The goal of this project is to track changes in the Amazon rainforest due to deforestation using satellite image data. It is a multi-label image classification problem introduced in a past Kaggle competition Planet: Understanding the Amazon from Space. The final solution consists of an ensemble of five deep learning models using transfer learning with the fastai library which would have placed 19th in the competition.

You can read my full project report here

You can see my original project proposal here

You can find the blog post of my project workflow here

To run the notebooks you can create a conda environment with all necessary software by running:

conda env create -f environment.yml

Alternatively, you can install the libraries listed in requirements.txt however you see fit.

The dataset consists of 40,479 training images and 61,191 test images, and each image is a 256x256 pixel jpeg. The filetrain_v2.csv is included which liststhe training file names and their accompanying labels. The labels can be broken down into three categories: cloud coverlabels, common labels, and less common labels. There are 17 labels in total: clear, partly cloudy, cloudy, haze, primary, water, habitation, agriculture, road, cultivation, bare ground, slash and burn, selective logging, blooming, conventional mining, artisanal mining, and blow down.

The data and all models I used to produce my final solution can be obtained by running the get_data.ipynb notebook.

To reproduce my solution by training from scratch, you first must download the data by running the get_data.ipynb notebook.

Then you can proceed to train the models by running the train.ipynb notebook. You will need to run this notebook five times, once for each of the following fastai models: resnet50, resnet101, resnet152, densenet121, densenet169. You can change which architecture you are using by changing the line in cell [8] to whichever pretrained model you wish to train, for example:

arch = models.resnet152

To create an ensemble of the models you wish to include, you can run the create_ensemble.ipynb notebook. Simply define the model_list in cell [4] to include the models you wish like so:

model_list = ['resnet50.pkl','resnet101.pkl', 'resnet152.pkl', 'densenet121.pkl', 'densenet169.pkl']

To make predictions and submit to the Kaggle competition to see your score, run the notebook predict_and_submit.ipynb. To submit directly from the notebook you will need the Kaggle command line tool installed. Directions on how to set this up can be found in the get_kaggle_data.ipynb notebook. Alternatively, you can manually submit to the competition by navigating to the competition page and select the .csv file you'd like to submit.

Note that in either case, you will need to accept the competitions terms before being able to submit to the competition.