Multi-temporal analysis of urban vegetation using deep learning and 3D reconstruction

Abstract

Abstract Context Urban green spaces play a vital role in enhancing environmental quality and human well-being. However, traditional assessment methods, such as the green view index, primarily quantify green coverage while neglecting vegetation diversity, color richness, and seasonal dynamics, which are critical for urban livability. Objectives This study develops a multi-temporal and multi-perspective analysis framework for urban green space visualization, introducing the Seasonal Species-Specific Plant View Index (S3PVI) to quantify plant coverage at the species level, capturing seasonal changes and visual diversity. Methods The framework integrates computer vision, deep learning, and 3D reconstruction technologies, including structure from motion and 3D Gaussian splatting. To validate the S3PVI, case studies were conducted in Suita City, Japan, analyzing real-world seasonal vegetation patterns and testing the framework in a virtual park environment to assess its applicability in urban design. Results The S3PVI effectively captured species-specific seasonal patterns, with cherry blossoms peaking at 45.61% visibility in spring and maples at 56.78% in autumn. Comparative analysis revealed distinctive vegetation strategies between streets, with Sanshikisaido showing higher seasonal amplitude but lower consistency than Nakayoshido. Virtual simulations confirmed that multi-species schemes optimally balanced seasonal impact with year-round visual stability. Conclusions The S3PVI framework advances urban vegetation assessment by providing species-specific and seasonally dynamic visual data, supporting evidence-based urban planning for ecological sustainability and livability. Potential applications include brownfield redevelopment, virtual park planning, and urban design simulations.

Plant phenotyping relevance

植物の種別・季節別被覆を定量化するS3PVIを開発し、深層学習・コンピュータビジョン・3D再構成で検証しており、植物状態の取得・抽出手法が中心である。

Code and data availability

The paper's street view imagery, 3D reconstructions, segmentation outputs, and S3PVI datasets are not publicly deposited; the Data availability statement says they are available from the corresponding author on reasonable request. No author code or model repository URL is provided.

No evidence-backed public reproduction asset is currently recorded.