LandCover Segmentation

Description

This repository is an example pipeline for training and evaluating the segmentation model for Deepness: Deep Neural Remote Sensing plugin for QGIS.

This repository contains code for training and evaluating the LandCover segmentation model based on the LandCover.ai dataset.

Ready-to-use model and example inference in QGIS are available in our LandCover segmentation example.

Requirements

Create your Python 3 virtual environment and install requirements:

pip3 install -r requirements.txt

Usage

• Prepare dataset for training and evaluation

  # create a directory for the dataset
  mkdir data/LandCover
  
  # move to the dataset directory
  cd data/LandCover
  
  # download dataset
  wget "https://landcover.ai.linuxpolska.com/download/landcover.ai.v1.zip"
  
  # extract dataset
  unzip landcover.ai.v1.zip
  
  # (optionally) remove the zip file
  rm landcover.ai.v1.zip
  
  # create train-val-test split
  python3 split.py
  

• Logging configuration

  • NEPTUNE

    If you want to use Neptune logger, you have to set your keys:

    export NEPTUNE_API_TOKEN=""
    export NEPTUNE_PROJECT_NAME=""
    

  • NONE

    If you don't want to use any logger, just add ~logger in the commands below.

• Run training pipeline

  python run.py name=landseg
  

• You can run an evaluation pipeline if you have trained model

  python run.py name=landseg eval_mode=True trainer.resume_from_checkpoint=./path/to/model
  

• Run export pipeline to export model to ONNX format to use it in QGIS

  python run.py name=landseg eval_mode=True trainer.resume_from_checkpoint=./path/to/model export.export_to_onnx=True