Deployment of models predicting compressed sward height on Wallonia: quality and validity of the predictions

Currently, there is a high interest to integrate remote sensed data and machine learning algorithms to develop pastures management tools. In this context, over the past two years, we published models predicting the available compressed sward height (CSH) in pasture using Sentinel-1, Sentinel-2, and meteorological data. Those scalable models were developed to be the basis of a decision support system (DSS) available for Walloon farmers. A platform predicting CSH over Wallonia was developed and this presentation aims to provide some insights in its prediction capabilities by detailing the values and their variability at the parcel level for the year 2021. The first prospect was the distribution of the predictions without considering a mean per parcel: 2% of the predicted values were out of the [0:250] mm of CSH range on which model were trained and the median was around 56mm. If we were to consider a mean per parcel, the values fall back in training range and the median is around 60mm and coefficient of variation (cv) of the prediction within each parcel, which indicated the variability, ranged from 0 to 986. These cv values indicated that the predictions might be too variable and that further training data are required to get more stable outputs. However, given that the range of predictions respects the training CSH range, the predicting platform was considered mature enough to be ready for being a data provider for a future DSS, the main interest in this future tool being to provide everyday food availability to the farmer in order to help him manage his feed wedge and therefore improving cattle welfare.