Automated quantification of greenspace maintenance level assessment using Street View images and deep learning methods

Keywords: Urban greenspace, Street View images, Deep learning
Abstract Type: Paper Abstract

Abstract

Well-maintained urban greenspaces may attract users and enhance passive enjoyment of nature, leading to health benefits. However, traditional assessment of greenspace conditions is often resource-intensive and limits the ability to study a large area. Google Street View (GSV) imagery offers the potential to scale up assessment of greenspaces. Here, we adopted a deep learning method to assess signs of disorder (i.e., lack of maintenance, litter, vacancy, and graffiti) found in greenspaces. To do this, we annotated GSV images for signs of disorder (higher scores represent more disorder) for a set of 606 recreational, empty, or ‘other’ greenspace lots (e.g., urban farm, cemetery) in Detroit, MI. We used Pytorch to train our model on 500 images, leaving 106 test images. Our results showed better performance in estimating the level of maintenance (67% accuracy, average class F1 score of 0.66), compared to estimating the level of litter (64%, average class F1 score of 0.44). Although the results were better than random (33% accuracy), they were still suboptimal for estimation of greenspace disorder. This may be due to two factors: 1) the small size of the training dataset with only 500 images, and 2) class imbalance in the training set with few images showing high levels of litter. To improve performance, increasing the dataset size and striving for balance is crucial. If accuracy is improved, such an approach may assist urban planning efforts to maintain greenspaces by providing large-scale assessments, identifying maintenance areas, and enabling timely interventions for residents.