Awesome

Forest Change Detection: baseline models

Source code for "DEEP LEARNING FOR HIGH-FREQUENCY CHANGE DETECTION IN UKRAINIAN FOREST ECOSYSTEM WITH SENTINEL-2 (K. Isaienkov+, 2020) paper

Project structure info

• data_prepare - scripts for data download and preparation
• segmentation - investigation about model approach, model training and model evaluation of clearcut detection

Credential setup

This project needs several secure credentials, for peps.cnes.fr and sentinel-hub. For correct setup, you need to create peps_download_config.ini (it could be done by example peps_download_config.ini.example) and feel auth, password and sentinel_id parameters.

Model Development Guide

Data downloading

1. Create an account on https://peps.cnes.fr/rocket/#/home

2. Specify params in config file input/peps_download_config.ini : most valuable parameters are start_date,end_date,tile, which have to be set to exact values, and latmin,latmax,lonmin,lonmax, which could be specified approximately for a given tile.

3. Download an image archive with python peps_download.py.

4. Unzip the archive.

5. Merge bands with python prepare_tif.py --data_folder … --save_path … for a single image folder, or ./PREPARE_IMAGES.sh "data_folder" "save_path" for the catalogue of images.

6. Run prepare_clouds.py (by defaults, this script is executing with ./PREPARE_IMAGES.sh "data_folder" "save_path" script). This scripts works for Level-C images only (detection clouds with sentinel-cloud-detector API).

7. Download global land cover map (for the Central Europe region): wget https://s3-eu-west-1.amazonaws.com/vito.landcover.global/2015/E020N60_ProbaV_LC100_epoch2015_global_v2.0.2_products_EPSG-4326.zip

   • Unzip archive
   • Run script python prepare_landcover.py --save_path ... --data_path .../E020N60_ProbaV_LC100_epoch2015_global_v2.0.2_discrete-classification_EPSG-4326.tif

The output of scripts are .tif merged bands, .png image, each separately for each band, _clouds.tiff images with clouds detections, and land_cover.tiff file with land cover classification.

If you want to use cloud detector for Level-A images, or prepare more bands, you can use the scripts from the data_prepare directory in time-dependent folder.

Also, data could be downloaded manually from https://scihub.copernicus.eu/dhus/#/home:

1. Create an account

2. Specify the tileID in search query (e.g., 36UYA)

3. Choose the wanted dates of imaging, and download with the arrow-shaped buttom (Download Product)

Keep in mind, that scihub.copernicus.eu allows to download 3 images maximum at once.

Data preparation

1. Create folder where the following data are stored:

   • Source subfolder stores raw data that has to be preprocess (source)
   • Input subfolder stores data that is used in training and evaluation (input)
   • Polygons subfolder stores markup (polygons)
   • Subfolder containing cloud maps for each image tile and land cover classification map (auxiliary)

2. The source folder contains folders for each image that you downloaded. In that folder you have to store TIFF images of channels (in our case, 'rgb', 'b8', 'ndvi' channels) named as f”{image_folder}_{channel}.tif”.

3. If you have already merged bands to a single TIFF, you can just move it to input folder. But you have to create the folder (it may be empty) for these images in the source folder.

4. The polygons folder contains markup that you apply to all images in input folder. Polygons have to be as a single file.

Example of data folder structure:

data
├── auxiliary
│   ├── land_cover.tiff
│   ├── image0_clouds.tiff
│   └── image1_clouds.tiff
├── input
│   ├── image0.tif
│   └── image1.tif
├── polygons
│   └── markup.geojson
└── source
    ├── image0
    │   ├── image0_ndvi.tif
    │   ├── image0_b8.tif
    │   └── image0_rgb.tif
    └── image1
        ├── iamge1_ndvi.tif
        ├── image1_b8.tif
        └── image1_rgb.tif

5. Run preprocessing on this data. You can specify other params if it necessary (add --no_merge if you have already merged channels with prepare_tif.py script).

python preprocessing.py \
 --polys_path ../data/polygons/markup.geojson \
 --tiff_path ../data/source
 --save_path ../data/input
 --land_path ../data/auxiliary
 --clouds_path ../data/auxiliary 

The output of this scipt are subfolders in the input path, which contain divided tile images \ tile masks into the small pieces, with specified --width and --height.

Example of input folder structure after preprocessing:

input
├── image0
│   ├── geojson_polygons
│   ├── image0.png
│   ├── image_pieces.csv
│   ├── images
│   ├── masks
│   ├── clouds
│   └── landcover
├── image0.tif
├── image1
│   ├── geojson_polygons
│   ├── image1.png
│   ├── image_pieces.csv
│   ├── images
│   ├── masks
│   ├── clouds
│   └── landcover
└── image1.tif

6. Run data division script with specified split_function (default=’geo_split’, which divides the data with geospatial split and removes empty masks) to create train/test/val datasets.

python generate_data.py --markup_path ../data/polygons/markup.geojson

Model training and evaluation

1. If it necessary specify augmentation in pytorch/dataset.py

2. Specify hyperparams in pytorch/train.py

3. Run training, prediction, and evaluation. Example:

#! /bin/bash

name=example
echo "$name"
mkdir ../data/predictions/$name
python train.py --lr 1e-3 --network unet50 --name $name --dataset_path ../data/input/ --train_df ../data/train_df.csv --val_df ../data/val_df.csv

python prediction.py --data_path ../data/input --model_weights_path ../logs/$name/checkpoints/best.pth --test_df ../data/train_df.csv --save_path ../data/predictions/$name --network unet50
python evaluation.py --datasets_path ../data/input --prediction_path ../data/predictions/$name/predictions --test_df_path ../data/train_df.csv --output_name $name'_train'


python prediction.py --data_path ../data/input --model_weights_path ../logs/$name/checkpoints/best.pth --test_df ../data/val_df.csv --save_path ../data/predictions/$name --network unet50
python evaluation.py --datasets_path ../data/input --prediction_path ../data/predictions/$name/predictions --test_df_path ../data/val_df.csv --output_name $name'_val'


python prediction.py --data_path ../data/input --model_weights_path ../logs/$name/checkpoints/best.pth --test_df ../data/test_df.csv --save_path ../data/predictions/$name --network unet50
python evaluation.py --datasets_path ../data/input --prediction_path ../data/predictions/$name/predictions --test_df_path ../data/test_df.csv --output_name $name'_test'