Use Case 1

Classification model portability

1

Training a classification model in one place / one year and investigating which method to adopt to make the trained model valid in another place and/or another year which have no or less in situ data

Using the Bhattacharyya distances between different site gives a metric for the invariability of the set of features

Targeting one of the key bottlenecks in agriculture remote sensing which is the in situ data scarcity

AOIs in France and Belgium, 2019 and 2020

Source domain is Belgium 2019 and target domains are Belgium 2020 and France 2019

Working at parcel-level, using the Land Parcel Identification System dataset

Selecting the most portable set of features (Sentinel bands and/or indices) by using statistical distances.

Using the Bhattacharyya distances between crops on the Belgium 2019 gives a metric for the separability of the classes

Using separability for how good the model will classify and using invariability for how good the model will transfer to another place/year

Selecting the set of features based on the Pareto-optimal candidates

Invited stakeholder: JECAM community