Conducted a 2-hour live training with 140 registrants on using AI segmentation to delineate GIS features for hydraulic modeling, where attendees followed along on Python notebooks in the cloud. Three separate methods were taught for performing this delineation, depending on the use case.
This was prefaced by a live demo of these technologies with over 1000 people in attendance, which can be viewed in the above video.
Building Footprint Detection and Segmentation System
Deep Learning Model and Production Framework
Trained a custom fine-tuned deep learning building footprint detection model from satellite imagery, and built out a framework for it which allows the user to automatically detect and segment building footprints using AI remote sensing for an arbitrarily large area, and save the results back to a GIS shapefile. To accomplish this scalability, two separate spatial chunking, or "tiling" schemes were implemented, based on the respective optimal image resolution that the detection and segmentation models were each trained on. The tool is freely available, linked in the above image.
Synthetic Internal Boundary Conditions through NTR Timeseries Prediction
LSTM Regression Deep Learning Model
After developing synthetic gridded rainfall patterns and downstream surge timeseries' for the LWI Lower Calcasieu HUC 8 model's design storms, methodology was required to produce realistic inland surge boundary forcings, as the model was quite sensitive to these forcing inputs. To bridge this gap, I designed a minimal LSTM regression model to test if inland Non-Tidal Residual (NTR) forcings could be deterministically predicted from precipitation and coastal NTR, targeting ≤ 1 ft MAE/RMSE at model validation stations.
• Compiled & preprocessed 40 years of USGS/NOAA gauge records:
• Merged and bias-corrected adjoining station time series
• Detrended tidal constituents to yield non-tidal residuals (NTR)
• Validated synthetic NTR outputs on 3 excluded storms (Delta 2020, June 2018, May 2021), meeting performance criteria and confirming deterministic BC behavior
• Architected a 2D CNN model (Convolutional Neural Network) to ingest cumulative 72-hour gridded rainfall fields and distill 64 predictive features per hourly timestep
• Employed 5th-order polynomial sample weighting, L2 regularization, and dropout to "learn" and accurately predict extreme hydrodynamic phenomena (e.g., negative surge in Sabine Lake)
• Integrated CNN feature extractor with a 24-hour sequence LSTM, appending downstream surge to precipitation features
• Achieved validation MAE ≤ 0.279 ft across validation events, demonstrating reliable synthetic inland surge time series for boundary-condition generation