IEEE Geoscience and Remote Sensing Magazine - December 2022 - 155

inference. Hashtags were used as anchor words to work out
topics around certain hashtags and therefore support the
topic inference. This method could be useful in geospatial
research, e.g., event detection, where certain hashtags are
related to a specific event. The tweets without that hashtag
could therefore be utilized as well.
The LDA has been used in various studies in the geosciences.
For example, the authors in [37] used tweets and Flickr
images for a multilabel land use classification in New York
and San Francisco. They applied the LDA to extract relevant
topics related to Foursquare venues in specified clusters. The
relevant clusters were calculated by Hierarchical DensityBased
Spatial Clustering of Applications with Noise [38] to
identify local hot spots in the Flickr images. Before they applied
the LDA, the text was preprocessed by removing stop
words and applying lemmatization. Only those tweets within
the clusters detected before were used. The LDA revealed
the relevant topics for each cluster, which led to good classification
performance. The LDA was also used in [39] to identify
relevant topics of the 2012 Olympic Games in London
within the context of city planning. First, the tweets were preprocessed
with the aforementioned methods; thereafter, they
used the LDA to filter the tweets that were about the Olympic
Games and transportation. Tweets that were related to the
mentioned topics were used to perform a sentiment analysis
and a spatial-temporal analysis. The work in [40] investigated
the tracking and monitoring of Twitter topics related to a disaster
over time. As the LDA is not suitable for tracking topics
over time, they used a dynamic topic model [41], which is
based on the LDA.
Another classic but still very popular [42] algorithm is term
frequency-inverse document frequency (TF-IDF) [43]. Basically,
TF-IDF measures the relevancy of a term (e.g., a word)
in a document and weighs the document's importance, for
example, for a search query. This weighing process involves
the combination of two different steps. First, the standard way
of determining the TF of a certain term t in a document d is to
calculate the quotient of the total count of t appearing in d by
the complete count of all terms in d. Second, a stop word like
" the " would distort document weighing because it is likely to
appear very frequently within a document. Therefore, the IDF
decreases the importance of the frequently recurring terms.
This is achieved by taking the logarithm of the quotient of the
total number of documents and the count of d including that
certain term t. The final TF-IDF score for a term is computed
by the product of the TF and the IDF.
SENTIMENT ANALYSIS
As pointed out in the " Social Media Data Availability " section,
social media data can also contain emotions. One rulebased
algorithm used to determine sentiments in English
(but not limited to) Twitter text data is the Valence Aware
Dictionary for sEntiment Reasoning (VADER) [44]. The
authors compiled a list of sentiment-loaded terms. Those
terms can be words like " happy " or emoticons like " :-). " Furthermore,
the authors added support for emoji-sentiment
DECEMBER 2022 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
detection. The terms are validated and ranked by humans
(Amazon Mechanical Turk). In the end, the list comprises
7,520 of such terms. Every term in the list comes with a
mean sentiment intensity ranging from −4 (very negative)
to 4 (very positive) and 10 independent human ratings.
Next, a qualitative approach was used to detect the main
textual drivers of the perceived sentiment intensity. With
this, five heuristics incorporating grammatical and syntactical
clues are derived to determine the sentiment intensity
of a string or possible changes of the sentiment. For example,
exclamation points add some sentiment intensity to a
string, as does using all caps to stress words that express
the intended sentiment. Furthermore, booster words like
" very " also contribute to the computation of the sentiment
intensity. VADER achieved good performance and was able
to outperform human reviewers in some cases.
EMBEDDINGS
Today, the field of NLP has seen a strong shift toward ML
and particularly deep learning approaches which, in many
cases, now outperform previous methods based in linguistics
[45]. The topic of social media analysis is no exception
here, and many large-scale systems now employ deep neural
networks [46].
A crucial issue that caused the relatively late introduction
of deep learning into text data analysis (as opposed to
other forms of data, like images) lies in how to represent
words numerically. Traditionally, this has been done with
one-hot encodings that do not capture semantic meaning.
The key development here was the introduction of word
embeddings, which are neural networks themselves and
a part of the complete classification network. Some very
successful early approaches that are still in use today are
word2vec [47], GloVe [48], and fastText [49]. One of the
earliest, word2vec is a neural-based architecture that can
embed large vocabularies from huge text corpora into an
n-dimensional feature space very quickly. It can preserve
semantic and syntactic features of words and embeds
similar words close together in the feature space. This development
was followed by GloVe, a statistics-based approach.
To solve the issue of out-of-vocabulary (OOV)
words, Joulin et al.[49] introduced the fastText algorithm,
which divides words into subwords. This process facilitates
the approximation of OOV word vectors by composing
a word vector based on the word's subwords, which is
particularly useful for irregular text as it is often found
on social media. In each case, it is very common to use
pretrained embeddings, such as those provided for fastText
in 157 languages [50].
Urban areas are multilingual spaces [51], [52], [53], and
the set of languages discovered in social media posts in cities
is diverse [54] (of course, English is dominant on Twitter
[55] and Chinese on Weibo). However, multilingual approaches
are rare in earth sciences and thus offer interesting
research opportunities, such as ensemble models covering
all languages to classify building functions in urban areas
155
IEEE Geoscience and Remote Sensing Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2022
