Ahmad Yateem
← Project index3.5 / Machine learning research

Biodiversity Computer Vision Research

State-of-the-art plant-species identification, proven in the lab and monitored in an admin dashboard.

Machine Learning Researcher
University biodiversity research group across two phases with separate faculty supervisors
Completed research work

The product

A benchmark-beating biodiversity classifier across 57 plant species and 18,400+ images, paired with an admin dashboard that monitored experiment runs, class-level failure, confidence intervals, and dataset audits.

Admin dashboard / 70+ monitored runs

State-of-the-art accuracy trajectories, class-level failure, confidence intervals, and dataset audits monitored from one research dashboard.

Research admin dashboard
Live experiment monitor · 070+
Validation accuracy91.7%
+14% relative
Admin dashboard monitoring runs, accuracy, failure, and dataset audits.VGG16 · ResNet-50 · Wald confidence intervals
57plant species classified
18,400+images processed
91.7%best validation accuracy
0.893macro F1 score
14%relative improvement over the initial transfer-learning baseline
35%reduction in redundant samples after diversity selection
70+experiments automatically benchmarked
22 mindataset audit and report workflow, reduced from nine hours

This research pushed biodiversity computer vision past a transfer-learning baseline and into a monitored, reproducible research system. Across two supervised phases, the work combined CNNs, VGG16 transfer learning, custom feature extraction, augmentation, class balancing, and geolocation and temporal context.

Across 57 species and 18,400+ images, the strongest model reached 91.7% validation accuracy and 0.893 macro F1, improving 14% over the initial transfer-learning baseline. That lift was treated as a research result, not a slide metric: class-level failure, confidence, and dataset quality had to explain why the model improved.

The second phase treated dataset quality as a first-class research problem. ResNet-50 embeddings and farthest-point sampling reduced redundant training samples by 35%. Automation benchmarked 70+ experiments, calculated accuracy, precision, recall, F1, and Wald confidence intervals, and compressed repeated dataset auditing and reporting from nine hours to 22 minutes.

To keep that system honest in practice, I built an admin dashboard that monitored experiment runs, accuracy trajectories, class-level failure, confidence intervals, and dataset-audit status so the research group could see what was improving, what was stalling, and what needed another pass.