Simulation of satellite reflectance data using high-frequency ground based hyperspectral canopy measurements for in-season estimation of grain yield and grain nitrogen status in winter wheat

摘要

Remote sensing allows the assessment of biomass and grain yield of cereals. An enhanced potential to detect specific traits is offered by proximal hyperspectral sensing, in contrast the potential of satellite remote sensing might still be leveraged up seen the degree of spatial and temporal resolution. Consequently, hyperspectral satellite data are still not readily available for most agricultural applications and ground truth validation done at high frequency throughout the season remains scarce. Here, we simulate multispectral satellite data through resampling ground based hyperspectral reflection (400-1000 nm). Spectral data were collected at 3 nm resolution at high temporal frequency during 5-14 growth stages over three winter wheat growing seasons for investigating the influence of year and seasonality. The field experiment comprised 24 different genotypes, which varied in morphology and phenology, grown at varying nitrogen application and additionally by varying the sowing time and fungicide intensity during the third season. Ground based reflectance data were resampled to fit the spectral resolution of the satellite sensors Landsat-8, Quickbird, RapidEye, WorldView-2 and Sentinel-2. The resulting spectral bands were used for calculating all possible normalized difference vegetation indices, which were correlated to grain yield, grain N uptake and grain N concentration. The index performance depended substantially on the growth stages and years. For grain yield, maximum linear relationships (R-2) obtained from hyperspectral sensing over the season ranged from 0.65 in 2017 to 0.88 in 2015 compared to 0.40 to 0.79 for the best simulated multispectral sensor, Sentinel-2. In most cases, indices performed better for grain N uptake and less well for grain N concentration. Typically, correlations peaked at early to medium grain filling but decreased around ear emergence. Index performance from multispectral compared to hyperspectral data decreased over time during grain filling. The sensor ranking remained consistent with Sentinel-2 followed by Worldview-2 and RapidEye clearly outperforming the other sensors. Advantages are attributed to the red edge band for N-related traits and the better coverage of the NIR range between 800 and 1000 nm by the Sentinel-2. The results can possibly be extrapolated to the application of UAV and satellite sensing by elucidating optimized measurement stages and enhancing spectral properties of sensors.