One pixel changes meaning
GSD controls how much physical ground each pixel represents, changing visible canopy detail.
AIT Master's thesis · Geospatial AI
Teaching a drone to see.
A six-model instance-segmentation benchmark for detecting and measuring oil-palm canopies across eight UAV image resolutions.
6Models compared
8GSD levels tested
0.77Best mean IoU
>80%Annotation time saved
The operational problem
Altitude changes the image—and the model.
A detector trained at one drone height can fail over the same plantation at another. Ground Sample Distance is the hidden variable connecting flight planning to model reliability.
Accuracy shifts with altitude
A model trained at 0.03m GSD performed 40% worse at 0.20m on the same plantation.
Flight plans become AI inputs
Deployment requires choosing altitude, model, speed, and canopy precision as one system.
Six models enter
The simplest pipeline won.
YOLOv11 standalone delivered the best mean IoU while running roughly ten times faster than the YOLO+SAM hybrid.
ModelIoUF1ms/img
YOLOv11l-seg0.770.69856
Hybrid-v11-SAM0.740.681~600
Mask R-CNN0.7120.727~340
YOLOv8l-seg0.710.67268
Hybrid-v8-SAM0.690.651~580
Agent-orchestrated pipeline
From gold labels to canopy area.
The best trained model becomes a Teacher Agent, scaling annotation before the full benchmark and physical measurement stages.
Gold standard
500 manually annotated UAV images establish trusted labels.
Teacher Agent
YOLOv11 propagates annotations across tiled GSD variants.
Model suite
Six architectures train and evaluate under the same protocol.
Biometry
Polygon masks become canopy area and equivalent diameter.
.03m.05m.07m.10m.12m.15m.17m.20m
Reproduce the work
Open research, end to end.
The repository contains model code, evaluation results, notebooks, and the full multi-GSD pipeline.
git clone https://github.com/Sai21112000/oil-palm-instance-segmentation.git
cd oil-palm-instance-segmentation
pip install -r requirements.txt
# Explore notebooks, models, src, and resultsResearch made operational
Choose the flight and the model together.
Explore the source, results, and complete six-part technical story.