A take on the Statoil/C-CORE Iceberg Classifier Challenge. Kaggle Competition
(Sourced from Kaggle)
In this competition, you will predict whether an image contains a ship or an iceberg. The labels are provided by human experts and geographic knowledge on the target. All the images are 75x75 images with two bands.
train.json, test.json The data (train.json, test.json) is presented in json format. The files consist of a list of images, and for each image, you can find the following fields:
id - the id of the image band_1, band_2 - the flattened image data. Each band has 75x75 pixel values in the list, so the list has 5625 elements. Note that these values are not the normal non-negative integers in image files since they have physical meanings - these are float numbers with unit being dB. Band 1 and Band 2 are signals characterized by radar backscatter produced from different polarizations at a particular incidence angle. The polarizations correspond to HH (transmit/receive horizontally) and HV (transmit horizontally and receive vertically). More background on the satellite imagery can be found here. inc_angle - the incidence angle of which the image was taken. Note that this field has missing data marked as "na", and those images with "na" incidence angles are all in the training data to prevent leakage. is_iceberg - the target variable, set to 1 if it is an iceberg, and 0 if it is a ship. This field only exists in train.json.
- Remove columns with 'na'.
- Remove 'inc_angle' from features as more than %10 of training examples don't have a value.
- Apply data normalization.
- Apply median filter. (Since the images are pixelated, applying median filter helps to achieve more accurate edgemaps.)
- Extract edgemaps of both bands.
- Stack band images and edgemap images seperately.
CNN architecture involves a two-stream CNN. One for actual stacked band images and another for stacked edgemap images. These two networks are fused together before the last fully-connected layer. (Early-fusion)
- Learning rate: 0.001
- Batch size: 64
- Number of epochs: 200
- Loss function: Cross-Entropy loss
- Optimizer: Stochastic Gradient Descent