AI & Computer Vision · Field Sensing to Decision · Precision Agriculture
I develop machine learning and computer vision systems that read crops the way an agronomist would — non-destructively, in the field. My research turns hyperspectral and 3-D imagery into vine- and tree-level decisions on nutrition, yield, and fruit quality, advancing precision agriculture from field to table.
About
I'm a Ph.D. candidate in Biological & Environmental Engineering at Cornell University, advised by Prof. Manoj Karkee. My dissertation builds in-field hyperspectral sensing systems that generate vine-level nutrient maps without cutting a single leaf.
Before Cornell I spent three years as a Graduate Research Assistant at Washington State University and two more as a Farm Operations Engineer in Pakistan. I'm a member of ASABE and PSAE, an FAA Part 107 UAS pilot, and I'm based in Ithaca, NY.
Research
Breaking the work into focused projects — each one instruments something unmeasured, then turns the reading into an action in the field.
Modeling grapevine macro- and micronutrient status at leaf and canopy scales from ground-based, in-field VNIR hyperspectral imaging — pairing feature selection and machine learning with a convolutional-transformer for simultaneous multi-nutrient prediction, contextualized by growth stage.
A motorized hyperspectral rig scans potted waterhemp on a conveyor beneath a Headwall Nano-Hyperspec push-broom sensor (400–1000 nm) — the same camera platform used across the vineyard work, repointed at a weed-management question: can spectral imaging catch herbicide resistance before it becomes visible? Two populations, Hunter (glyphosate-resistant) and Nebraska (glyphosate-susceptible), were sprayed and imaged across five timepoints (0, 3, 6, 10, and 14 days after treatment), comparing each plant back to its own baseline to isolate the spray response from built-in biotype differences. Resistance became spectrally separable at 6 days after treatment, not before, with a 22% injury gap concentrated in the red edge (700–780 nm). Plant-validated LDA and SVM classifiers (split by plant, never by pixel) reached about 73% held-out accuracy, climbing to 81% by day 10, and a stepwise-selected 15-band subset — clustered in the green, red, and red-edge regions — matched full 343-band accuracy, pointing toward a low-cost multispectral field sensor.
Every hyperspectral scan has to become one clean reflectance spectrum per leaf or per canopy before it's useful for nitrogen modeling. This project builds that extraction pipeline end to end. It starts from a shared radiometric calibration: each pixel's raw digital number is corrected against dark- and white-panel scans and rescaled by the panel's known reflectance, turning raw sensor counts into absolute reflectance before any segmentation happens. At leaf level, each tagged leaf is manually outlined and its pixels averaged into one spectral observation. At canopy level, three complementary pipelines pull the same signal at scale: hand-drawn ROI boxes averaged across sampled leaves; an unsupervised route that auto-detects the white reference panel and fuses NDVI, ReNDVI, Excess Green, and Spectral Angle Mapping into a leaf mask; and a supervised route — a YOLO model trained on the cube's RGB channel — that isolates only sun-exposed leaf pixels for a cleaner canopy average. Together they turn a raw pushbroom cube into the leaf- and canopy-scale reflectance behind the nitrogen models.
Grapevine leaf nitrogen varies by cultivar, canopy position, and growth stage, so no single spectral index transfers across a vineyard. This project builds that pipeline end to end. Every scan first passes through a radiometric calibration step — dark- and white-reference correction into absolute reflectance, kept consistent across changing field light — before hierarchical clustering and an eight-method ensemble feature selection (SelectKBest, Lasso, Ridge, Elastic Net, Random Forest, Extra Trees, Gradient Boosting, and a Random Frog search) rank and stabilize the wavelengths most sensitive to nitrogen. Validated across four cultivars (Chardonnay, Pinot Noir, Concord, Syrah), two growth stages, and two seasons in commercial Washington and Oregon vineyards, the pipeline collapses 274 hyperspectral bands to as few as 5–14, reaching R² = 0.82 at the leaf level and R² = 0.72 at the canopy level on raw reflectance. Bands selected at the leaf level transferred to canopy-scale prediction with little loss in accuracy — evidence the selected wavelengths track nitrogen physiology rather than measurement scale.
Apple orchards are fertilized block-by-block, yet every tree has its own nitrogen needs. This project closes that gap end to end. On the sensing side, a ground vehicle's RGB-D stereovision scans each tree and a machine-vision pipeline segments the canopy point cloud into green, yellow, and trunk points to compute a yellowness index of fall color — a gradient-boosting model estimates it at R² = 0.72 and identifies roughly the 29th week after bloom as the best window to read a tree's nitrogen status from canopy color. On the action side, an autonomous orchard robot navigates the rows, localizes itself to within half a tree spacing, keeps a per-tree history, and — through an onboard decision-support system built on ROS — delivers a variable, tree-specific rate of liquid nitrogen in real time, validated in both lab and field trials. Together they turn canopy color into a precise, individual-tree fertilization decision.
A low-cost yield-estimation pipeline for wine grapes that runs entirely from a smartphone a grower already carries. YOLO11 detects and counts grape clusters and shoots from phone RGB imagery, depth maps give per-cluster surface area, and the phone's LiDAR reconstructs a 3-D point cloud of the canopy to compute canopy volume. Those features — cluster count, canopy volume, neighbouring-vine volume, adjusted surface area, and shoot count — feed a multi-feature linear regression that estimates vine-level yield (R² ≈ 0.52, RMSE ≈ 1.35 kg/vine); aggregated to the block, the prediction landed within ~2–3% of the hand-harvested total, with accuracy improving as more vines were sampled. I contributed to the machine-vision system and data analysis.
Read Chapter 4 (dissertation) →
Behind every model is a season in the field. I orchestrated multi-site campaigns across commercial vineyards in Washington, Oregon, and New York — hyperspectral and RGB-D imaging from dormancy through harvest, leaf-tissue sampling and lab processing, and grapevine pruning and hand-harvest.
Publications
Experience & Education
Developing advanced computer vision, machine learning, and hyperspectral remote sensing systems for precision agriculture — engineering decision-support tools and automated workflows to optimize grapevine nutrient modeling, robotic pruning, and weed management.
Led computer vision, machine learning, and sensor-fusion research for precision viticulture and orchard management, engineering automated data pipelines and remote-sensing frameworks to optimize crop nutrition and yield estimation under USDA-NIFA funding.
Managed agricultural engineering operations and automated climate-control systems across multiple high-density production facilities.
Supervised feasibility, hydraulic design, and execution of high-efficiency drip and sprinkler irrigation networks, modernizing water-use efficiency across regional farming command areas.
Managed a high-value agricultural credit and microfinance portfolio across rural branches, driving lending growth while maintaining strict asset quality and regulatory compliance.
Thesis: wastewater characterization and GIS-based subsurface contamination mapping.
Awards & Honors
Connect
I'm open to research collaborations, industry partnerships, and new projects in sensing and precision agriculture. Feel free to reach out — email works best.