Mar 25 2020
2:00 pm - 3:30 pm
Advanced Lidar Computing
Track Names: ADVANCED LIDAR COMPUTING, Wednesday 2:00 - 3:30
Session Date: Mar 25 2020 2:00 pm - 3:30 pm
Rethinking Spatial Database Advantages in a Petabyte Geospatial Environment
In recent years, spatial databases have fallen out of favor as a solution for the storage of large geo-spatial data sets from laser scanning and other remote sensing sources. In their place, file-based systems have gained prominence. Researchers and practitioners in this area have touted rapid data ingestion and availability of compatible software as prevailing justifications for a near abandonment of spatial database solutions, with the resulting side effect of such negativity being only minimal research work being undertaken in this area in the past decade. However, the increasing availability of both national aerial laser scans with densities up to 20 points/m2 and more localized kilometer-scale scans at densities of hundreds of data points per square kilometer challenge the presumption that a file-based system is superior to a spatial database solution, particularly with respect to random data access. Arguably, such querying is the most fundamental activity undertaken by most geo-spatial data users. As such, this presentation presents a novel spatial database solution for random access of three-dimensional point cloud data in a distributed computing environment.
New York University
A New Approach to Storing High Density Spatial Data
The LAS file format has been in use for nearly 20 years. The basic format has been updated through the years to accommodate new hardware collection systems as they have been developed. Over time, implementation has become more cumbersome as the format has grown to support data generated by the large variety new hardware platforms. LAS 1.4 introduced 64-bit capabilities to allow for extremely large file sizes. While this addition has helped in the management of large high-density projects, it has complicated the software developers’ task of quickly extracting regions of data for analysis and display. This work explores new approaches to storing large volumes of elevation data using a priori knowledge of underlying spatial data. The proposed solution, built on the industry standard HDF5 file format, seeks to address the needs of both software developers and large data aggregators. This is accomplished by providing high levels of data compression while maintaining the ability to quickly extract arbitrary regions of data with on the fly decompression.
Exploiting Deep Learning Techniques for Topo-Bathy Lidar Data Classification and Point Cloud Editing
2:30 PM - 2:45 PM
Airborne lidar topo-bathy systems are precision mapping instruments used to map the coastal zone on a regular basis to generate a wide range of products that include nautical charting, seamless topo-bathy maps, benthic habitat classification maps. Automatic land-water classification of lidar shots is required in order to apply refraction correction to water shots to generate accurate topo-bathy point cloud products. Typically, airborne lidar topo-bathy systems employ an infrared channel to map water surface and to discriminate land and water returns. However, low reflective environments like, wetlands, pose a significant challenge to commonly used statistical-based techniques that primarily exploit Infrared waveform return features to classify land and water shots. We have developed Land-water classifier based on Recurrent Neural Network (RNN) architecture to automatically discriminate land and water shots. Long short-term memory (LSTM) architecture is introduced to take advantage of the intensity relationship within a signal time series and subsequently a pooling operation is applied to incorporate all information together for the final prediction. Another aspect of Deep Learning application is the automation of the manual editing process, which is a critical part of the data processing workflow to generate noise-free final point cloud products. We have developed a Noise Classifier based on the 3D Convolutional Neural Network that has a strong capability to learn the feature representation of noise to automate the editing process by invalidating system noise and water column noise returns. We believe that application of Noise Classifier would significantly reduce the manual data editing time. In the presentation, we will discuss the results and improvements achieved by the application of the above models to lidar topo-bathy datasets.
Q&A and panel discussions with session presenters
There will be a 30 minute Q&A /panel discussion with the presenters of the Advanced Lidar Computing session.