Automating Visual Blockage Classification of Culverts with Deep Learning

Blockage of culverts by transported debris materials is reported as the salient contributor in originating urban flash floods. Conventional hydraulic modeling approaches had no success addressing problem primarily because unavailability peak floods data and highly non-linear behavior at culvert. This article explores a new dimension to investigate issue proposing use intelligent video analytics (IVA) algorithms for extracting blockage related information. The presented research aims automate process manual visual classification from maintenance perspective remotely applying deep learning models. potential using existing convolutional neural network (CNN) (i.e., DarkNet53, DenseNet121, InceptionResNetV2, InceptionV3, MobileNet, ResNet50, VGG16, EfficientNetB3, NASNet) investigated over dataset three different sources images culvert openings (ICOB), hydrology-lab (VHD), synthetic (SIC)) predict given image. Models were evaluated based on their performance test accuracy, loss, precision, recall, F1 score, Jaccard Index, region convergence (ROC) curve), floating point operations per second (FLOPs) response times single instance. Furthermore, models was benchmarked against conventional machine SVM, RF, xgboost). In addition, idea classifying features extracted CNN MobileNet) also implemented this article. From results, NASNet most efficient with 5-fold accuracy 85%; however, MobileNet recommended hardware implementation its improved time comparable 78%). Comparable standard achieved case where classified approaches. False negative (FN) instances, false positive (FP) instances layers activation suggested that background noise oversimplified labelling criteria two contributing factors degraded algorithms. A framework partial automation proposed, none able achieve high enough completely process. detection-classification pipeline higher 94%) has been proposed future direction practical implementation.