Validation of a workflow based on Sentinel-2, Sentinel-1 and meteorological data predicting biomass on pastures

This study develops the validation of the four best promising models resulting from a workflow
processing Sentinel-1, Sentinel-2 and meteorological data through 13 different machine learning
algorithms that led to 124 models predicting biomass under the form of compressed sward height on
square sub-samples of paddocks (i.e., pixel-based estimation with a resolution of 10 m). The training
and validation data were acquired in 2018 and 2019 in the Walloon Region of Belgium with a rising
platemeter equipped with a GPS. The cubist, perceptron, random forest and general linear models had
a validation root mean square error (RMSE) around 20 mm of CSH. However, the information relevant
for the farmer and for integration in a decision support system is the amount of biomass available on the
whole pasture. Therefore, those models were also validated at a paddock-scale using data from another
farm (117 CSH records acquired with a different rising platemeter) based on input variables expressed
at paddock scale or predictions aggregated at paddock scale. The resulting RMSE were higher than
before. To improve the quality of prediction, a combination of the outputs of the models might be needed.