Jake Lee

@jakehlee.com

303 Followers 130 Following

Data Scientist @ NASA/Caltech Jet Propulsion Laboratory, Machine Learning for Earth and Space Science. EMIT Imaging Spectroscopy, Euclid/Roman Cosmology. Volunteer robotics head coach FRC 2404.

Top Left: YouTube thumbnail for the video "The Race to Win Staten Island", but the colors are inverted: the land is blue and the ocean is orange. Top Right: "Fable of the Dragon-Tyrant", a square crop in the center is mirrored horizontally. Bottom Left: "Is Pluto a planet?", the entire thumbnail has gaussian noise added, resulting in a grainy image. Bottom Right: "Death to Pennies", the thumbnail has noise added, as well as a small square flipped horizontally on the left.
Caption: "DDPM-generated Euclid-like galaxies" Content: a grid of black-and-white images of galaxies generated by a diffusion model.
About 30 students and mentors of First Robotics Competition Team 2404 standing around their competition robot, about 2ft across and 3ft tall.
Screenshot of a web portal visualizing airborne instrument imagery of the Eaton Fire. JPL located on the upper left. Burned areas cover the right half, and is outlined in red.
Three visualizations of an AVIRIS-NG scene. a) An RGB scene with clouds in the left and middle. b) A falsecolor scene that highlights the presence of clouds. c) A cloud mask overlaid in red over the RGB. It masks most, if not all, clouds present in the scene.
A plot with "wavelength (nm)" on the x-axis and "Attention Score" on the y-axis. A higher attention score indicates that the model thought the wavelength was more important for the cloud detection task. Some wavelengths are annotated: O2-A, H2O, and CO2. These are known absorption spectrums that SpecTf has learned from the data alone.
2 rows and 4 columns of pictures of satellite imagery. The second row is a zoom-in of the first. Left: an RGB image of a scene covered in clouds. Center-Left: a falsecolor image of certain wavelengths that highlight the presence of clouds, even those that are partially transparent. Center-Right: the current EMIT cloud mask overlaid on the RGB image. A significant number of clouds are not detected. Right: a cloud mask produced by SpecTf. Almost all clouds are detected.
a) A global map indicating the distribution of sampled scenes for the dataset used in this work. b) Visualization of the spectra in the "clear" and "cloud" classes. c) Visualization of the data cube of a single EMIT scene. d) A diagram of the Spectroscopic Transformer model architecture, going from the spectrum input to the "clear" and "cloud" class probabilities.
Two overlaid graphs showing the per-iteration loss while training a deep learning model. A previous model in red stays at the same loss over 60,000 iterations. A newer model in green significantly lowers in loss after 10,000 iterations, indicating better convergence.
A screenshot of the Google Scholar PDF Reader with an AI-generated paper outline on the left.