Project
Earth-Science Data Tooling — Packaged Scientific Python
I developed scientific-data tools through NASA internships and UMBC research, including conversion to netCDF4 and parsing into pandas DataFrames. cosmic-crunch and noaa-gml-file-reader are published on PyPI; I also contributed to collaborative cloud-anomaly analysis.
Atmospheric research produces two kinds of code: the notebook you run once to make a figure, and the tooling everything else depends on. This is the second kind — the packaged layer under my NASA and UMBC remote-sensing work that got raw satellite and observatory data into a shape the science could use.
Context
Across two NASA Goddard internships and research at UMBC, the bottleneck was never the analysis — it was the data plumbing. Satellite and ground-station data arrive in awkward formats (HDF4, fixed-width ASCII, vendor text dumps) at sizes that punish naive code. A cluster of focused tools grew up to handle ingestion, conversion, regridding, and visualization.
The problem
Scientific data formats are their own discipline: MODIS products as HDF4, reanalysis and soundings as netCDF or raw ASCII, observatory records as idiosyncratic fixed-width text. Before any analysis, all of it has to be parsed, put on common grids, and stored to read back fast — and at NASA data volumes, “read it all into memory and loop” isn’t an option.
Approach
The work is split across small packages, each doing one job:
- cosmic-crunch
(
pip install cosmic-crunch) — crawls the JPL GENESIS COSMIC archive and converts ASCII soundings to netCDF4, behind a singlecosmic-crunchCLI (get/convert) with multiprocessing. - NOAA observatory readers —
noaa-gml-file-reader(pip install noaa-gml-file-reader) parses NOAA GML station data intopandasDataFrames. - MODIS / CALIPSO anomaly analysis — the analysis layer for a cloud-opacity discrepancy between the MODIS and CALIOP instruments.
- Dark Target ingestion + regridding (private) — a toolkit around NASA’s
Dark Target aerosol retrieval: AERONET/VIIRS ingestion (ASCII → netCDF4),
congrid-style resampling, and optical-depth processing onpyhdf/HDF4. The substance is the ingestion and regridding, not a from-scratch reimplementation of the retrieval. - netCDF/HDF-EOS2 conversion — a small recompression tool from the same era, since rebuilt into ncarnate, a published converter that reads HDF-EOS2 granules and reconstructs their geolocation as CF-standard coordinates.
Common thread: pyhdf (HDF4) and netCDF4 for the formats, NumPy/SciPy/pandas
for the work, cartopy/Matplotlib for the maps, and real packaging so the next
person could install and run.
At HPC scale
The same MODIS data also drove the distributed-computing side of my UMBC CyberTraining work. I aggregated MODIS products with Apache Spark on a SLURM-scheduled HPC cluster and ran it as a scaling study — the same job across a growing node count, serial versus parallel — to find where the parallelism actually paid off. The machine-learning coursework alongside it covered distributed supervised learning (Random Forest, logistic regression, and SVM with feature pipelines and cross-validated tuning on Spark MLlib). In the 2020 CyberTraining coursework, I also used TensorFlow/Keras to build, train, and evaluate a Sequential network on MNIST.
Outcome
The two readers/crawlers are published on PyPI under the MIT license —
modernized in 2026 with dual-format parsing, a packaged CLI, tests, and CI — so
each is one pip install cosmic-crunch or pip install noaa-gml-file-reader
away from running. The MODIS/CALIOP analysis stays on GitHub under AGPL-3.0, and
some research-specific tooling stays private.
Collaborative research and sources
The MODIS/CALIOP anomaly-analysis work was collaborative: Christine Abraham, Olivia Norman, and I authored the 2020 technical report listed by UMBC, with Jianyu Zheng as research assistant and Zhibo Zhang as faculty mentor. The March 2020 presentation credits all five contributors. These are distinct research outputs.