Hyperspectral Signal Unmixing Techniques for Site Specific Plant State Monitoring

In precision farming, field management is based on observing and responding to intra-field variations. Hyperspectral remote sensing has shown great potential in providing timely and accurate information on the spatial variability of field- and plant conditions. However, due to the discontinuous open canopies typical of most (perennial) cropping systems, the size of the image pixels will exceed, in many cases, the size of the objects of interest. The reflectance signal of a pixel is thus the integrated result of spectral contributions of both the crop and non-crop components (i.e. soils, weeds and shadows) building up a pixel footprint. As a consequence, image interpretation and the extraction of the biophysical parameters from the measured hyperspectral signature is hampered. Accurate site-specific monitoring of the crop thus requires removing all the undesired background effects from a measured mixed pixel, resulting in a purified vegetation signature, which can be used to derive the desired information regarding the plant. To this end, Signal Unmixing (SU) methodologies are presented in this dissertation, deriving the pure spectral signature of the crop component on a per-pixel basis. The basis is Multiple Endmember Spectral Mixture Analysis (MESMA). Spectral libraries or Look-up Tables (LUTs) are used, i.e. a collection of spectra representing the possible reflectance values of the different endmembers. The MESMA algorithm is an iterative process that selects endmember combinations from spectral libraries, and the combination which results in the lowest reconstruction error of the mixed signal is selected as the best representation of the components present within the pixel. The selected signature can on its turn be used to derive the desired information regarding the crop s vigour status.In the first part of this work, solutions for the major bottlenecks of the MESMA methodology are presented. As the accuracy of MESMA is determined by the adequacy of the spectral library, it is crucial that the spectral library used for the unmixing of the image is representative of all endmembers present. An extensive tree LUT was thus created using a radiative transfer model, incorporating a high level of detail in the spectral signatures. However, variability in endmembers may lead to more than one pure spectrum combination resulting in the same mixture spectrum, a problem commonly referred to as ill-posedness. In Chapter 3, the integration of in situ measured soil moisture content into the SU model is therefore proposed to provide an estimation of the soil signature, as such reducing the number of possible solutions. This integration leads to a better extraction of the vegetation spectra, which on its turn results in an improved estimation of the trees vigour. Finally, the large size of the LUTs restricts the computational efficiency of the SU model. Incorporating geometric unmixing principles into MESMA enables a more efficient evaluation of all the different endmember combinations (Chapter 4). Whereas the traditional MESMA explores alldifferent endmember combinations separately, and selects the most appropriate combination as a final step, our approach selects the best endmember combination prior to unmixing, as such increasing the computational efficiency of MESMA.In addition to MESMA, two other unmixing methodologies are presented. Alternating Least Squares (ALS) unmixing is proposed as a Signal Unmixing methodology in Chapter 5. While MESMA requires extensive LUTs from which the most representative signal can be selected, ALS only needs an initial estimate of the spectral signature of each of the components present in the mixed pixel. This initial estimate is further optimised by ALS, and the pure spectral signature of the tree can thus be extracted from the mixed pixel signal. All the previous methodologies are tested on mixtures comprised of trees, soil and shadows. As the high spectral similarity between the trees and weeds hampers an accurate extraction of the tree signature, the performance of shape-based unmixing for separating spectrally similar endmembers is evaluated in Chapter 6. These insights can then be used to develop shape-based unmixing further into an SU model. Overall, this work provides a conceptual framework for the operational implementation of SU methodologies in a precision farming context. New methodologies are presented to extract the pure tree signature from hyperspectral mixed pixels, as well as to improve the computational efficiency and the accuracy of these methods. With the extracted tree signatures, an improved monitoring of the trees condition is achieved. SU thus provides a new avenue to explore the use of hyperspectral imagery in a precision farming context.nrpages: 151status: publishe