As in many industries, the last few decades of Hydrographic surveying has seen exponential leaps in technology. Multibeam and phase discrimination swath sonar systems have evolved to the point where they collect several thousand samples every second, providing unprecedented coverage and quality in surveying the world’s oceans and inland waters. With this increase in collected data volume comes a corresponding increase in the time required to post-process this data. More recent industry trends have been correspondingly pushing towards simplification and automation to try and mitigate this significant bottleneck.
Despite these incredible leaps in technology, acoustic sonars can still produce a lot of noise in challenging environments, or when operated by inexperienced users. This means, in the worst-case scenario, much of the data must still be manually reviewed, or run through a processing algorithm like CUBE. Generally speaking, most users apply a combination of tools and algorithms to meet the particular challenges of their working area and hardware configuration. Given the volume of data involved, this can compound the time needed to process the collected datasets.
We have created a novel approach to tackle the challenge of cleaning noisy sonar data by using Deep Learning techniques, leveraging state-of-art architecture with a 3D convolutional neural network, trained using publicly available dataset meticulously cleaned in-house. We have focused primarily on identifying noise in our first model, and have since extended its use to other tasks, including identifying noise in other remote sensing paradigms. Our use of Deep Learning has also expanded to Bathymetric LiDAR, not only in identifying noise, but also classifying pulses as land or water. This ensures proper corrections are applied in processing and we achieve very high-quality data products.
Burns Foster
Burns Foster is the Innovation Manager for Teledyne Geospatial. His primary responsibility involves developing new products and services outside of Geospatial’ s core competencies, with a focus on novel applications of Machine Learning technologies across the entire Geospatial domain.