Boaz Berhanu Tulu

Research

My PhD is about precision agriculture at more than one scale: using generative AI to get more out of drone and satellite imagery, from a single plot up to a whole region. I work with Dr. Haimanote K. Bayabil and Dr. Yiannis Ampatzidis, and the projects below were built in Dr. Bayabil's Water Resources Research Lab. Most of them have a demo you can open in a browser.

The dissertation also takes the super-resolution idea up a scale: co-located drone flights teach GAN and diffusion models to add realistic detail to satellite images of the same fields, so that decisions made for one field can be made across a region. I presented that work at the AI in Agriculture Conference at NC State in 2026.

On this page: PhD projects · Software · Earlier work · Talks

Projects from my PhD

PID-LDM: thermal images from ordinary drone photos

An RGB drone photo of crop rows next to the canopy temperature map predicted from it, on a scale from 0 to 50 °C
An RGB drone photo (left) and the canopy temperature that the online demo model predicts from it (right).

The Crop Water Stress Index (CWSI) tells a grower when plants are short of water, but it needs canopy temperature, and thermal cameras are expensive and need careful calibration. RGB cameras, on the other hand, ride on almost every drone. PID-LDM is a physics-informed latent diffusion model that predicts canopy thermal images straight from RGB photos and then turns the predicted temperature into a CWSI map.

Try the demo (it runs a simpler Pix2Pix model on sample plots) · Related paper in Computers and Electronics in Agriculture (2026)

GeoDiff-SR: sharper drone images with diffusion

A blocky low-resolution plot image in a white frame, laid over sharp drone imagery of crop rows
A low-resolution input, zoomed four times, over real drone frames from UF/IFAS TREC.

GeoDiff-SR is a conditional denoising diffusion model that raises the resolution of agricultural UAV images four times while staying faithful to what is actually in the field. It recovers fine canopy detail without the invented texture that GAN super-resolvers tend to add, so measurements taken from the sharper images stay reliable.

It was trained on 2,136 sweet corn and green bean images collected over three growing seasons (2020 to 2023) at the UF/IFAS Tropical Research and Education Center.

Try it on sample plot images

Sabal Field Advisor: irrigation advice with a receipt

Sabal app screen: the decision is Skip, with the next irrigation date, a root zone 86 % full and a forecast of soil water depletion
The irrigation decision screen.

Sabal brings the public data for a Florida field into one place: FAWN and National Weather Service weather and forecasts, USDA soil surveys, USGS groundwater wells, drought and flood maps, and state water rules. A daily FAO-56 root-zone water balance decides whether to irrigate or skip, and when the next irrigation is due.

Every recommendation comes with a receipt listing its sources, data dates and steps, so it can be checked. Growers can draw their fields, add their own records and ask questions in plain language.

Still in testing: the web app asks for an access code, and Android and iPhone versions are being tested too.

Open the test site

Shamba: rainy-season advice for East African farmers

Shamba app screen: a map around Lake Tana and season advice for Bahir Dar, Ethiopia
Season advice for Bahir Dar, Ethiopia.

Shamba (Kiswahili for farm) turns public satellite rainfall and climate data into season information for any point in Ethiopia, Kenya and Tanzania: when the rains usually start, whether they have started yet, false starts, dry spells, and which crops the season can carry. You pick a place on the map or search for a town or village.

The answers are short and come in English, Amharic, Afaan Oromoo or Kiswahili, and every number shows its source and data date. Rainfall comes from the CHIRPS and TAMSAT satellite records, and the crop water outlook covers maize, sorghum, wheat and other crops.

Still in testing: the web app asks for an access code, and Android and iPhone versions are being tested too.

Open the test site

SpectraSelect: spectral filter design with language-model agents

Chart of ten bell-shaped filter curves between 400 and 1000 nm, drawn over the reflectance spectra of crop and weed species
The ten filters NSGA-II chose for one data partition, drawn over the reflectance of the crop and weed species.

Hyperspectral cameras separate crops from weeds well, but they are too expensive to put on a sprayer. A compact sensor can measure a few bands through optical filters instead, and then the question is which centres and widths to pick. I searched filter designs with NSGA-II on WeedCube, a public set of 160 greenhouse images with 224 bands covering three crops and four weeds, and tested whether a team of language-model agents (built on the raia platform) could steer that search.

Interactive results page, including a filter designer that retrains the classifier in your browser

Software

AgriSenAI

A desktop platform that takes raw UAV thermal and multispectral flights to plot-level numbers: stitching, calibration, plot extraction and metrics, now with diffusion super-resolution and RGB-to-thermal translation built in. It started as a tool for my own experiments. The University of Florida has exerted its rights to the software and its source code (© 2026 University of Florida Research Foundation, Inc.) for licensing and commercialization.

Paper in SoftwareX (2025)

IrrigSense

A decision support tool that streamlines the processing of soil-moisture sensor data for irrigation scheduling. I presented it in 2024 at the 9th UF Water Institute Symposium and at the North Florida Research and Education Center's spring showcase.

Strawberry pruning robot

I was on the University of Florida team that entered a strawberry pruning robot in the ASABE 2024 Robotics Competition in Anaheim, California.

Earlier work, in Ethiopia

Debo Buna

Launch poster for the Debo Buna app: a hand holding a phone that scans a coffee leaf
Launch poster for the app.

A phone app from Debo Engineering that detects coffee leaf diseases, made for Ethiopian smallholder farmers. Debo's coffee disease tools (web, mobile and desktop) now serve over 4 million farmers, and the work won the GIZ Coffee Innovation Award in 2022.

Debo Buna on Google Play

Soil classifier and crop recommender

A field device from Omishtu Joy that measures soil properties and uses AI to suggest suitable crops and management for a plot. I led both the model and the hardware, and it won the Sankalp World Bank innovation prize in Nairobi in 2022.

Omishtu Joy · News coverage

Coffee grading device

Automated grading of coffee beans, to help producers judge quality before they sell.

Watch the demo

Gebeta-playing robot

A robot that plays Gebeta, the traditional Ethiopian board game, combining perception, actuation and game strategy.

Watch it play

Amharic script and Ethiopian dance

Two of my 2026 papers are about Ethiopian writing and culture: AmharicScrib-Diff, a class-conditional diffusion model that writes handwritten Amharic Fidel characters, and EthioDance-ViT, a video vision transformer that recognizes Ethiopian traditional dances.

Talks