Deployment of models predicting compressed sward height on Wallonia: results and feedback

There is currently high interest in integrating data linked to remote sensing and methods from the
machine-learning domain to develop tools to support pasture management. In this context, over the
past two years, we have published models predicting the available compressed sward height (CSH) in
pastures using Sentinel-1, Sentinel-2, and meteorological data. These scalable models could provide the
basis of a decision support system (DSS) available for Walloon farmers. A platform performing the CSH
prediction was developed and this paper aims to provide some insights in its prediction capabilities and
tackle the challenge of using data acquired at different moments in time. Predictions were made from
the beginning of January until the end of October 2021 using our most promising published models.
After data cleaning, the coefficient of variation of CSH predictions, calculated for each studied date
(n=35) and parcel (n=192,862), ranged from 0 to 986. This extreme variation suggests some prediction
imperfections. Before the integration of the platform in a DSS, the main task to solve is the issue of
missing or non-operational S1 or S2 data. Indeed, even if a gap filling method was applied, only 62% of
potentially exploitable dates were usable.