mintusf/land_cover_segmentation

Segmentation for land cover using raw satellite data (Sentinel 2La)

Summary

Land cover detection using Sentinel satellite data with instance segmentation.

How to run

Environment setup

1. If you want to create a local environment (e.g. on AWS), please run env/setup_env.sh
2. If you want to create a docker container, please run env/setup_docker.sh)

Training

1. Create your yml config file in config folder, according to parameters in default.py
2. Run training with python train.py --cfg CONFIG_PATH
3. Experiment can be tracket in comet_ml. In order to do so, please add your API_KEY in .comet.config
4. Weights are saved to weights folder
5. Logs can be found in logs folder

Testing

1. In order to test and get class-wise metrics on test set, run evaluate.py --cfg CONFIG_PATH --checkpoint WEIGHTS_PATH

Inference

1. In order to make inference, run infer.py --cfg CONFIG_PATH --checkpoint WEIGHTS_PATH --sample_list PATH_TO_SAMPLES_TXT --outputs SELECTED_OUTPUTS. Possible outputs are: alphablend (png), raster (mask with geographic coordinates), alphablended_raster (alphablend with geographic coordinates) and raw_raster (input image).
2. Weigths for 3 bands input (results shown below) are available here

Folders:

• config: training configs, dataset configs
• env: scripts to build docker environment
• logs: training logs
• models: models implementation
• tests: code tests
• tools: visualization tool
• train_utils: utils for training
• utils: I/O, visualization, raster utils
• weights: for weights

Dataset

Reference:

• https://github.com/chrieke/awesome-satellite-imagery-datasets/blob/master/README.md
• https://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/IV-2-W7/153/2019/isprs-annals-IV-2-W7-153-2019.pdf

Used datasets:

SEN12MS (LINK)

• Spatial resolution: 10m
• Patch size: 256 x 256
• Classes: 33 (converted to 12)
• According to the paper, the accuracy of annotation is ~80%, which is also limitation for the training.
• Channels: 12 bands of Sentinel L2C (Only RBG bands used for training and its results are shown in this document)
• Dataset for class snow was created by this repository creator using EO Browser

Models

• DeepLab v3
• HRNet (implementation, paper)

Results

Metrics (test set)

class	recall	precision	f1
shrublands	0.85	0.73	0.79
urban	0.70	0.60	0.65
agricultural	0.90	0.76	0.82
barren	0.43	1.00	0.60
snow	0.91	0.93	0.92
water	0.92	0.97	0.94
dense_forest	0.42	0.77	0.54
open_forest	0.53	0.54	0.54
wetlands	0.66	0.23	0.23
grasslands	0.48	0.38	0.42

Confusion matrix (test set)

Generated masks example

Real-life Web App

The model is used in this repository as a backend for land cover classification.