Automating River Channel Mapping: Lidar and AI Revolution

By Aryaman PattnayakReviewed by Susha Cheriyedath, M.Sc.Jan 8 2024

An article recently published in the journal Water Resources Research outlines new methodologies to accurately delineate the bank full extent of river channels using high-resolution lidar data and artificial intelligence techniques. Defining the bankfull level, the stage at which rivers start to spill onto their floodplains, provides a consistent benchmark for analyzing channel geometry and evolution critical to river management. However, conclusively identifying this morphologic boundary can prove challenging, especially in densely vegetated environments with obscured aerial imagery.

Manually surveying rivers through field mapping or digitization from aerial photography can be highly time-consuming and challenging where overhanging vegetation conceals the channel banks. Computational approaches to delineate bankfull hold the promise of efficiency, consistency, and objectivity. However, existing techniques still need to be revised regarding accuracy and applicability to complex multithread rivers.

The researchers aimed to develop automated approaches purpose-built to leverage high-resolution topographic data from lidar surveys. The goal was to delineate the bank's full extent through computer vision and spatial analysis, even where traditional optical remote sensing fails due to occlusion from vegetation. The ability to rapidly characterize stream dimensions over large areas aids various modeling and analysis needs to support flood risk mapping, stream rehabilitation initiatives, and tracking channel evolution.

The Study

The study area comprised 13 km of Cardinia Creek and 3 km of its tributary, Officer Road Drain, in Victoria, Australia. Cardinia Creek is a lowland alluvial system flowing through agricultural and peri-urban areas. The five study reaches ranged from 2.2 to 4.5 km in length with variable surrounding land use, riparian quality, channel planform, and evidence of historical modification through straightening or incision. Bankfull widths derived from manual digitization of lidar data spanned from 7.5 m for the engineered Officer Road Drain tributary up to 26 m for a wandering, recovering reach of Cardinia Creek.

The researchers tested two computational methods capable of ingesting lidar digital elevation model grids and derivatives:

• An enhanced one-dimensional cross-sectional approach called HydXS that delineates bankfull elevation by maximizing the hydraulic depth, calculated as the ratio of wetted cross-sectional area to surface width
• A convolutional neural network performing semantic image segmentation to classify river channel pixels in derived lidar raster "imagery."

The heterogeneous study with diverse fluvial geomorphology offered an ideal testing ground to understand the relative strengths of each technique.

Methods Used

The HydXS method detects bankfull elevation along cross-section profiles by finding the elevation that produces the maximum wetted area per unit width, known as hydraulic depth. This concept builds on prior formulations for automated bankfull delineation but substantially improves results by allowing multiple in-channel depressions per cross-section instead of just one main channel. The algorithm also implements several checks for consistency and overrides potentially spurious elevations reaching the data boundaries.

Meanwhile, the deep learning approach utilizes transfer learning to retrain a convolutional neural network (CNN) called ResNet-18 on augmented lidar-derived imagery. Transfer learning sidesteps full CNN training by retaining preset hierarchies of feature detection layers, thereby enabling accurate image classification with small training datasets. Fourteen combinations of elevation, slope, and curvature raster layers were constructed from the lidar digital elevation model and supplied as pseudo-red, green, blue (RGB) inputs to assess the optimal data formulation for differentiating bankfull extent.

The HydXS algorithm and CNN used the same set of manually digitized bankfull channel centerlines and boundaries derived from the lidar data to serve as training and test data. The researchers split the study reaches such that 90% of each reach lent examples to train the models, reserving 10% for final quantitative testing.

Key Findings

• The HydXS method demonstrated higher overall accuracy with a Dice coefficient of 0.83 compared to 0.78 achieved by the best-performing CNN input. The Dice metric balances precision and recall to measure the overlap between predictions and ground truth.

• Supplying locally normalized elevation as input rasters significantly enhanced CNN performance using regionally normalized lidar, indicating that accentuating subtle within-image topographic variability aids differentiation of the bankfull edge.

• The CNN approach proved advantageous for reaches exhibiting inset floodplain features, where it outperformed the cross-section-based HydXS method, which tended to identify lower elevation benches and terraces as backfill instead of the actual channel margin.

• Both computational methods showed markedly improved accuracy over existing algorithms leveraging automated cross-sectional analyses or pixel-wise classifications to extract bankfull extent from high-resolution topographic data.

Significance and Applications

The study demonstrates the powerful capability of AI-enabled techniques to automate the intricate task of mapping diverse fluvial landforms at large scales using lidar surveys. The data extraction methods grant an efficient alternative to time-consuming and potentially hazardous field mapping where overhead vegetation precludes accurate digitization from optical imagery.

The convolutional neural network requires only seamless lidar-derivative raster mosaics spanning the study region, enabling rapid application across expansive areas. Meanwhile, the objective HydXS approach yields precise bankfull elevations directly from river cross-sections at regular intervals along the channel centerline. The choice of appropriate methodology depends on specific project aims and stream types.

The algorithms facilitate large-scale topographic and hydraulic analyses by extracting key parameters, including water surface elevation, channel width, depth, and cross-sectional area over kilometers of stream length. Potential use cases include:

• Flood inundation modeling.
• Stream rehabilitation design for incised or unstable channels.
• Gauging sediment transport capacities.
• Tracking geomorphic changes in response to shifting land use practices or climate shifts.

Regional bankfull delineation supports river managers in evaluating current channel capacities relative to historical baselines.

Future Directions

While the techniques proved robust across diverse fluvial environments within the study domain, further testing across geographic contexts is warranted to characterize transferability and limits in performance fully. The researchers suggest investigating the dynamic scaling of CNN input imagery proportional to stream width and depth to improve generalizability across stream orders. Centering images on an automated centerline may also boost accuracy.

This research puts forth trailblazing progress in mapping the bankfull boundaries of river channels under dense riparian canopies. The computational methods showcase the immense potential of lidar-derived terrain data combined with artificial intelligence for automated feature extraction. The study adds to a growing movement applying data science techniques to unlock previously unfeasible environmental mapping workflows. The bankfull channel delineations support a multitude of predictive, analytical, and planning applications to evaluate and sustain fluvial system function.