IEEE Geoscience and Remote Sensing Magazine - March 2023 - 103

SOFTWARE AND DATA SETS
BURAK EKIM , TIMO T. STOMBERG ,
RIBANA ROSCHER , AND MICHAEL SCHMITT
MapInWild
A remote sensing dataset to address the question of what makes nature wild
T
he advancement in deep learning (DL) techniques
has led to a notable increase in the number and size of
annotated datasets in a variety of domains, with remote
sensing (RS) being no exception [1]. Also, an increase in
Earth observation (EO) missions and the easy access to
globally available and free geodata have opened up new
research opportunities. Although numerous RS datasets
have been published in the past years [2], [3], [4], [5], [6],
most of them addressed tasks concerning man-made environments,
such as building footprint extraction and
road network classification, leaving the environmental
and ecology-related subareas of RS underrepresented.
Nevertheless, environmental protection has always been
an important topic in the RS community, with RS being
a useful tool to support conservation policies and strategies
combating challenges such as deforestation and loss
of biodiversity [7], [8], [9]. Thus, in this article, to meet
the pressing need to better understand the nature we are
living in, we introduce a novel task of wilderness mapping
and advertise the MapInWild dataset [10]-a multimodal
large-scale benchmark dataset designed for the
task of wilderness mapping from space.
INTRODUCTION
Automated classification of image data has a long tradition
in EO. In this community, the classification task
can be addressed in different ways, most notably as a
scene- or patch-wise classification in contrast to pixelwise
classification. In scene classification, full scenes are
assigned by the classifier with single or multiple class labels,
whereas in pixel-wise classification (usually called
semantic segmentation by the computer vision community),
the task outputs densely annotated prediction
maps on a pixel scale by separating the input into distinct
and semantically coherent segments.
Digital Object Identifier 10.1109/MGRS.2022.3226525
Date of current version: 31 March 2023
MARCH 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
In general, image classification approaches either
are based on hand-crafted feature engineering and
subsequent machine learning (ML) models or feature
learning incorporated into the ML model in the form of
deep neural networks. In the RS area, the long-standing
tasks of scene classification and semantic segmentation
have been approached in a variety of settings [11],
ranging from metric learning [12] to multitask learning
[13]. While some methods frame the RS-related tasks
within the context of perturbation-seeking generative
adversarial networks [14], some others have made use of
uncertainty estimation applied to deep ensembles [15]
and self-attention context networks under adversarial
attacks [16].
The success of DL models comes at the expense of
decreased interpretability, which means the ability to
understand the decision process of the model and the
reason why a specific outcome was derived. This is
mainly caused by their formation of hundreds of successive
layers, leading to a high number of parameters.
In recent years, several studies addressed the lack of
explainability and interpretability of DL models and
proposed methods to overcome this challenge. Modelagnostic
approaches are independent of the used model
and can be applied post hoc. Popular approaches are,
for example, occlusion sensitivity maps [17], which
observe the change in the output while systematically
occluding small parts of the input; local interpretable
model- agnostic explanations [19], which approximate
the model with smaller models with fewer parameters;
and gradient-weighted class activation maps [18], which
combine the activation maps in a convolutional neural
network with class-specific gradients. Besides modelagnostic
approaches, model-specific approaches are tailored
to specific models. They mainly use parts of the
model, such as the weights, to analyze the decision process
or to interpret the model outcome. One example
is activation space occlusion sensitivity (ASOS), which
103
https://orcid.org/0000-0001-7014-1907 https://orcid.org/0000-0002-5452-4104 https://orcid.org/0000-0003-0094-6210 https://orcid.org/0000-0002-0575-2362
IEEE Geoscience and Remote Sensing Magazine - March 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2023
