Farm Monitoring Drone - Pranav Bonthu

Project Overview

I'm leading a student team building an all-purpose agricultural drone. It flies over a farm, collects camera and LiDAR data, and turns it into a field heatmap the farmer can actually use: where pests are showing up, how crops are growing, and how much fruit is on the trees.

The idea grew out of my pest-detection work at Owl Vision and a list of problems a farmer I work with wanted solved. We've partnered with a farm and are designing the drone alongside its farmers so it fits how they really work.

Long term, flying the same drone across many farms and seasons builds a crop-health dataset that no single farm could collect on its own.

System Architecture

A Raspberry Pi 4 rides onboard as the companion computer. It runs detection on a Pi camera, with an OpenMV RT1062 as a second onboard vision module, and a BerryGPS-IMU tags every frame so results land in the right spot on the map. Multispectral imaging and LiDAR for canopy and fruit density are the next sensors on the roadmap.

System block diagram

First Model: Counting Apples

The first detection task is counting apples. I built it as a tabletop demo on the Raspberry Pi 4 and camera, with a live dashboard that shows apples in view, a running count of unique apples (each one gets a tracking ID so it is only counted once), the active model, frame rate, and CPU temperature.

The stock COCO-trained YOLO11n found apples, but with confidence only around 0.5 to 0.7, and it happily labeled a tennis ball plush as an apple. To fix this I fine-tuned YOLO11n on public apple datasets in Google Colab and exported it to NCNN so it runs on the Pi at about 2 FPS.

YOLO11n

Fine-tuned on public apple datasets

0.73 to 0.92

Detection confidence after fine-tuning

NCNN

Exported for on-device inference

Before: baseline COCO model

Baseline YOLO11n detecting apples at 0.52 to 0.70 confidence and mislabeling a plush toy — Cluttered scene: 0.52 to 0.70 confidence, and the tennis ball plush counted as an apple

Baseline YOLO11n detecting four apples on a table — Top-down view with the baseline model

After: fine-tuned on apple data

Fine-tuned model detecting a red apple at 0.91 and ignoring a plush toy — 0.91 on the red apple, and the plush is no longer flagged

Fine-tuned model detecting two apples at 0.92 and 0.88 next to a dark toy — 0.92 and 0.88, with the dark toy correctly ignored

Fine-tuned model in a cluttered scene detecting apples at 0.73 to 0.87 — Same cluttered scene: real apples rise to 0.73 to 0.87, though the tennis ball plush still fools it

Grid of validation images with apple predictions from training — Validation predictions from the Colab training run

Flight Hardware

The flight side is kept separate from the sensing computer. A 6S LiPo feeds an inline smoke stopper, then the Aero Selfy F405 flight controller and 4-in-1 ESC stack driving four 2450 KV motors over DShot. A 5 V BEC powers the Pi, which talks to the flight controller over UART, and a CRSF radio receiver handles manual control. The stack is soldered, wired, and connected to the configurator; next up is choosing flight firmware, with ArduPilot as the front runner for GPS waypoint missions.

Drone electrical diagram showing battery, smoke stopper, BEC, ESC, flight controller, radio receiver, Raspberry Pi 4, Pi camera, BerryGPS-IMU, and OpenMV RT1062 — Power and data wiring for the drone

Next Steps

Move apple counting from the tabletop to real trees, recording fast footage and processing it afterward
Integrate multispectral imaging and LiDAR on the Pi
Pick flight firmware and airframe, then fly first survey missions over the partner farm
Generate the first field heatmap for the farmers to review