Automatic detection of meliorations

Project aims to automatically detect meliorations on orthophoto maps from http://geoportal.cuzk.cz using machine learning. Images from given coordinates are exported using GDAL in QGIS. Meliorations are detected on the images by binary semantic segmentation with a convolutional neural network. Pixels on the images are classified into two classes: is/is not a melioration; the output of the network is a binary segmentation mask. The network used is U-Net with EfficientNetB0 or VGG16 backbone pretrained on ImageNet. Use to detect whether meliorations are present on an image; accuracy is too low for precise location of all pipes due to low contrast and lack of training data.

How to use

• run in terminal unless stated otherwise
• provide coordinates of image centers in a .csv file (see data.csv for formatting)
• run export.py in QGIS Python Console
• run delete.py to delete images from undefined layers
• download a trained model (trained on GPU in Linux, not guaranteed to work on other systems)
  • EfficientNetB0 backbone: https://drive.google.com/file/d/1O3S760vMr4afAHM_XOiaHAQCZpsx3knN
  • VGG16 backbone: https://drive.google.com/file/d/1DRrk0IqLxqjXEHFiF7wz83qs-4VwM1TF
• run melio_finder_effnet.py or melio_finder_vgg.py to find meliorations on images

How to train (optional)

• find training images and create matching binary segmentation masks
• run train_effnet_unet.py or train_vgg_unet.py to train melio finder