Grassland Yield Estimation Using Transfer Learning from Remote Sensing Data

Grassland is one of the most important resources for dairy farmers around the world. Deeper insights into the properties of grassland enable new applications. In particular, site-specific yield information is valuable for objective farm resource planning, fertilization, and field logistics. The Sentinel-2 satellites provide multi-spectral images with a spatial resolution of \(10 \times 10\) m. According to recent studies, these satellite data are successfully used to predict the yield of arable crops. The biggest challenges for satellite data in the visible and near-infrared spectrum are atmospheric disturbances, such as clouds or fog. Current methods for approximating data between undisturbed satellite scans do not take weather data into account. We developed a novel approach to predict vegetation indices such as NDVI, EVI, NDWI, LAI, and FAPAR using multispectral satellite and weather data. Based on this model, transfer learning was introduced to train a grassland yield model. We compared artificial neural network architectures for predicting vegetation indices and grassland yields, including a multi-task formulation to additionally classify the crop types. The training samples for biomass prediction (n = 292) were collected in 2021. The crop prediction in the grassland crop category has an accuracy of 47.6%. The prediction of the vegetation indices and rgb values for three different time periods, ranging from 0 to 20 days after the last satellite scan, was done. The prediction of the leaf area index, for example, achieves a Pearson correlation of \(r=0.904\) and a mean absolute error (mae) of \(0.324 \frac{\text{m}^2}{\text{m}^2}\) for the period from 10 to 20 days from the latest satellite image. Finally, the Pearson correlation of the grassland fresh mass yield prediction was \(r=0.891\) with \(mae=1.245 \frac{\text{kg}}{\text{m}^2}.\)