Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Master → Pays Bas → 2021 → Dam Detection using Convolutional Neural Networks applied to Satellite Imagery

UNESCO-IHE Institute for Water Education, Delft (2021)

Dam Detection using Convolutional Neural Networks applied to Satellite Imagery

Ñustes Tovar, Paola Andrea

Titre : Dam Detection using Convolutional Neural Networks applied to Satellite Imagery

Auteur : Ñustes Tovar, Paola Andrea

Université de soutenance : UNESCO-IHE Institute for Water Education, Delft

Grade : Master of Science (MS) 2021

Résumé
Dams are one of the most popular hydraulic infrastructures, being used for thousands of years. They have allowed managing water resources for different purposes. While a dam is a very safe infrastructure, there is still the risk of incidents and failure, leading to extensive damage and casualties. According to the significance of this infrastructure and its associated risks and benefits, it is essential to keep a complete and accurate inventory of existing dams that facilitate dam safety investigations, hazard assessments, monitoring of the structure and its surroundings, among others. Nowadays, a significant part of the world-dams have been included in global databases such as the World Register of Dams (WRD) developed by the International Commission on Large Dams (ICOLD), the Global Reservoir and Dam database (GRanD), and the Global Georeferenced Database of Dams (GOODD). ICOLD, being the one including more dams, reports 58,000 large dams. However, just in the USA, there are about 90,000 dams listed in the National Inventory of Dams (NID), which means that a considerable percentage of the global dams remain unreported. Considering that the existing databases have been done mainly manually, which is time-consuming, and despite the efforts made, it is estimated that there is still a significant amount of unregistered dams, there is an urgent need for automatizing dam’s detection. This research provides an algorithm that, based on satellite imagery, can identify dams using image classification and neural networks ; this combination of techniques provides a compelling opportunity to access a large amount of accurate information worldwide. More specifically, we investigate the features that are important to classify dams in satellite imagery correctly. The development of a transparent methodology makes it replicable and opens the door to further applications, such as detecting dam elements and their dimensions, dam monitoring using high-resolution images, and ultimately generating an automated global dams’ dataset. Initially, the methodology has focused on a control group of dams in the USA since this region provides access to good quality data in terms of satellite imagery through the National Agriculture Imagery Program (NAIP) administered by The United States Department of Agriculture (USDA), digital elevation model from USGS National Elevation Dataset, and dams’ information from the NID. The satellite imagery and digital elevation models can be accessed through the Google Earth Engine platform. An algorithm to extract images of ‘dam’ and ‘no-dams’ is developed, allowing the creation of an input dataset to train, validate, and test a model for dams’ detection. The model is trained using MobileNetV2 architecture and is tested in 27 scenarios. The wrong predictions found are mainly attributed to images that are confusing even to humans. However, the algorithm’s main objective is to locate dams ; therefore results are satisfying since the results from the different scenarios result in accuracies between 85% and 94%.

Sujets  : dams satellite imagery neural networks dam detection

Présentation et version intégrale

Page publiée le 22 décembre 2021