Hyperspectral Remote Sensing for Winter Wheat Leaf Area Index Assessment in Precision Agriculture

Remote sensing provides temporal, spectral and spatial information covering a wide area. Therefore, it has great potential in offering a detailed quantitative determination of the leaf area index (LAI) and other crop parameters in precision agriculture. The spatially differentiated assessment of LAI is of utmost importance for enabling an adapted field management, with the aim of increasing yields and reducing costs at the same time. The scientific focus of this work was the investigation of the potential of hyperspectral remote sensing data of different spectral resolutions, which were acquired at different spatial scales, for a precise assessment of wheat LAI. For this reason, three research experiments were conducted: 1) a comparison of different empirical-statistical regression techniques and their capabilities for a robust LAI prediction; 2) a determination of the required spectral resolution and important spectral regions/bands for precise LAI assessment; and 3) an investigation of the influence of the ground sampling distance of remote sensing images on the quality of spatial LAI predictions. The first part of this thesis compared three empirical-statistical regression techniques – namely, partial least-squares regression (PLSR), support vector regression (SVR) and random forest regression (RFR) – and their achieved model qualities for the assessment of wheat LAI from field reflectance measurements. In this context, the two different validation techniques – leave-one-out cross-validation (cv) and independent validation (iv) – were applied for verifying the accuracy of the different empirical-statistical regression models. The results clearly showed that model performance markedly depends on the validation technique used to assess model accuracy. In the case of leave-one-out cross-validation, SVR provided the best results, while PLSR proved to be superior to SVR and RFR when independent validation was applied. In the second part of this thesis, the spectral characteristics of the hyperspectral airborne sensor aisaDUAL (98 spectral bands) and the upcoming hyperspectral satellite mission EnMAP (204 spectral bands) were investigated to show their capability regarding the precise determination of wheat LAI. Moreover, the feature selection algorithm RReliefF, combined with a randomized sampling approach, was applied to identify the spectral bands that were most sensitive to changes in LAI. The results demonstrated that only three spectral bands of aisaDUAL, as well as EnMAP, at specific locations within the investigated spectral range (400–2,500 nm) were necessary for an accurate LAI prediction. The third part of this thesis dealt with the influence of the spatial resolution of aisaDUAL (3 m) and simulated EnMAP (30 m) image data on the assessment of wheat LAI. While the ground sampling distance (GSD) of aisaDUAL allowed a robust regression model calibration and validation, LAI predictions based on simulated EnMAP image data led to poor results because of the distinct difference in size between the EnMAP pixels (900 m2) and the sampled field plots (0.25 m2) for which the LAI was measured. In order to enable a more precise determination of wheat LAI from EnMAP image data, the two different approaches of image aggregation and image fusion were examined. In this context, the fusion approach has proven to be the more suitable method, which allowed a more accurate LAI prediction compared to the results based on the EnMAP data with a GSD of 30 m. In summary, the findings of the research reported in this thesis demonstrated that the accuracy of spatial LAI predictions from remote sensing data depends on several factors. Besides the applied empirical-statistical retrieval- and validation method, the spatial and spectral characteristics of the used image data sets played an important role. With the forthcoming hyperspectral satellite missions (e.g., EnMAP, HyspIRI), the area-wide assessment of LAI and other crop parameters (e.g., biomass, chlorophyll content) will be strongly supported. The moderate spatial resolutions of these satellites systems, however, require a combined use with spatial higher resolution multi- or superspectral satellite data (e.g., RapidEye, Sentinel-2). This multisensoral approach offers great potential for the prompt identification of spatial variations in crop conditions on sub-field scale, which is a mandatory prerequisite for precision agricultural applications