Predicting tree failure likelihood for utility risk mitigation via a convolutional neural network

Apostolov, Atanas, Oke, Jimi, Suttle, Ryan, Arwade, Sanjay, Kane, Brian 2023. Sustainable and Resilient Infrastructure 8(6):572–588.

Abstract

Critical to the resilience of utility power lines, tree failure assessments have historically been performed via costly manual inspections. In this paper, we develop a convolutional neural network (CNN) to predict tree failure likelihood categories (Probable, Possible, Improbable) under three classification strategies. The CNN produced the best performance under the Probable/Possible vs. Improbable strategy, achieving a recall score of 0.82. We also perform a visual analysis of the predictions via Grad-CAM++ heatmaps, indicating an approach for incorporating interpretability into model selection. Benchmarking the results of our model against those produced by two state-of-the-art CNNs (ResNet-50 and Inception-v3), we show that our relatively simple model produces better results in a computational time that is three times faster. Via this novel framework, we demonstrate the potential of artificial intelligence to automate and consequently reduce the costs of tree failure likelihood assessments in proximity to power lines, thereby promoting sustainable infrastructure.