Crop reflectance operational models for agriculture (CROMA)

3D structure model of canopy were developed, so as to simulate homogeneous and clump canopies, as well as realistic maize canopies. The characteristics of the canopies change as a function of the crop stage. The reflectance of these canopies is then computed by the ray-tracing model PARCINOPY (Chelle, 1997). Such 3D models provide highly accurate radiative transfer models, which can serve as references to evaluate other models.

The description of the contribution of the multiple scattering inside the leaf canopy is an important aspect of the modelling of canopy reflectance. In the near infrared, leaves absorb less than 10 percents of the incident radiation. Therefore, multiple scattering may provide the dominant contribution to the reflectance in this part of the spectrum. However, the effects of the multiple scattering are difficult to model. The methods that have been developed up to now suppose that the scattering medium is isotropic. This hypothesis does not hold in the case of leaf canopies. The N+2 stream method was developed by Verhoef (1998) to provide an accurate description of the radiative transfer in non-isotropic media. The SAIL model (Verhoef, 1984, 1985) was improved so as to take into account the N+2 stream method for the description of the multiple scattering in homogeneous canopies.

The a priori distribution of canopy architecture variables for maize crops was derived from several experiments. Some distributions were derived directly from ground measurements or from 3D crop architecture models. Others were derived from the inversion of radiative transfer models. This last approach was preferred, since it keeps a high degree of consistency on the actual meaning of the variables with respect to the radiative model transfer assumptions.

The impact of sub-pixel non-homogeneities of the maize canopy on the reflectance was studied by comparing four crop reflectance models. The influence of the vertical heterogeneity of the canopy was studied by comparing the model of homogeneous canopies MCRM (Kuusk, 1995) and the two-layer crop reflectance model (Kuusk, 2001). The influence of the canopy patchiness was analysed by comparing simulations of MCRM and the forest reflectance model FRT (Kuusk and Nilson, 2000). The row effects were assessed from an improved version of MCRM2, which was adapted to row-structured canopies. The study demonstrated that the use of a simple homogeneous model to estimate the biophysical variables of heterogeneous canopies might lead to very erroneous results.

The SAIL model (Verhoef, 1984, 1985) was cross-compared to the model MCRM2-Row developed from MCRM2 (Kuusk, 2001) so as to account for the row effects. The results show that row effects are noticeable as long as the rows are not contiguous. They mainly lead to a large underestimation (~ 55 %) of the LAI and to an overestimation of the CHL at low LAI. They strongly depend on the viewing conditions with respect to the row orientation.

Several methods (Quasi-Newton, Look-Up Table and Neural Network) were compared on the estimation of key crop biophysical variables by inverting a model of the crop reflectance. The study focused on the retrieval of the LAI and of the leaf chlorophyll content. The study used both synthetic and actual remote sensing data. A particular methodology was developed so as to include an optimisation of the prior information within the inversion process.

Several inversion techniques were compared based on the estimation of maize crop key biophysical variables (LAI, leaf chlorophyll content) by inverting a crop reflectance model. The Look-Up-Table technique and the neural network were compared on a single data set. Great attention was paid to the way prior information is introduced in these two techniques. Although neural network should theoretically provide better performances when the training data set is properly constructed. Results show that the performances of the two techniques are rather similar. This also shows the limitations of results when derived from experiments because the measured canopy variables are not always equivalent to the variables estimated by radiative transfer models.

The mechanistic crop functioning model PlantSys version 1.0 was forced with temporal remote sensing estimates of the LAI and of the leaf nitrogen content through the graphical user interface MoReSens version 1.0. This allowed quantifying the impact of actual uncertainty on the remote sensing estimates on the estimations of the final maize grain yield.

Simple crop functioning models were developed, which are empirical models describing the evolution of the crop LAI as a function of the cumulated mean air temperature from sowing. Existing models published in the literature as well as new models based on botanic rules were compared. The model of Baret (1986) was chosen, since it displayed the best fit to the CROMA experimental database. It was demonstrated that the coupling of this model to a crop reflectance model significantly improve the LAI estimates.

Inverting SAIL with the simulations computed with SAIL+XBS, supposing that the soil is shaded, leads to a strong overestimation of the LAI (~ 100 %) and of the CHL (~3 mg/cm2) at low LAI. Hence, SAIL+shaded XBS accounts for similar effects as MCRM2-ROW when the soil between the rows is shaded. It was demonstrated that SAIL+enlightened XBS accounts for the same effects as MCRM2-ROW when the soil between the rows is enlightened. Inverting SAIL+XBS by itself demonstrated that xbs can be estimated only when the bare soil is shaded. In this case, estimating xbs allows to strongly reduce the bias on the LAI, and at a lesser extent the bias on the CHL.

"Forcing methods" refer to the way model variable values are adjusted to externally determined variable values. In the present case, the model is PlantSys version 1.0, which is a complex mechanistic crop functioning model. External measurements can be done in classical ways or, as in this case, by remote sensing signals to obtain biophysical variable values. Several forcing methods were developed, so as to calibrate the model during run-time and to assess the impact of external factors on the crop development.

The SAILH (Verhoef, 1998) was developed based on the SAIL model developed by Verhoef (1984, 1985). The former SAIL model does not describe the hot spot effect, whereas SAILH simulates the hot spot effect. The description of the hot spot effect is based on the theory developed by Kuusk (1985).

Six empirical models of the temporal evolution of the LAI were compared. Four of these models were extracted from the literature, among which the model developed by Baret (1986). Two others are new models developed based on botanic rules. The six models were compared based on their ability to fit the maize LAI cycles of the experimental database of the CROMA project. This database is representative of various maize varieties, cultural practices and climatic conditions. The model of Baret (1986) proved to be the most appropriate model to fit this database.

The cross-analysis of the models SAIL and SAIL++ showed that the hemispherical approximation leads to a systematic but weak overestimation of the LAI. The impact increases from LAI = 0 to LAI = 3, up to a bias of 5 %, and decreases from LAI = 5. The diffuse stream may first increase with the density of the canopy until it reaches a density within which the various diffuse flux components tend to homogenise. The hemispherical approximation also leads to a slight and stable overestimation of the CHL content (~+3.5 mg/cm2) at low LAI (LAI ~ 1).

The internal consistency of the SAIL++ model (Verhoef, 1998) was checked, and it was compared to other models. In particular, SAIL++ was compared to the former SAIL model (Verhoef, 1984, 1985). It was shown that the differences between the two models are rather modest; they are only significant in the near infrared, at high LAI and erectophile leaf angle distribution. However, SAIL++ is more precise, more stable and only moderately slower than SAILH. Therefore, its use is recommended for simulating canopy reflectance spectra at hyperspectral resolution.

A common source of sub-pixel canopy heterogeneity in the agricultural field is the seeding along rows. The rows are especially visible at the beginning of the growing season in maize fields. A new canopy reflectance model was developed, which accounts for row structures. Plant rows are described as long lines with soil between lines, which dimension changes as a function of the LAI, until the rows merge. The rows are "filled" with homogeneous vegetation, whose parameters depend on time. The bi-directional reflectance of the canopy is computed using the two-layer crop reflectance model MCRM2 model (Kuusk, 2001). MCRM2 was completed with geometrical optics related to row structures to estimate the proportion of shadowed and enlightened soil and plant canopy in the field of view.

Three reflectance models of homogeneous canopies were compared, which are SAIL (Verhoef, 1984, 1985), MCRM2 (Kuusk, 2001) and the ray-tracing model PARCINOPY coupled to homogeneous 3D canopy architectures. The results underscore large discrepancies between them. The differences between SAIL and MCRM2 would be explained by the use of different leaf angle distributions. All the discrepancies were not totally explained, and there is still no consensus on the most appropriate modelling approach.

The new empirical model of the LAI temporal evolution (EvolLAI) was inverted with more than 300 LAI cycles measured within the framework of the experimental CROMA database. This database is representative of various maize varieties, cultural practices and climatic conditions. Globally, EvolLAI correctly fitted the LAI cycles of the CROMA database, in particular when the number of measurements per cycle was high (~10) and the measurement errors were low (~5 %). However, EvolLAI displays more difficulty to fit the cycles having a priori noisy measurements, too little measurements per cycle and the cycles that display accidents in the LAI evolution.

The forcing methods developed within CROMA were applied to calibrate the mechanistic crop functioning model PlantSys version 1.0 with actual remote sensing estimates of the crop biophysical variables. Some of the remote sensing data are only time-dependent, others consist of 2D maps of the crop fields. They come from the experimental database of the CROMA project. This study allowed studying the impact of the uncertainty of the remote sensing estimates on the evaluation of the maize grain yield.

The inversion performance of the new empirical model of the LAI temporal evolution (EvolLAI) was checked. Several synthetic "measured" LAI cycles were generated with the model, which can be distinguished by their "measurement" noise. The model was dynamically inverted with these datasets. The result show that this model allow to estimate the duration of the maize development phases, its growth and senescence rates, the maximum LAI and the sowing date.

The forest reflectance model FRT (Kuusk and Nilson, 2001) was adapted to simulate patchy maize canopies. The resulting model was cross-compared to the SAIL model (Verhoef, 1984, 1985). The results show that the patchiness effects are identical to the row effects, before the patches merge (LAI £ 2). Moreover, they are perceptible at high LAI, where they induce an underestimation of the LAI up to 20 % and an over- or under-estimation of the CHL up to 8mg/cm2. These effects may be explained by the roughness of the upper surface of the canopy and by the shade effects between plants.

A new reflectance model is being developed based on a 3D structure model of the crop canopy, which takes into account the canopy clumping. Three levels of clumpiness are considered: the sowing of the stems in rows or randomly, the attachment of leaves to stems and the particular disposition of leaves, which prevents their overlapping. Moreover, the model takes into account varying leaf sizes and shapes. The reflectance of such canopies will be computed with the ray-tracing model PARCINOPY (Chelle, 1997).

The a priori distribution of canopy architecture variables for maize crops was derived from several experiments. Some distributions were derived directly from ground measurements or from 3D crop architecture models. Others were derived from the inversion of radiative transfer models. The results proved to strongly depend on the method used for their derivation. Methods that are based on the use of the radiative transfer model for the derivation of such information have to be preferred because they keep a high degree of consistency on the actual meaning of the variables with respect to the radiative model transfer assumptions. Great care has to be taken when calibrating the prior information by confronting variables derived from radiative model inversion to the actual values measured in the field.

The PlantSys model version 1.0 (Jongschaap, in press) was developed within the framework of CROMA based on the Rotask simulation model (Jongschaap, 1996) and on the crop growth and development approach of Yin et al. (2001). It is a dynamic model for fallow land, single crop and crop rotation systems, including soil organic matter fluxes calculated in time-steps of a single day. The model accounts for farmers' interventions like tillage, sowing, fertiliser application and harvesting. It can be forced with ground measurements or with remote sensing data through the graphical user interface MoReSens version 1.0.

The objective was to study the influence of the number and date of the LAI observations over the retrieval of the parameters of the empirical model of the temporal LAI developed by Baret (1986). This was investigated by (i) simulating a dataset with various LAI courses, then by (ii) selecting observation dates with different temporal resolutions, which are consistent with the revisit frequency of a dedicated sensor and with the cloud cover statistics and by (iii) Computing the global relative RMSE for sets of 10 or 20 equally spaced observations. The results demonstrated that 3 to 4 observations are necessary to estimate the parameters of the growth and senescent periods to ensure a good and reliable fit.

The objective was to study the influence of the number and date of the LAI observations over the retrieval of the parameters of the empirical model of the temporal LAI developed by Baret (1986). This was investigated by (i) simulating a dataset with various LAI courses, then by (ii) selecting observation dates with different temporal resolutions, which are consistent with the revisit frequency of a dedicated sensor and with the cloud cover statistics and by (iii) Computing the global relative RMSE for sets of 10 or 20 equally spaced observations. The results demonstrated that constraining the whole set of parameters requires 5 to 8 observations well spread out over the whole LAI cycle.

XMCRM2 is a graphical front-end for the two-layer crop reflectance model MCRM2 (Kuusk, 2001), which should run in all Unix environments. It provides a convenient way for creating input files for MCRM2 and for viewing the results. Model parameters are inserted in the main panel; configuration parameters are specified in another preference panel. The computed reflectance spectra or the angular distribution of the reflectance is plotted on distinct graph windows.

The SAIL+l model was cross-compared to the SAIL model. Inverting SAIL with the simulations of SAIL+l leaded to an underestimation of the LAI (~- 10 %) and to an overestimation of the CHL (~ 8 mg/cm2) as long as the leaves are aggregated. Hence, SAIL+l accounts for the patchiness effects at low LAI. Inverting SAIL+l with itself demonstrated that the aggregation parameter l couldn't be estimated from remote sensing observations.

The SAIL variant RowSAIL was cross-compared to the former SAIL model. Inverting SAIL with the simulations performed with RowSAIL leads to a significant underestimation of the LAI (20-30 %) and to a significant overestimation of the CHL (5-15 %) before the merging of the rows. The impact is the strongest when the sun and the row directions are in the same plan. The equivalence of SAIL and RowSAIL when the canopy is homogeneous requires introducing a description of the hot spot effect, which will be the object of future work. Hence, it was demonstrated that RowSAIL accounts for the same row effects as MCRM2-ROW.

The FRT model (Kuusk and Nilson) allows simulating the reflectance of patchy canopies, where the plants may be randomly or regularly positioned. Their position is generated by an electrostatic algorithm (Gusakov and Fradkin, 1990). Each plant is described by a stem and by a crown envelop, which is made of two halves of ellipsoids. A tool was developed to visualise such canopies. It is based on the software package POV-Ray V. 3.1 by Persistence of Vision Development Team. The software requires an input file, which describes the plants and lighting in the scene. This input file is generated by the FRT model.

Exploitable results

Share this page

Download