Comparison and assessment of crop yield estimation from satellite-derived vegetation indices, solar-induced chlorophyll fluorescence, and evapotranspiration in southern Sweden

Meeting the food demand for the rising global population with current agricultural resources is a great challenge for the 21^st century. Accurate crop yield estimation is crucial for food security planning. Traditional ground field measurements can be time-consuming and costly, which has limitations for providing temporally consistent yield information in large areas. During the past decades, satellite-based observations have become an important input for crop yield estimation, since they can capture eco-physiological conditions on the ground consistently and frequently over extensive areas.

Three main independent satellite-based variables for estimating agriculture production can be summarized as: (1) vegetation indices (VIs); (2) biophysical variables; and (3) abiotic environmental factors. NDVI has been widely used in estimating agricultural production. The recently developed Plant Phenology Index (PPI) is shown highly related to vegetation productivity. PPI does not suffer from saturation effects in dense vegetation, and hence, has the potential to address the underestimation problems of crop productivity using other indices. Solar-Induced Fluorescence (SIF) is a direct proxy of plant photosynthesis, which is closely linked to crop yield. Furthermore, satellite-derived evapotranspiration (ET) integrates the effects of multiple environmental factors (e.g., precipitation, temperature, and wind speed) with soil moisture conditions, which has been demonstrated as an essential variable in crop growth monitoring.

However, the performance of these satellite-based datasets for crop yield estimation in Sweden remains insufficiently explored, which motivates us to conduct this study. We will investigate relationships between four satellite-derived variables (i.e., NDVI, PPI, SIF, and ET) and ground-based crop yield data (e.g., wheat, barley, sugar beet) in Skåne county during 2003–2022. Both NDVI and PPI are derived using satellite imagery data from Landsat (16-day, 30 m temporal/spatial resolutions) and Sentinel-2 (~5-day, 10 m) missions. Contiguous SIF (4-day, 0.05°) is selected for providing long-term high-frequency data. The gridded ET dataset (8-day, 0.25°) is merged by multiple ET datasets (i.e., FLUXCOM, GLASS, PML-V2) based on the triple collocation method. The ground measurements of crop yield in multiple agricultural fields in Skåne are obtained from Statistics Sweden.

For methodology, all satellite-derived variables will be firstly harmonized with the crop type map. The ground-based crop yield data will be divided to the training and testing samples in terms of temporal periods and spatial distribution. The linear/nonlinear relationships between four satellite-based variables (both individually and in varying combinations) and crop yield data (training samples) will be explored with different models including machine learning methods. Then the testing samples will be used for independent validation to determine the best variables/variable combinations and models for crop yield estimation. Finally, we will estimate crop yield at the regional scale and analyze its temporal and spatial patterns.