← Projects Research · Independent · 2025

Crop Disease Detection Drone

Independent research  ·  Trained on UT Dallas HPC

A custom-built drone under $500 that flies autonomous routes over crops, photographs the leaves and runs a CNN that identifies 10 diseases across corn, potato, rice and wheat.

Published · IEEE IPCCC 2025 Affordable Crop Disease Detection with Drone Imagery and Artificial Intelligence Vedant Mohapatra, Sasmita Mohapatra 2025 IEEE International Performance, Computing, and Communications Conference (IPCCC)
1 citation 149 full-text views
Read on IEEE Xplore → DOI: 10.1109/IPCCC66453.2025.11304633
Assembled drone with labeled motors, flight controller, camera, video antennae, carbon fiber chassis and 3000 mAh battery
The assembled drone
<$500total build cost
27 minmax flight time
10diseases, 4 crops
5.9×faster training on H100

Why it matters

  • Pests and diseases cost the global economy about $220 billion in crops every year.
  • Current detection relies on lab soil tests or sensor arrays, which are slow, expensive and hard to maintain.
  • Goal: a cheap, repeatable flyover that flags diseased plants early, so farmers can treat only what's infected and use less pesticide.

The drone

Frame
Carbon fiber chassis with 3D-printed TPU parts
Propulsion
Four 1500 kV brushless motors, 60 A ESC, 3000 mAh battery
Electronics
Flight controller, GPS, DJI O3 video transmission
Onboard compute
Raspberry Pi 5 for real-time inference
Autonomy
Point-to-point waypoint routes planned in Litchi
Field tests
Clear close-up leaf images even in light rain and clouds; 27 min max flight

Dataset and model

Sample leaf images: corn and rice, healthy and diseased
Corn and rice classes
Sample leaf images: potato and wheat, healthy and diseased
Potato and wheat classes
  • 13,024 labeled leaf images (public Kaggle data plus drone captures), split 80/20: 10,426 train, 2,598 test.
  • 14 classes: a healthy class plus 2–3 diseases for each crop.
  • PyTorch CNN: three convolution + ReLU + max-pooling blocks on 128×128 images, a fully connected layer and softmax. About 5.6M parameters.

CPU vs. GPU training

Trained on UT Dallas's Ganymede2 cluster, once on an AMD EPYC 7453 CPU and once on an Nvidia H100 GPU. The GPU finished 488% faster (2,405 s vs 14,153 s) while running 67% more epochs.

EpochsTraining timeTrain acc.Test acc.
AMD EPYC 7453 (CPU)1514,153 s85%76%
Nvidia H100 (GPU)252,405 s89%71%
CPU model loss vs epochs
CPU loss
CPU model accuracy vs epochs
CPU accuracy
GPU model loss vs epochs
GPU loss
GPU model accuracy vs epochs
GPU accuracy

The longer GPU run started to overfit (higher train, lower test accuracy). Its speed leaves room for more data and a deeper model.

Results by disease

75–95% accuracy for most classes. Rice was the exception: its thin leaves give fewer visible features and are easily hidden in the frame. Neck blast, which leaves clearer marks, was still detected well.

CropClassCPU modelGPU model
CornHealthy
100.0%
100.0%
Common Rust
97.5%
100.0%
Gray Leaf Spot
88.1%
64.4%
Northern Leaf Blight
81.7%
95.9%
PotatoHealthy
96.7%
96.7%
Early Blight
99.5%
93.0%
Late Blight
98.5%
95.5%
RiceHealthy
11.1%
6.7%
Brown Spot
7.4%
22.9%
Leaf Blast
50.3%
19.5%
Neck Blast
83.5%
50.5%
WheatHealthy
93.3%
98.2%
Brown Rust
90.0%
81.7%
Yellow Rust
74.5%
93.5%
CPU model (15 epochs)GPU model (25 epochs)

Detection pipeline

  1. Fly
    • Autonomous waypoint route over the field, recording video
  2. Train
    • CNN trained on HPC, then exported for offline use
  3. Detect
    • Inference on the Raspberry Pi 5 or in the cloud (AWS) flags healthy vs. diseased plants

Acknowledgments

Thanks to the UT Dallas High Performance Computing Center and its researchers for feedback and for compute time on Ganymede2.