Self-Supervised Learning for Hyperspectral Imagery
Doctoral thesis · Machine Learning and Sensing Lab
Hyperspectral images carry hundreds of spectral bands per pixel, and labels are scarce and expensive. Masked autoencoders are the obvious way to learn from the unlabeled majority, but they were designed for natural images, where masking a patch removes information that genuinely has to be inferred.
Spectral bands do not behave that way. Adjacent bands are highly correlated, so masking one is trivially recoverable from its neighbors. Isolated absorption features are the opposite: nothing else predicts them, so masking one makes reconstruction impossible. Band-level masking therefore produces a training signal of wildly uneven difficulty, most of which teaches the model nothing.
My thesis, Beyond Band Masking: Decomposition-Guided Self-Supervised Learning for Hyperspectral Images, argues that the lever is not a cleverer masking rule but the representation space you mask in. Decomposing the spectrum first, then masking whole components rather than individual bands, turns an uneven reconstruction task into a uniform one that exploits the redundancy structure instead of fighting it.
Benchmarking and Routing Large Language Models
Applied Scientist Internship · Amazon, Seattle · Summer 2026
A deployment rarely has a single right model. It holds many, spanning an order of magnitude in price and several-fold in speed, and the model that answers a hard query best is seldom the one that answers an easy query most cheaply. A router turns that spread into an advantage by sending each query to a model chosen for it.
Evaluating routers offline needs an answer key: a record, for every query and model, of what came out and whether it was right. Building one is expensive, so released keys label whichever axes their authors needed, and the axis consistently missing is time. Without it you cannot ask whether a cost-optimal route is also latency-optimal, which is the question an operator actually has.
I built an answer key that carries correctness, dollar cost, and served latency in the same cell, over a pool spanning both open-weight and closed commercial models, and a router that trades the three off at a ratio the operator sets. Measured together, the two resource axes turn out to disagree often enough that treating them as one number leaves real savings unclaimed.
Hyperspectral Imaging for Plant-Soil Systems
Machine Learning and Sensing Lab
Traditional soil moisture sensors give point measurements, but the gradients that drive rhizosphere processes happen at the millimeter scale. I developed the first demonstrated approach for pixel-level soil water content mapping in rhizosphere systems, training a DenseNet-inspired CNN on controlled hyperspectral images of soil with measured water contents, then transferring it to rhizoboxes containing growing peanut and sweet corn.
Two problems had to be solved for that transfer. A systematic intensity difference between setups, which I addressed with a spectral scaling correction, and interference from root spectral signatures, which turns out to be species-dependent: sweet corn's thick roots produce a 73% prediction range reduction when filtered, against 34% for peanut's thinner architecture.
The model reaches R²=0.89 and MAE=1.58% at quadrant level across 28 rhizoboxes, mapping water content at roughly 0.5mm effective pixel spacing, a 26 to 45% improvement over traditional regression methods. The work builds on the HyperPRI dataset, which I led the creation of.