Reproduce “Small data deep learning methodology for in-field disease detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
David Herrera-Poyato, Jacinto Domínguez-Rull, Rosana Montes, Inés Hernánde, Ignacio Barrio, Carlos Poblete-Echeverria, Javier Tardaguila, Francisco Herrera and 1 more
A small data deep learning methodology (ISD⁴L) was developed for in-field potato late blight detection using high-resolution RGB images. The approach accurately identified all nine diseased cases and correctly classified 12 of 13 healthy plants in a challenging field environment, achieving 95.45% overall accuracy.
Key insights
- Random patch generation with varying rotations and zooms effectively creates a diverse and robust training set from a few high-resolution images, enabling deep learning on small datasets.
- Focal Loss is critical for effectively handling class imbalance and improving model robustness by emphasizing hard examples, which is crucial when dealing with subtle symptoms and complex field backgrounds.
- Aggregating predictions from overlapping sliding windows provides reliable image-level disease detection, mitigating the impact of individual patch misclassifications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.