Vegetation is an important component of the Earth’s biosphere and therefore plays a crucial role in the carbon exchange of terrestrial ecosystems. Vegetation variables, such as leaf area index (LAI) and leaf chlorophyll content (Cab), can be monitored at global scale using remote sensing (RS). There are two main categories of approaches for estimating the vegetation variables from RS data: empirical and physically-based approaches. Physically-based approaches are more widely applicable because they rely on radiative transfer (RT) models, which can be adapted to the observation conditions and to the observed vegetation. For estimating the vegetation variables, however, the RT model has to be inverted, and this inversion is usually an ill-posed and under-determined problem. Several regularization methods have been proposed to allow finding stable and unique solutions: model coupling, using multi-angular data, using a priori information, as well as applying spatial or temporal constraints.
Traditionally, radiance data measured at top-of the atmosphere (TOA) are pre-processed to top-of-canopy (TOC) reflectances. Corrections for atmospheric effects, and, if needed, for adjacency, directional, or topographic effects are usually applied sequentially and independently. Physically, however, these effects are inter-related, and each correction introduces errors. These errors propagate to the TOC reflectance data, which are used to invert the canopy RT model. The performance of the TOC approach is therefore limited by the errors introduced in the data during the pre-processing steps.
This thesis proposes to minimize these errors by directly using measured TOA radiance data. In such a TOA approach, the atmospheric RT model, which is normally inverted to perform the atmospheric correction, is coupled to the canopy RT model. The coupled canopy-atmosphere model is inverted directly using the measured radiance data. Adjacency, directional and topographic effects can then be included in the coupled RT model. The same regularization methods as used for TOC approaches can be applied to obtain stable and unique estimates. The TOA approach was tested using four case studies based on mono-temporal data.
A) The performance of the TOA approach was compared to a TOC approach for three Norway spruce stands in the Czech Republic, using near-nadir Compact High Resolution Imaging Spectrometer (CHRIS) data. The coupled model included canopy directional effects and simulated the CHRIS radiance data with similar accuracy as the canopy model simulated the atmospherically-corrected CHRIS data. Local sensitivity analyses showed that the atmospheric parameters had much less influence on the simulations than the vegetation parameters, and that the sensitivity profiles of the latter were very similar for both TOC and TOA approaches. The dimensionality of the estimation problem was evaluated to be 3 for both approaches. Canopy cover (Cv), fraction of bark material (fB), Cab, and leaf dry matter content (Cdm) were estimated using look-up tables (LUT) with similar accuracy with both approaches.
B) Regularization using multi-angular data was tested for the TOA approach, using four angular CHRIS datasets, for the same three stands as used in A). The coupled model provided good simulations for all angles. The dimensionality increased from 3 to 6 when using all four angles. Two LUTs were built for each stand: a 4-variable LUT with fB, Cv, Cdm, and Cab, and a 7-variable LUT where leaf brown pigment concentration (Cs), dissociation factor (D), and tree shape factor (Zeta) were added. The results did not fully match the expectation that the more angles used, the more accurate the estimates become. Although their exploitation remains challenging, multi-angular data have higher potential than mono-angular data at TOA level.
C) A Bayesian object-based approach was developed and tested on at-sensor Airborne Prism Experiment (APEX) radiance data for an agricultural area in Switzerland. This approach consists of two steps. First, up to six variables were estimated for each crop field object using a Bayesian optimization algorithm, using a priori information. Second, a LUT was built for each object with only LAI and Cab as free variables, thus spatially constraining the values of all other variables to the values obtained in the first step. The Bayesian object-based approach estimated LAI more accurately than a LUT with a Bayesian cost function approach. This case study relied on extensive field data allowing defining the objects and a priori data.
D) The Bayesian object-based approach proposed in C) was applied to a simulated TOA Sentinel-2 scene, covering the area around Zurich, Switzerland. The simulated scene was mosaicked using seven APEX flight lines, which allowed including all spatial and spectral characteristics of Sentinel-2. Automatic multi-resolution segmentation and classification of the vegetated objects in four levels of brightness in the visible domain enabled defining the objects and a priori data without field data, allowing successful implementation of the Bayesian object-based approach.
The research conducted in this thesis contributes to the improvement of the use of regularization methods in ill-posed RT model inversions. Three major areas were identified for further research: 1) inclusion of adjacency and topography effects in the coupled model, 2) addition of temporal constraints in the inversion, and 3) better inclusion of observation and model uncertainties in the cost function. The TOA approach proposed here will facilitate the exploitation of multi-angular, multi-temporal and multi-sensor data, leading to more accurate RS vegetation products. These higher quality products will support many vegetation-related applications.