Adsa Logo White Adsa Title White

Prediction of compressed sward height of Walloon pastures from sentinel-2 images using machine learning algorithms.

C. Nickmilder




Prediction of compressed sward height of Walloon pastures from sentinel-2 images using machine learning algorithms.
C. Nickmilder*1, A. Tedde1, P. Lejeune1, I. Dufrasne2, F. Lessire3, B. Tychon4, F. Lebeau1, H. Soyeurt1. 1TERRA, ULiege Liege, Belgium, 2Departement de gestion veterinaire des Ressources Animales (DRA) / Nutrition des animaux domestiques, ULiege Liege, Belgium, 3Fundamental and Applied Research for Animals and Health (FARAH), ULiege Liege, Belgium, 4Departement des sciences et gestion de l environnement (Arlon Campus Environnement), ULiege Liege, Belgium.

ROADSTEP is a Walloon research program aiming to develop decision tools to help farmers in their daily herd monitoring on pastures. One of the aims is to develop a modeling tool to predict the availability of pasture feeding based on satellite images, meteorological variables and soil characteristics. 7737 compressed sward heights (CSH) were measured on 2 farms recorded with Jenquip EC20G platemeter in July and August 2019. They were used to calibrate and validate 73 predictive models of CSH. The tested algorithms were linear regression, lars, cubist, generalized linear model, neural network, random forest and linear support vector machine. The explaining variables were the 11 sentinel-2 reflectance bands at the bottom of atmosphere. Those bands and CSH were introduced directly in the model but also through their logarithm, square-root, square and cube forms to test the possible nonlinear relationships between them. The reduction of dimensionality of X-matrix through the estimation of principal components as well as partial least squares factors was also tested. To guarantee independence between calibration and validation, calibration was made on CSH (ranging from 12 to 158 mm with an average value of 59.4+-22.3 mm) measured on a farm and validation on CSH (ranging from 13 to 247.5 mm with an average value of 53.2+-21.6 mm) measured on an other farm. The model that performed the best was a generalized linear model from the gamma family using an inverse link function. Calibration and validation RMSE were respectively equal to 17.4 and 20.7 mm or 29.3 and 28.9% of their respective mean. These results are only preliminary. Additional sampling periods and pastures are needed to improve the models robustness. Moreover, the second step of this research will consist in adding information related to meteorological data and soil characteristics to enhance the prediction power of the developed models.

Keywords: remote sensing, compressed sward height, machine learning.