Case Study: Jacques Marais Media achieves faster model training (from three weeks to three days), improved elephant-detection accuracy (56% → 67%) and 97% less manual verification with Valohai

Data scientists are changing nature conservation with deep learning

Jacques Marais Media applied machine learning to detect African forest elephants from aerial infrared and RGB imagery but faced large datasets (typical training sets ~30 GB), slow model training and high manual verification costs. To scale and speed up research while keeping results reproducible, the team turned to Valohai and its Valohai platform to manage infrastructure and experiments.

Using transfer learning (InceptionV3 with TensorFlow/Keras), cloud compute and Valohai to orchestrate hyperparameter optimization across nine AWS instances, the project cut a three‑week retraining campaign down to three days, raised detection accuracy from 56% to 67%, and dramatically reduced overdetection. The infrared region‑of‑interest model trimmed candidate regions from 55,507 to 1,875 (a 97% reduction, missing only 1% of manually counted elephants), while Valohai provided reproducible experiment tracking and efficient large‑scale resource use.

Jacques Marais Media

Jacques Marais

Faculty