This session explores how artificial intelligence and advanced analytics are reshaping lidar processing workflows. Presentations demonstrate deep learning approaches for hydrographic and forestry feature extraction alongside new methodologies for evaluating ground support, surface smoothness, and overall data quality across diverse landscapes.
10:30 AM – 10:45 AM
Using Deep Learning to Extract Hydrographic Features from Lidar Data
Presented by Colin Flynn, Dewberry
Traditionally, hydrographic features are extracted manually through a digitization process. In geographies with complex hydrography, for example braided river deltas, this inefficient process can become expensive and time-consuming. Various methods for automating the extraction of hydrographic features exist for satellite and aerial imagery; these methods typically take advantage of spectral indices and thresholding. However, when working with lidar data, it is common practice to utilize the derived terrain in this endeavor; unfortunately, the results from such efforts are inconsistent and unreliable. By utilizing the source lidar data as a starting point and deriving various raster products, we develop custom composite images that allow us to consistently visually identify hydrographic features. We are able to solve this complex image classification problem using the intrinsic nature of lidar scanners and hydrographic principles to emphasize target features. This allows for the subsequent training of a deep learning model for robust pixel-wise classification. The result of the model is a classified raster from which polygons are derived. Since the model classifies each pixel, we have additional control over how model predictions are filtered and post processed. This allows us to provide many hydrographic features well below specification for our end users, resulting in greater accuracy for any subsequent modeling based on the lidar data. Over the past years, this process has been refined and streamlined. Our team has leveraged open-source libraries and multiple programming languages, leading to dramatic improvements at production scale. In this newly improved process, lasR (developed by r-lidar), Whitebox Tools, and GDAL are used for rigorous and efficient processing of 1000’s of square miles of data at a time, while ensuring our compute resources are utilized effectively.
10:45 AM – 11:00 AM
AI-Powered Forest Classification and Tree Inventory Workflows in LP360
Presented by Thadeu Lugarini, GeoCue Group
Artificial Intelligence is transforming lidar-based forest analysis by automating the extraction of meaningful information from dense point clouds. This presentation will demonstrate how LP360 leverages AI-driven classification to automate vegetation analysis, enabling users to classify forested environments, identify individual trees, extract tree stems, and generate high-value forestry products directly from lidar data. Attendees will learn how AI models can be used to separate vegetation classes, detect and classify tree trunks, perform automated tree counting, and estimate key forest metrics such as tree height, canopy dimensions, and Diameter at Breast Height (DBH). The session will also discuss the data acquisition and processing requirements necessary to achieve reliable results, including point density, trajectory quality, and optimal lidar collection strategies. Real-world project examples will illustrate complete workflows within LP360, from raw point cloud processing through AI classification and forestry analytics, demonstrating how these tools improve productivity, consistency, and scalability for forest inventory and environmental monitoring applications.
11:00 AM – 11:15 AM
A Practical Plane-Fitting Method for Quantitative Evaluation of Lidar Surface Smoothness
Presented by Qassim Abdullah, Woolpert
Surface smoothness or repeatability is an important indicator of lidar data quality because it reflects the combined effects of sensor performance, system calibration, data processing, and measurement noise. Despite its importance, the geospatial industry lacks a standardized methodology for objectively quantifying local surface smoothness. Current practices often rely on the standard deviation of elevations within a neighborhood, a metric that is highly sensitive to terrain slope and curvature and therefore may not accurately represent the intrinsic quality of the lidar measurements.
This paper presents a practical plane-fitting methodology for evaluating local surface smoothness using least-squares estimation. For each analysis neighborhood, a best-fit reference plane is computed and the orthogonal point-to-plane residuals are used to characterize surface smoothness through their mean, standard deviation, and root mean square (RMS) values. Because the residuals are measured normal to the fitted surface, the proposed method largely removes the influence of surface orientation and provides a more physically meaningful measure of local measurement quality.
The methodology is evaluated using synthetic lidar datasets representing flat, tilted, and curved surfaces with identical simulated measurement noise. Results demonstrate that conventional elevation standard deviation varies significantly with terrain slope and curvature, while the proposed orthogonal residual statistics remain stable across planar surfaces regardless of orientation. The metric increases only when the local surface geometry cannot be adequately represented by a planar model, providing a clear distinction between measurement noise and actual surface characteristics.
The proposed approach offers a repeatable and defensible framework for lidar quality assessment that is suitable for production workflows, quality assurance, and specification development. By providing a terrain-independent measure of local surface smoothness, the methodology has the potential to support more consistent evaluation of lidar datasets across diverse landscapes and applications.
