In this project I created a map of aquaculture facilities in British Columbia, Canada. Starting from a starting set of images with nets present and images without nets, a classification model is trained.
Then, new aquaculture candidates are extracted by predicting on randomly sampled images and manually selecting the correct ones from the candidates with highest probability. This process was repeated 3-5 times until a total of 112 instances of aquaculture nets and 7514 images without nets were extracted.
I visualized it in QGIS below. Blue dots are randomly sampled positions without nets and the red dots show the 112 instances of nets.
- Download images with nets present with 2019-11-16-pp-Download_net_images_and_around_nets.ipynb.
- Download random images (most of them will be without net) with 2019-11-16-pp-Download_random_images.ipynb.
- Resize images and train a ResNet18 Classification model with 2019-11-16-pp-Modeling.ipynb.
- Predict the net probability for the random images and manually select the ones with nets present by looking at those with highest probability with 2019-11-17-pp-Predict.ipynb.
- Use the remaining random images that you did not select as having nets present as training examples for "no net present".
- Repeat steps 3-5 until you have covered the search space.