Automated Fault-Detection in Irrigation Distribution Networks using Convolutional Neural Networks and Drone Imagery

Abstract

Automated detection of faults in irrigation networks can facilitate network maintenance planning and reduce water loss and environmental impacts. The method proposed here combines deep learning with UAV hyperspectral imaging to detect damaged elements of irrigation networks. Convolutional neural networks are trained using a small dataset of images obtained under optimal conditions and labeled manually. The trained model is evaluated with images collected in subsequent field campaigns, and the results indicate a mean average precision of 91.53%, demonstrating the great potential of state-of-the-art deep learning models for detecting faults in irrigation distribution networks.

Localized fault detection constitutes a fundamental step towards preparing a maintenance action plan. Whether caused by environmental factors or undetected operational mistakes, these faults contribute to plant water stress, severe water loss, and environmental hazards while improving the spread of phytopathogens and pests. State-of-the-art deep learning algorithms detect faults in irrigation networks through the analysis of multitemporal images. To achieve real automation, current protocols combined with the training of a deep learning detection model must be transferred to convolutional neural networks.