Stephen Huysman

I developed and optimized a computationally intensive, high-resolution (1-meter) spatial hydrology model. This project involved translating complex mathematical equations into high-performance code capable of calculating flux variables, such as Actual Evapotranspiration (AET) and Climatic Water Deficit (CWD), which represent daily water and energy fluxes.

Because running simulations at a 1-meter spatial resolution requires processing immense volumes of high-dimensional gridded data, the primary engineering challenge was strict memory management and algorithmic efficiency. To achieve this, I navigated the full spatial data lifecycle, utilizing my field experience with drone-based LiDAR collection to process downstream LiDAR-derived DEMs. I also developed programmatic workflows to extract and refine complex vector objects (points, lines, and polygons) for hydrological topology and ecological boundaries.

Architected and deployed a fully automated, cloud-native data pipeline on AWS to process high-dimensional spatial and environmental data into daily predictive wildfire forecasts. The system ingests and transforms gridded historical climate data and ensemble weather forecasts, applying custom rolling-window calculations and empirical statistical mapping (eCDF) against a 40-year baseline.

This system relies on rigorous data engineering and a serverless "compute-on-demand" architecture. I designed the orchestration using Docker, AWS Fargate, and AWS Step Functions to ensure highly parallelized, reliable daily execution. The pipeline handles the entire flow from raw data ingestion to the automated generation of static HTML dashboards, interactive maps, and GIS-ready NetCDF and Cloud-Optimized GeoTIFF (COG) outputs.

Engineered a recurrent neural network (RNN) using PyTorch and long short-term memory (LSTM) architecture to forecast complex streamflow dynamics and flood risks. This project required processing high-dimensional temporal datasets to predict hydrological events with high precision.

Key technical challenges included handling variable-length time sequences, engineering features from raw environmental data, and optimizing the model to prevent overfitting on historical baselines. By leveraging LSTM architecture, the model successfully captures long-term dependencies in the data that traditional statistical models often miss. The resulting system provides critical, data-driven insights for flood warning and risk management, demonstrating the practical application of deep learning to real-world safety challenges.