A Landscape-Aware Spatial Representation Learning for Understanding Plural Values of Nature (Poster)

Abstract

Landscape Ecology (LE) is an academic field that quantitatively analyses land use and cover patterns to elucidate ecological systems and interpret human sociocultural and economic activities. This study aims to apply the Natural Language Processing (NLP) methodology in the LE. LE analyses a) spatial arrangements of landscape elements within unit spaces, b) landscapes as aggregates of these units, and c) their significance in natural and social-ecological systems. This is similar to NLP’s approach: a’) sentences as word aggregates, b’) texts as sentence aggregates, and c’) document and contextual meanings. While NLP has evolved from count-based methods to advanced techniques like word embeddings and large language models, LE has primarily relied on count-based analyses of landscape elements. This limitation has made it challenging to capture the full spectrum of nature’s values, particularly the relational value between humans and nature, one of the important values of nature (intrinsic, instrumental, and relational). To address this gap, this study reviewed spatial representation learning techniques developed primarily in transportation and explored their potential applications in LE. Then, the hex2vec algorithm, a word2vec-based representation learning using OpenStreetMap, was selected to embed the landscape structure of the entire Japan and tested some downstream tasks of LE. From the results, limitations of the current spatial representation learning and future directions were discussed.