P. Aebischer: Measuring sward height and dry matter yield of pastures using multispectral imagery from UAV and a random forest algorithm

Bio Information

Philippe Aebischer is a Scientific collaborator in Physics and New Technologies at the School of Agricultural, Forest and Food
Sciences (HAFL) in Zollikofen, Bern. He is partly involved in teaching and partly in research projects with focus on image processing and model development.

Presentation Abstract

In Switzerland, more than 80% of agricultural land consists of semi-natural grassland with a very diverse species composition and a heterogeneous growth pattern. This heterogeneity is reinforced by the hilly terrain
in which most of the pastures are located. A prerequisite for an efficient pasture management is the regular measurement of the sward height by means of a Rising Plate Meter (RPM). The performance of representative measurements with the RPM can be very time consuming, depending on the area.

Moreover, processing the data obtained from the manual RPM measurements usually lacks automation. With the latest generation of unmanned aerial vehicles (UAV) equipped with a Real Time Kinematic (RTK) positioning system and a multispectral camera, it should be possible to measure sward heights and to estimate dry matter yields over large areas with high accuracy (Kaivosoja et al., 2018)). In order to automate data processing as much as possible, we were looking for a solution without complex georeferencing. For this purpose, we developed an algorithm that allows us to calculate a digital terrain model (DTM) out of the digital surface model (DSM) of pastures. This makes it possible to measure the vegetation height without prior marking the area of interest with ground control points (GCPs) and thus without a subsequent referencing of the image. This is a significant improvement of the degree of automation and thus reduces time and costs of the whole process.

Sward height and dry matter yield are estimated using a random forest estimator. To provide the model with reliable data and to make it as robust as possible against seasonal changes, grass growth experiments of
intensively managed pastures at two different locations were flown over with a UAV (DJI P4 Multispectral) weekly from April to October 2020. Different vegetation indices calculated from the reflectance in the red,
green, blue, red-edge and near-infrared channel were used as model features. First results show that the method of measuring sward height without GCPs is comparable to the measuring method including GCPs with a coefficient of determination of 91%. This allows us to use sward height and the resulting statistical quantities out of the 3D-data as additional reliable features in our model.

On a test set splitted from the training data, the Random Forest estimator achieves a prediction accuracy of about 87 % with a deviation of 190 kg/ha. First estimates at a site never seen before, achieved an average
deviation of 11.7 % (240 kg/ha) on an average dry matter yield of 2050 kg/ha. Compared to the RPM measurements with a deviation of 254 kg/ha (Schori, 2020), this first result is very promising and shows that
the approach can be further developed and automated as a practical agricultural application.

Authors: Philippe Aebischer,
Michael Sutter: Scientific collaborator Grassland Management and Ruminant Production Systems (HAFL);
Prof. Dr. Beat Reidy:
Grassland Management and Ruminant Production Systems (HAFL)


Ask the Speaker

You can post your questions for the speaker of this talk here until Nov 8. Your questions are visible to all other registered participants of the conference and will be visible together with your username that consists of the beginning of your email address. Participants can “like” your question. Questions with more like are more likely to be discussed during the live Q&A session on November 10, 2020.