Overview

Problem Statement

Purpose of the project is to create deep learning based system for creating maps of urbanized areas from satellite images.

Approach

The concept of the entire project is quite simple, we take a satellite image, divide it into small pieces of 50x50 pixels, which are then classified by a previously trained model, and finally, based on the classified data, we create a map.

Used technologies

Entire projects is written in python, satellite images are fetched from qgis, ml part is made using PyTorch.

Project structure

   ├── data
   |     ├── processed            # Data for training
   |     |       ├── images
   |     |       ├── test.csv
   |     |       ├── train.csv
   |     |       └── validation.csv
   |     ├── raw                     # Data to label and split
   |     |    └── images
   |     ├── labeler.py            # PyQt gui for labeling images
   |     ├── split_dataset.py        # script spliting data to train, val and test sets
   |     └── split_image.py         # script spliting satellite image to 50x50 pixels tiles
   ├── maps                   
   ├── src
   |     ├── ml
   |     |   ├── dataset            # Pytorch datasets, dataloaders and transformations
   |     |   |      ├── dataset.py
   |     |   |      └── transformations.py
   |     |   ├── models                   # Models definitions and corresponding weights files
   |     |   |  
   |     |   ├── scripts                  # scripts for training, evaluating and predicting
   |     |   |      ├── predict.py
   |     |   |      ├── test.py
   |     |   |      └── train.py
   |     |   └── utils                  # Everything else used for training
   |     |          ├── early_stopping.py
   |     |          ├── metrics.py
   |     |          ├── optimizers.py
   |     |
   |     ├── map_generator             # Code for generating map from model predictions
   |     |         ### TODO ###
   |     ├── qgis_api                  # Api for fetching setallite images
   |               ### TODO ###

Prototype