Deep learning and data integration for detecting trees entangled with utility lines

Abstract

Urban image classification is a challenging task due to factors such as occlusions, diverse appearances, and ambiguous visual contexts. This paper addresses these challenges by consolidating and discussing methodologies focused on tree-powerline interactions, a critical problem for urban safety and infrastructure planning. Using state-of-the-art Deep Learning Networks (DLNs), prior studies evaluated performance on a curated dataset of 11,000 labeled street-level images collected and annotated with the INvestigate and Analyze a CITY (INACITY) platform and the Street-Level Image Labeler (SLIL) tool. These evaluations revealed significant limitations in baseline models, with a maximum accuracy of 74.6\% on this dataset. To address these limitations, earlier work introduced a ``Challenging'' class and applied semi-supervised learning techniques, including the Noisy Student protocol and Focal Loss, achieving recall rates of 83.7\% for positive cases and 78.8\% for negative cases. Dimensionality reduction using the Self-Supervised Neural Projection (SSNP) method was explored for clustering and visualization tasks, while the Supervised Decision Boundary Maps (SDBM) technique improved classifier interpretability by addressing scalability and clarity in decision boundary visualizations. This paper consolidates these contributions, presenting an integrated narrative that highlights tools, datasets, and methodologies for advancing urban image classification. These efforts provide scalable, interpretable, and accurate solutions for urban safety and infrastructure management challenges.