This repository contains an implementation of Polsarnet, which was created in the context of the module 'Hot topics in computer vision' at TU Berlin.

The goal of the project is to evaluate the Paper PolSARNet: A Deep Fully Convolutional Network for Polarimetric SAR Image Classification. Therefore the implementation is tested on two POLSAR datasets.

These datasets are:

The L-band POLTOM dataset of the region Oberpfaffenhofen, as used in Exploiting GAN-Based SAR to Optical Image Transcoding for Improved Classification via Deep Learning()

The Sentinel-1 radar images of the city Wroclaw (Poland), as used in Exploiting SAR Tomography for Supervised Land-Cover Classification

A list of dependencies is found in requirements.txt.

The dataloader package contains all the files needed for loading these two datasets.

The models package contains the implementation of the networks (Polsarnet.py) and the implementation of the complex building blocks (complexNetwork.py).

The code used for the metrics is found in evaluation.py.

The training is done by calling either oberpfaffenhofen.py or poland.py.