We address the problem of whale counting in large scale areas represented by a large number of images by combining two models, an image classification model and a whale detection model in images. To build these models, we need to build two training datasets, one for each model.
Convolutional Neural Networks with Tensorflow API
We used TensorFlow, an open source library for numerical computation, specializing in machine learning applications.
This codelab was tested on TensorFlow 1.12
pip install --upgrade "tensorflow==1.12.*"
Step 1. Presence/absence with Tensorflow (https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/#0) Architecture: Inception v3 Transfer learning: ImageNet Data-augmentation:flip left right, random crop, random scale, random brightness
parser.add_argument( '--flip_left_right', default=True, help="""\ Whether to randomly flip half of the training images horizontally.\ """, action='store_true' ) parser.add_argument( '--random_crop', type=int, default=1, help="""\ A percentage determining how much of a margin to randomly crop off the training images.\ """ ) parser.add_argument( '--random_scale', type=int, default=1, help="""\ A percentage determining how much to randomly scale up the size of the training images by.\ """ ) parser.add_argument( '--random_brightness', type=int, default=1, help="""\ A percentage determining how much to randomly multiply the training image input pixels up or down by.\ """ )
Step 2. Counting with Tensorflow API object-detection (https://github.com/tensorflow/models/tree/master/research/object_detection) Architecture: Inception-Resnet v2 Transfer learning: Coco
For ships, we selected 400 images from 100 ports around the world. For water and submerged rocks, we selected 400 coastal images randomly around the world, excluding ports and polar regions. For whales, we focused on 10 coastal areas that are known for marine mammal diversity and whale watching (13,348 images in total).
Automatic whale counting in satellite images with deep learning Emilio Guirado, Siham Tabik, Marga L. Rivas, Domingo Alcaraz-Segura, Francisco Herrera doi: https://doi.org/10.1101/443671
- Arkive photographers, NOAA fisheries photo database
- Tensorflow developers
- Google Earth developer team