Kaggle Planet ConvNet

Convolutional neural network developed for the Planet: Understanding the Amazon from Space Kaggle Competition

Overview

This project is an implementation of a convolutional neural network that classifies satellite imagery of the Amazon Rainforest into 17 categories. Both Keras and Tensorflow versions of the convnet are implemented. Learns to classify images with >90% accuracy.

Setup Instructions

Python 2.7.1 or above is required, preferably the Anaconda Distribution

Dependencies

Make sure to have the following dependencies installed:

NumPy (included in Anaconda)
Pandas (included in Anaconda)
CV2
TQDM
Keras
Tensorflow

All of these modules can be installed via pip.

Getting the Dataset

For the network to run, training and test data and the correct training labels will be needed. All can be downloaded from the competition website. Look for the files train_v2.csv.zip, train-jpg.tar.7z and test-jpg.tar.7z.

Running the Model

Extract the train and test datasets to any directory. Ensure the two extracted folders are named train-jpg and test-jpg respectively.

Keras Version

In line 12 of Preprocessing_Keras.py AND line 18 of Planet_Keras.py, assign the relative path to the train data directory (i.e. the parent directory of train-jpg) to pre_filepath.
Run Preprocessing_Keras.py. You should get numpy arrays of saved data in the pre_filepath directory.
Run Planet_Keras.py.

Tensorflow Version

In line 14 of Planet_TF.py, assign the relative path to the train and test dataset to the variables pre_filepath_train and pre_filepath_test
Extract the CSV file and place it in the same directory as Planet_TF.py. Ensure the CSV file is named train_v2.csv.
Run the network. Note that it may take a while to train especially on CPU only.

Note: After training, the EPOCHS constant may need to be adjusted, depending on how many epochs the network takes to converge.

chriswang101/Kaggle-Planet