Deepglobe Landcover Classification Challenge 2018

The Deepglobe Landcover Classification Challenge 2018 was one of three challenges launched by DeepGlobe in 2018 in order to push forward the research of computer vision technologies in the field of satellite images.

Setup

Dependencies

The current code base works in python=3.9

• torch >= 2.0.0
• torchvision >= 2.0.0 The rest can be installed with pip running:

pip install -r requirements.txt

Running the training script in kaggle

In case you don´t have a machine with GPU, which is highly needed for our case, you can optionally run the training script on kaggle and user their free GPUs. To do that simply go to the train_kaggle.ipynb notebook and click on the "Open in kaggle" button on the top left. This will take you to the Kaggle website where you can edit and run the notebook with ease.

Dataset download and preparation

Note: In case you decided to run the training script form Kaggle, you can skip this step.

The dataset from the DeepGlobe 2018 Landcover Classification challenge is hosted in Kaggle. For ease of use you can just run the python script prepare.py that is located in the data folder. This script will download the dataset using the python Kaggle API and prepare it for training. Remember that in order for this to work you need to have the Kaggle API downloaded (pip install -r requirements.txt) and also have your kaggle private key properly set up in your machine. After that you can run the following command from the root of the project:

python3 data/prepare.py

This will fill the data

data
├── images    -> RGB satellite images
├── masks     -> gts segmentation masks in label encoding format 
└── raw_masks -> gts segmentation masks in RGB format 

Log training to Weights and Biases

Locally: If you want to log to W&B you need create a file named key.py in the root of the project and add a variable named wandb_key with your API key. After doing that you simply need to add the argument --log to the training script train.py in order to create a new run to wandb.

When running in Kaggle: If you want to log the training to Weights and Biases you should first add their API key as a Kaggle secret. To do this, go to the "Add-ons" tab and click on "Secrets", then add a new secret with the label "wandb_key" and the value of your private API key. After that is done simple add the --log when running the train.py script.

Dataset

The dataset consists of a total of 803 satellite images of 2448x2448 pixels, each image comes with a segmentation mask that classifies the pixels of the image in the following classes:

File Name	Explanation / Function
Urban land	Man-made, built up areas with human artifacts
Agriculture land	Farms, any planned (i.e. regular) plantation, cropland, orchards, vineyards, nurseries,and ornamental horticultural areas; confined feeding operations
Rangeland	Any non-forest, non-farm, green land, grass
Forest land	Any land with at least 20% tree crown density plus clear cuts
Water	Rivers, oceans, lakes, wetland, ponds
Barren land	Mountain, rock, dessert, beach, land with no vegetation
Unknown	Clouds and others

Model

Vanilla UNet architecture from the original paper.

• Reproducibility: To ensure the reproducibility of trainings in PyTorch, it's necessary to take some measures. First, a manual seed is set for generating random numbers, which ensures that the numbers generated in each run are always the same. In addition, the cuDNN optimization is disabled, which although it can improve performance, can produce non-deterministic results. By disabling this optimization, performance is ensured to be more consistent across different systems. Finally, the cuDNN determination is activated, which guarantees that the same results are generated in each run.

Training

The training parameters are stored in a YAML file so that each parameter of the experiments is documented and the same experiment can be reproduced at any time.

Evaluation

Inference

References