Monitoring and modeling of citrus plant production systems by integrating hyperspectral remote sensing data and in situ data in a mathematical framework

摘要

Plant production systems are governed by site management and by biotic and abiotic factors. For annual crops, many of these have been described, modeled and used as a basis for production steering. Existing biophysical production models however lack components for the correct modeling of fruit crops, such as integrated submodels for flowering and fruit growth, adaptation and carbohydrate allocation. Dynamic inputs to these models are being delivered by meteorological records and - increasingly - by remote sensing due to its high spatial and temporal resolution. Specifically hyperspectral satellite data sets offer a high added value as they are capable of detecting a large number of relevant biophysical variables, such as leaf area index, chlorophyll content, soil and crown water content, pigment ratios and biomass production. The major scientific challenge is the robust and unbiased extraction of these variables and their subsequent assimilation in production models.The first objective of this dissertation, throughout which citrus is used as a model crop, is the development of a bottom-up approach to model, at the different scales, the physiological, structural and optical processes leading to changes in signals as retrieved by hyperspectral satellites. The first step herein encompasses a description of the critical internal processes that lead to detectable changes in biochemical components with a focus on the carbohydrate flows (chapter 2). In the second step, a dorsiventral leaf model (DLM) is developed (chapter 3) capable of accurately simulating optical properties (reflectance and transmittance) of leaves with known structural properties and biochemical composition. Contrary to existing isolateral leaf models, DLM is adapted to model dorsiventral leaves, typical for dicot species such as citrus. The next step in the up-scaling process establishes the relation between optical properties at the leaf level and those at the canopy (tree crown) level. Using 3D virtualization techniques (ray-tracing), in chapter 4 a model of an existing citrus orchard was constructed and validated using field data. This virtual environment enables simulations of hyperspectral sensor measurements with high accuracy. Innovative aspects are the use of calibrated 3D tree models with an explicit description of the geometry of all leaves, twigs and stems and realistic simulation of both diffuse and direct sunlight. This virtual environment has subsequently been used as a reference for the (in)validation of different assumptions commonly made by simpler but faster algorithmic canopy reflectance models. Tested assumptions concern the shape, distribution and glossy reflections of leaves, the properties of the incident light and the row structure of orchards. This relation between leaf and tree has not only been studied using simulations, but has also been investigated using a hyperspectral time series of field measurements in a commercial citrus orchard (chapter 5). Changes in leaf, fruit and crown spectra throughout consecutive growth seasons could be explainedby interpreting within-canopy mixtures of canopy components (leaf and fruit types) in different phenological stages (flowering, fruit set and growth and leaf drop), stress (sunburn) and management actions (pruning, irrigation and harvest). The last step in the up-scaling process, from crown to satellite level, was made in chapter 7 in which the impact of shadow, viewing geometry and pixel size were investigated.The second objective in this research was the search for more robust data extraction methods that follow the inverse path: from satellite measurements to biophysical model variables. Chapter 3 introduces an improved model inversion strategy for DLM that allows leaf biochemistry (chlorophyll, carotenoids, dry matter and water) to be determined with higher accuracy. Additional statistical model building (chapter 2) reveals that spectral contact measurements can be used to determine leaf starch and soluble sugar concentrations. For measurements at the crown level, in chapter 6, a new measurement protocol is developed that enables time series collection under variable environmental conditions. This substantially improves measurement opportunities as compared to existing field protocols that demand a cloud free sky. Finally, using the virtualization techniques from chapter 4, chapter 7 finds optimal viewing angles for off-nadir satellite imagery in row plantations that minimize the interfering influence of soil and weeds on the canopy spectrum.The simulations at the different scale levels of this research (biophysical process, leaf, crown and satellite) are some of the building blocks of a framework. Within this framework, remote sensing measurements capable of monitoring the production process of fruit crops can be simulated in a reliable and physically/physiologically based way. The insights that such an approach has delivered were used to make existing data extraction methods more robust and more accurate. Some parts of this research resulted in technologies that can be employed in an operational context, such as a fast assessment of leaf biochemistry and field protocols for canopy reflectance spectra. The full bottom-up approach as envisaged here, however, requires a substantial amount of sustained fundamental and applied research.