be set up as a benchmark
competition. To promote practical
innovation, the 2020 Data Fusion Contest will have two
sequentially organized challenge tracks:
◗◗ Track 1: land cover classification with low-resolution
labels
◗◗ Track 2: land cover classification with low- and highresolution labels.
To train the land cover prediction models, we use
the SEN12MS data set [4], which includes more than
180,000 triplets of corresponding Sentinel-1 syntheticaperture radar data, Sentinel-2 multispectral imagery,
and moderate-resolution imaging spectroradiometer
(MODIS)-derived low-resolution land cover maps [5]
sampled across the globe (Figure 2). While all data are
provided at a ground sampling distance of 10 m, the Sentinel images have a native resolution of approximately
10-20 m per pixel, whereas the MODIS-derived land
cover has a native resolution of 500 m per pixel. For
the contest, we use a simplified version of the international geosphere-biosphere program (IGBP) classification scheme [6], which has 10 land cover classes such
as forest, shrubland, savanna, grassland, wetlands, cropDigital Object Identifier 10.1109/MGRS.2020.2970124
Date of current version: 6 March 2020

154

land, urban/built-up areas, snow/
ice, barren terrain, and water. The
main challenge is to train classification models for high-resolution
land cover prediction (with a target
resolution of 10 m) from low-resolution and noisy annotations. The
validation and testing employ semimanually derived high-resolution
land cover maps of scenes with undisclosed geolocations. It should be
noted that the validation and testing samples are not contained in the
SEN12MS data set.
In Track 1, the semi-manually derived high-resolution land cover maps for the validation set are undisclosed. The objective is to predict land cover labels using
only low-resolution labels for training and hyperparameter tuning. In Track 2, we provide high-resolution labels for the validation set, and the goal is to train models
for land cover mapping using low-resolution labels and
a limited number of high-resolution labels. For both
tracks, performance is assessed through the average accuracy of all classes.
The 2020 Data Fusion Contest focuses on a fundamental issue related to weakly supervised learning: global generalization, which is a major challenge in a range
of fields, from Earth observation to computer vision and
machine learning. In conventional land cover classification, it is common to use human annotations for training
at the same resolution as the input image. However, this
approach is expensive and difficult to apply on a global
scale. The contest serves as a benchmark to evaluate the
best approaches in global land cover classification with
weak supervision. The large-scale training data set facilitates the development of deep-learning algorithms that
are generalized worldwide, and the globally sampled test
data set enables the generalization to be evaluated.
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

MARCH 2020

IEEE Geoscience and Remote Sensing Magazine - March 2020

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2020