By Anne van Gijsel
Summary
This thesis deals with the retrieval of canopy variables from remotely sensed data with the main objective to detect iron chlorosis in olive orchards. Iron chlorosis leads to a reduced uptake and distribution of iron in plants, which in turn hampers the creation of chlorophyll and thus the photosynthesis and tree’s production. Iron chlorosis can therefore be identified through monitoring for trees with a decreased chlorophyll a + b content (CAB). As changes in CAB have a pronounced effect on the leaf reflectance, we can monitor these CAB by means of remote sensing. CASI images were available for the three study sites at two resolutions: CASISPATIAL having 8 bands with 1 meter spatial resolution and CASISPECTRAL having 72 bands with a spatial resolution of 4 meter.
Two retrieval strategies using the CASI observations have been studied: the inversion of a coupled radiative transfer model (the leaf model PROSPECT + the canopy reflectance model FLIGHT) by means of neural networks and the application of statistical relations between canopy reflectance and the chlorophyll content.
The leaf reflectance and transmittance were modelled with PROSPECT and resampled to match the CASI bands. These resampled leaf reflectance and transmittance were used together with variables describing the canopy (leaf area index (LAI), tree dimensions, leaf angle distribution (LAD) etc.), scene conditions (fCOVER, background/soil reflectance etc.) and viewing geometry as an input for the FLIGHT model to obtain the canopy reflectance for that scene. Variation was introduced in the inputs for PROSPECT (N, CM and CAB) and for FLIGHT (LAI, fCOVER). For the CASISPECTRAL simulations, we have also created three classes of soil brightness to be able to study the effect of different background signals on the retrieval methods. The simulations where N and CM have a constant value (with all other variables being non-constant) are referred to as PF-set 1 and the simulations with variation in N and CM¬ will be referred to as PF-set 2.
Two different approaches using neural networks (NN) were tested for the retrieval of canopy variables (CV) from CASISPATIAL. The first approach consisted of the ‘classical’ inversion where the NN were estimating the CV from the canopy reflectance. In the second approach, the NN were doing the reverse: the CV were obtained by inverting this NN estimating the canopy reflectance from a first guess for the CV. The estimated CV were updated using an optimalisation algorithm minimising the difference between the estimated reflectance and the observed reflectance. It was found that the classical NN has a better performance, as the second method had greater uncertainties due the combination of the imperfect training and inversion of the NN. For the classical NN it was furthermore observed that the use of a-priori information improved the estimation of CAB from PF-set 2, especially knowledge of N and CM proved to be important. Fixing these two parameters (PF-set 1) resulted in an RMSE of 1.98 μg/cm2. Reducing the number of input bands from 7 (band 6 was not used as it corresponds to an oxygen absorption band) to 2 greatly increased the error and indicates that indices using only these 2 bands may have strong limitations. However, the estimation of fCOVER and LAI in addition to CAB from PF-set 1 resulted in an RMSE for CAB of 2.57 μg/cm2, for fCOVER it was 3.91% and for LAI 0.52 m2/m2. We concluded that theoretically it is well possible to retrieve multiple variables simultaneously using an inversion of a leaf+canopy model, provided that important crop characteristics (here N and CM) are known or can be considered constant.
In parallel to the NN approach, we have developed relations between vegetation indices and the chlorophyll content. It was found to be most important to study the effect of other non-constant factors, such as LAI or the soil brightness on the behaviour of an index, as changes in these other factors may induce a trend that could be confused with a change in the main variable under study (the chlorophyll content in this study). The best performing indices in terms of RMSE were the approximated MTCI and MERIS red edge indices and the developed CASI red edge index. These indices were found to have a fairly constant performance over all tested spatial resolutions (1, 4, 32 and 300 m).
Finally we have described the performance of the trained NN and the derived relations between VI and CAB when these were applied to the real CASI imagery. The results were very poor. We identified a mismatch between the simulated reflectances and the measured reflectances similar to (Tan et al. 2005). Possible sources of error are incorrect parameterisation of the PROSPECT+FLIGHT models such as approximations made on the dimensions of the olive trees or an unrepresentative soil spectrum, limitations of the FLIGHT model at high spatial resolutions, or calibration artefacts in the CASI images. It has shown that careful validation of the results should be done after application of model based relations.
Tan, B., Hu, J., Huang, D., Yang, W., Zhang, P., Shabanov, N.V., Knyazikhin, Y., Nemani, R.R., & Myneni, R.B. (2005). "Assessment of the broadleaf crops leaf area index product from the Terra MODIS instrument". Agricultural and Forest Meteorology, 135(1-4), 124-134