NIVA - Automatic field delineation

This repo contains code to generate automatic contours for agricultural parcels, given Sentinel-2 images. This code has been used ot generate contours for Lithuania and the province of Castilla y Leon.

You can find more information about this project in the blog post Parcel Boundary Detection for CAP.

Introduction

This sub-project is part of the "New IACS Vision in Action” --- NIVA project that delivers a suite of digital solutions, e-tools and good practices for e-governance and initiates an innovation ecosystem to support further development of IACS that will facilitate data and information flows.

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 842009.

Please visit the website for further information. A complete list of the sub-projects made available under the NIVA project can be found on gitlab

Requirements

The field delineation pipeline uses SentinelHub service to download Sentinel-2 imagery, in particular using the large-scale batch processing API. The batch processing allows to download images over large Areas of Interest in a very fast and efficient manner. The data is automatically stored in S3 buckets, which need to be adequately configured.

Installation

The fd Python 3.5+ package allows to execute the end-to-end field delineation workflow.

To install the fd package, clone locally the repository, and from within the repository, run the following commands:

pip install -r requirements.txt
python setup.py install --user

In addition, the field delineation workflow uses the following:

• Docker containers;
• psql PostgreSQL database;
• gdal geospatial processing library, version >2.4.0.

The numbered notebooks showcase how to execute the end-to-end workflow.

AWS set-up

To run the notebooks, the bucket-name bucket permission need to be set-up, as described below.

bucket_name = "bucket-name"
aws_access_key_id = ""
aws_secret_access_key = ""
region = "eu-central-1"

Sentinel-Hub credentials

Sentinel-Hub credentials need to be added to the download notebook.

Input data

In order to execute the entire workflow, including training of the deep learning model, the following files are required:

• file with geometry of the AOI (e.g. in GeoJSON format);
• file with reference GSAA parcel boundaries (e.g. in ESRI Shapefile format);
• the time-interval over which we want to estimate parcel boundaries. Predictions can be made for sub-intervals of this given time period (e.g. aggregation can be done over an arbitrary time interval as a post-processing step)

Content

This repository has the following content:

• fd: modules implementing each part of the workflow;
• input-data: folder storing the file defining the AOI and the consequent grid definition file;
• notebooks: folder storing the example notebook to execute the end-to-end workflow.

End2End Execution

The field delineation workflow has been designed to scale to large AOIs, by downloading data quickly and efficiently, and by parallelizing execution of pipelines over the tiled data.

The End2End notebook showcases the entire procedure to reproduce the entire end-to-end workflow. The following steps are executed:

• Data download: downloading the Sentinel-2 images (B-G-R-NIR) using Sentinel-Hub Batch API;
• Conversion of tiffs to patches: converts the downloaded tiff files into EOPatches (see eo-learn), and computes cloud masks from cloud probabilities;
• Vector to raster: adds reference vector data from a database to EOPatches and creates reference masks used for training of the model;
• Patchlets sampling: sample EOPatches into smaller 256x256 patchlets for each cloud-free time-stamp. The sampling can be done for positive and negative examples separately;
• Patchlets to npz files: the sampled patchlets are chunked and stored into multiple .npz files, allowing to efficiently access the data during training;
• Create normalization stats: compute normalisation factors for the S2 bands (e.g. B-G-R-NIR) per month. These factors will be used to normalise the data before training and evaluation;
• Patchlets split into k-folds: split patchlets into K-folds, allowing to perform a robust cross-validation of the models;
• Train model from cached npz: train k-models, one for each left out fold. The ResUnet-a architecture implemented within eo-flow is used as model. A single model can be derived by averaging the weights of the k-fold models;
• Predict eopatches: use the trained models to predict parcel boundary probabilities for the entire dataset;
• Post process predictions: merge the predictions temporally and combine the predicted extent and boundary probabilities. A time interval can be specified over which the predictions are temporally aggregated;
• Create vectors: vectorise the combined field delineation probabilities;
• Utm zone merging: combine spatially vectors from multiple UTM zone if applicable.