From sky to ground: Monitoring visible street greenery via multisource remote sensing imagery with deep learning

Abstract

The Green View Index (GVI) is a critical metric for assessing urban environments and has significant implications for enhancing residents’ well-being and promoting urban ecological health. Extracting GVI using street view images (SVIs) is a common practice. However, since urban street greenery generally undergoes a continuous change over time, solely using SVIs that are spatiotemporally sparse would fail to comprehensively capture the dynamics of GVI, potentially introducing biases into related research and policy-making. To fill that gap, this study proposes a new framework to estimate GVI using multi-source satellite imagery and deep learning technique. We first derive GVI information from SVI data collected along road networks and employ a spatiotemporal matching approach to establish paired samples linking GVI with remote sensing (RS) imagery. Subsequently, we develop a multi-task deep learning model with channel attention mechanisms for refined GVI estimation. Furthermore, leveraging the physical properties of GVI, we integrate multiple knowledge-based features with the raw RS bands as inputs to enhance the learning efficiency and overall performance of the model. Experimental results in Shanghai and Nanjing, China, show that the $R^2$ value of the GVI estimates against actual values can reach 0.716 at the point level and 0.834 at the street level. This study validates the feasibility of using satellite imagery to monitor GVI, offering robust technical and data foundations for evaluating urban greening and advancing environmental health studies.