Harvey Wu, Tyr Wiesner-Hanks, Ethan L. Stewart, Chad DeChant, Nicholas Kaczmar, Michael A. Gore, Rebecca J. Nelson, and Hod Lipson
Summary: using deep learning, we detect a corn disease, Northern Leaf Blight (NLB), through aerial images of corn fields captured by small UAVs.
Code organization:
/boom_transfer.py trains a CNN on images stored in the directory /new_data (you need to create it).
/test.py tests a trained CNN stored in /models on a test set of data in /new_data.
/boom_heatmaps.py generates heatmaps using a trained CNN stored in /models.
scripts/crop_lesions.py samples subimages containing lesions from a set of labeled images containing lesions.
scripts/crop_nonlesions.py samples subimages without lesions from a set of labeled images without lesions.
scripts/yesno.py splits the dataset specified in [1] into images with lesions and images without.
scripts/overlay.py creates a composite image with heatmap overlaid on top of the original.
scripts/drawlines.py draws the annotations onto the original images (semimajor axis of lesion).
scripts/*.pkl contain image names; they correspond to the train/val/test split used in our experiments.
Contact wu.harvey (at) columbia.edu with any questions.