Maize Plant Detection from UAV RGB Orthomosaics — M.Sc. Thesis

Problem

Measuring row spacing and inter-plant distance in maize stands is central to evaluating seed placement precision, emergence uniformity, and ultimately yield potential. Current methods rely on labour-intensive manual field measurements or sensor rigs that cannot scale across large trials. UAV RGB orthomosaics offer centimetre-resolution coverage of entire fields in a single flight, but automated extraction of per-plant metrics depends critically on how individual plants are labelled during training. No systematic comparison exists across annotation paradigms — bounding boxes (fast, rotation-agnostic), oriented bounding boxes (rotation-aware, morphology-preserving), and point annotations (minimal effort, position-only) — for this task, leaving practitioners without a principled basis for choosing their labelling strategy.

Approach

UAV RGB orthomosaics are acquired at early-to-mid vegetative growth stages (V3–V6) when individual maize plants are clearly separable before canopy closure. Three annotation strategies are prepared for the same imagery: (1) Axis-aligned bounding boxes — standard YOLO format, fast to label, rotation-agnostic; (2) Oriented bounding boxes (OBB) — rotated to align with each plant's elongation axis, encoding orientation alongside extent; (3) Point annotations — single centroid click per plant, converted to Gaussian heatmaps for density-map estimation. A YOLOv8/v11 detector is fine-tuned independently under each annotation regime. Post-detection, predicted centroids are reprojected from pixel to UTM coordinates using the orthomosaic GSD and RTK camera position metadata to yield sub-decimetre georeferenced plant positions. Row fitting (RANSAC-based line detection) and nearest-neighbour inter-plant distance computation produce per-plant and per-row agronomic measurements. Annotation effort (time per plot) and model performance (AP, localisation error vs. RTK reference) are evaluated jointly to quantify the accuracy–effort trade-off for each strategy.

Data

UAV RGB orthomosaics acquired with RTK-GNSS-positioned camera over maize field trials at University of Bonn research sites (≈1–2 cm/pixel GSD). Growth stages V3–V6 captured across multiple flight dates to assess detection robustness at different canopy sizes. RTK ground-truth plant positions measured with sub-centimetre accuracy serve as the localisation reference. Three parallel annotation datasets created from the same orthomosaics: bounding box labels, OBB labels, and point click labels — enabling direct per-image comparison of training signal quality.

Validation

Localisation accuracy evaluated as mean distance (cm) between YOLO-predicted centroids and RTK-measured reference positions, decomposed into systematic bias and random error components. Row spacing and inter-plant distance are computed from predicted plant maps and compared against manual tape-measure reference transects. Detection performance assessed with AP@0.5 and AP@0.5:0.95 per annotation strategy. Annotation effort logged per plot to quantify the accuracy-per-minute-of-labelling trade-off. Cross-flight generalisation tested by training on one growth stage and evaluating on another.

Results

In progress. Results pending completion of annotation pipeline and model training.

My Role

M.Sc. Candidate (Sole Researcher). UAV data acquisition planning, annotation pipeline design and execution across three labelling strategies, YOLO fine-tuning and evaluation, georeferencing pipeline implementation, and agronomic metric extraction.

Next Steps

Complete annotation of all flight dates across three strategies. Train and benchmark YOLOv8n/s/m variants per annotation type. Build RTK-reprojection pipeline for centroid georeferencing. Derive row spacing and inter-plant distance metrics and compare against field measurements. Publish dataset and code.

Key Outcomes

• Systematic annotation strategy benchmark (bbox vs. OBB vs. point)
• RTK-accurate georeferenced individual plant maps
• Row spacing + inter-plant distance from UAV RGB only
• Quantified accuracy–annotation-effort trade-off