UAV based mapping of variation in grassland yield for forage production in Arctic environments

Grassland cultivation for animal feed is the key agricultural activity in northern Norway. Even though the growing season has increased by at least a week in the last 30 years, grassland yields appear to have declined, probably due to more challenging winter conditions and changing agronomy practices. The ability for local and regional crop productivity forecasting would assist farmers with management decisions and would provide local and national authorities with a better overview over productivity and potential problems due to e.g. winter damage. Remote sensing technology has long been used to estimate and map the variability of various biophysical parameters, but calibration is important. In order to establish the relationship between spectral reflectance and grass yield in northern European environments we combine Sentinel-2 time series, UAV-based multispectral measurements, and ground-based spectroradiometry, with biomass analyses and observations of species composition. In this presentation we will focus on the results from the UAV data acquisition. We used a multirotor UAV with different sensors (a multispectral Rikola camera, and NDVI and RGB cameras) to image a number of cultivated grasslands of different age and productivity in northern Norway in June/July 2016 and 2017. Following UAV data acquisition, 10 to 20 in situ measurements were made per field using a FieldSpec3 (350-2500 nm). In addition, samples were taken to determine biomass and grass species composition. The imaging and sampling was done immediately prior to harvesting. The Rikola camera, when used as a stand-alone camera mounted on a UAV, can collect 15 bands with a spectral width of 10-15 nm in the range between 500-890 nm. In the initial analysis of the 2016 data we investigated how well different vegetation indices correlated with biomass and showed that vegetation indices that include red edge bands perform better than widely used indices such as NDVI. We will extend the analysis with partial least square regression once the 2017 data becomes available and in this presentation we will show the results of both the partial least square regression analysis and vegetation indices for the pooled data from the 2016 and 2017 acquisition.