Wind disasters adaptation in cities in a changing climate: A systematic review

When referring to the research questions, keywords related to three aspects (“city”, “wind disasters”, and “adaptation”) were developed. We did not include “climate change” in the list of keywords for two reasons. The first reason is not to limit wind disasters to climatic ones as the wind disaster itself is first and foremost, a type of natural disaster. Whether climate change could alter the frequency and intensity of wind disasters is yet to be determined. Therefore, adaptation to wind disasters does not necessarily take climate change into account. The second reason is not to connect wind disasters and climate change intentionally. We expect to see cli- mate change emerge as the new keyword in the results to examine if current research has inte- grated climate change into adaptation. Based on the authors’ experiences from browsing the literature, potentially relevant words were gradually gathered and refined. The results were sorted according to relevance after each search query was performed. Experts screened the titles and abstracts of the one hundred most relevant pieces of literature to examine the rele- vance and validity of the keywords. Phrases were connected by “OR” within the category and by “AND” between categories. The period was from the beginning of 1900 to the end of 2019. We chose all languages and all document types within the timeframe to comprehensively include all possibly relevant publications. It is important to note here that only literature with open access was downloaded together with their abstracts, authors, journals and year of publi-

cation, etc. The composition of keywords and the final search queries are provided in the S1 Appendix. After cleaning up the duplicate literature, the primary corpus was imported to ABSTRACKR, a website developed for screening publications for systematic reviews. Irrele- vant publications were identified and deleted after screening through the abstracts of all litera- ture. Before the formal screening, two experts reviewed 20 identical pieces of literature separately. Then, the results were compared and discussed to clarify the principles for litera- ture to be reserved. Generally, literature that has nothing to do with cities, wind disasters, or fails to focus on adaptation-related themes, was removed from the corpus. Currently, there is no standard quality assessment for individual environmental studies. Tools applied in the medical field are not suitable for this review because we target descriptive contents of literature that are relevant to our topic. Our quality control is reflected in the process of literature screen- ing where we made judgments according to three inclusion criteria for each article. They are i) The article’s discussion unfolds at the city level. ii) The main disaster discussed in the article are wind-related disasters. For example, tropical cyclones, tornadoes, sandstorms, and dust storms. iii) Adaptation should be the subject of the article. Topic modelling and content analysis We ran a topic model based on the Latent Dirichlet Allocation (LDA) to excavate topics dis-

cussed by the literature corpus. LDA is one of the most popular techniques in data mining, and latent topic exploring is applied across multiple fields [45]. LDA realizes dimensionality reduction to preserve the essential latent information by assigning a fixed number of topics and corresponding possibilities to each literature [46]. Previous literature reviews have done topic extraction in a similar way [47, 48]. LDA is a suitable algorithm for the analysis because it does not allow subtracted elements to be negative [49], and it assumes Dirichlet prior at the document level, therefore making it more flexible in representing and interpreting the litera- ture corpus [50]. As a pre-processing measure, we deleted letters that do not belong to the English language system (literature in Spanish has abstracts in both Spanish and English) and meaningless words such as stop words in detected root words. The model then calculated frequencies for the remaining root words and used them to generate topics with root words (set as 6 in this study) assigned to the same topic according to the frequency of two words occurring together. PLOS ONE Review on wind adaptation in cities PLOS ONE | https://doi.org/10.1371/journal.pone.0248503 March 17, 2021 4 / 26