Evaluation of leaf area index and dry matter predictions for crop growth modelling and yield estimation based on field reflectance measurements

Abstract

Leaf area index (LAI) and above ground biomass dry matter (DM) are key variables for crop growth monitoring and yield estimation. High prediction accuracies of these parameters are a vital prerequisite for sophisticated yield projections. The aim of the study was to examine the predictive ability of partial least squares regression (PLSR) for LAI and DM retrieval from hyperspectral (EnMAP), superspectral (Sentinel-2), and multispectral (Landsat 8, RapidEye) remote sensing data based on field reflectance measurements. Data was acquired from several crop types (wheat, rye, barley, rape, potato, sugar beet) during field campaigns in three different regions of Germany between the years 2011 and 2014. The field reflectance measurements were resampled to match the different spectral resolutions. Continuous reflectance and resampled data were transformed using five spec-tral pre-processing techniques. Continuous data were used for comparison and served as best case scenario. The predictive ability of the PLSR models for LAI and DM was examined with respect to the spectral resolution and the pre-processing techniques. To verify whether the composition of the data set had an effect on prediction quality, the entire data set (global) was divided in sub data sets (local) with respect to the region of acquisition, the year of acquisition and the crop type. Statistical models of the local data sets were compared with those based on the global data set. Generally, models were assessed with two validation strategies. R2 of the global PLSR models based on continuous field reflectance measurements and independent validation varied from 0.74 to 0.79 (LAI), and from 0.76 to 0.87 (DM). Root mean square error ranged between 0.70 and 0.74 m2 m-2, and between 1.64 and 2.56 t ha-1, respectively. There was no pre-processing method which consistently improved model performance. However, results pointed out that the technique should be chosen with respect to the sensor and the parameter of interest. Models based on hyperspectral information performed generally best. Prediction error increased with the superspectral sensor configuration by only 3% for LAI, and 16% for DM. Multispectral sensor configurations caused the prediction error to rise by up to 22% and 54%, respectively. A stratification into local data sets according to date of acquisition, sampling region and crop type partially increased the prediction performance. Cross-validation yielded higher prediction errors than independent validation in most cases.