The growing availability of high-resolution satellite image series offers
new opportunities in agro-hydrological research and modeling. We
investigated the possibilities offered for improving crop-growth dynamic
simulation with the distributed agro-hydrological model: topography-based
nitrogen transfer and transformation (TNT2). We used a leaf area index (LAI) map series
derived from 105 Formosat-2 (F2) images covering the period 2006–2010. The
TNT2 model (Beaujouan et al., 2002), calibrated against discharge and
in-stream nitrate fluxes for the period 1985–2001, was tested on the
2005–2010 data set (climate, land use, agricultural practices, and discharge
and nitrate fluxes at the outlet). Data from the first year (2005) were used
to initialize the hydrological model. A priori agricultural practices obtained from
an extensive field survey, such as seeding date, crop cultivar, and amount
of fertilizer, were used as input variables. Continuous values of LAI as a
function of cumulative daily temperature were obtained at the crop-field
level by fitting a double logistic equation against discrete
satellite-derived LAI. Model predictions of LAI dynamics using the a priori input
parameters displayed temporal shifts from those observed LAI profiles that are
irregularly distributed in space (between field crops) and time (between
years). By resetting the seeding date at the crop-field level, we have
developed an optimization method designed to efficiently minimize this
temporal shift and better fit the crop growth against both the spatial
observations and crop production. This optimization of simulated LAI
has a negligible impact on water budgets at the catchment scale (1 mm yr
Agro-hydrological modeling was first developed and applied to study the qualitative and quantitative impacts of agriculture on water resources in cropped land areas (Arnold et al., 1993, 1998; Breuer et al., 2008; Engel et al., 1993; Galloway et al., 2003; Leonard et al., 1987; Refsgaard et al., 1999; Whitehead et al., 1998). Hydrology and crop models were coupled to take into account the influences of both hydrological settings and agricultural practices on the water and nutrient cycle at the agricultural catchment scale: CWSS (Reiche, 1994), DAISY/MIKE-SHE (Refsgaard et al., 1999), NMS (Lunn et al., 1996), SWAT (Arnold et al., 1998), INCA (Whitehead et al., 1998), SHETRAN (Birkinshaw and Ewen, 2000), TNT2 (Beaujouan et al., 2002), DNMT (Liu et al., 2005), and STICS-MODCOU-NEWSAM (Ledoux et al., 2007). Subsequently these approaches have become widely used: hundreds of publications, among which the SWAT model is probably the most popular, report their use in studying the impact of (1) agriculture in terms of stream-water quality, e.g., nitrate contamination (Durand, 2004; Ferrant et al., 2011); (2) agricultural land use scenarios in assessing agricultural policy efficiency in terms of achievement of environmental objectives (Volk et al., 2009); (3) best agricultural practices in terms of stream-water quality (Ferrant et al., 2013; Laurent et al., 2007); (4) climate change impacts on surface water (Franczyk and Chang, 2009) or groundwater and irrigation withdrawal (Ferrant et al., 2014); and (5) hydrologic impoundments and wetlands on water resources (Bosch, 2008; Perrin et al., 2012).
Most of these applications require spatially distributed models, where information on soil–crop location within slopes as well as hydrological settings (topography, groundwater storage, reservoir location, and irrigation pumping) is included, to provide spatially explicit information on water uses (Ferrant et al., 2014; Perrin et al., 2012), nutrient transfer, and transformation within the catchment (Arnold et al., 1998; Beaujouan et al., 2002; Ferrant et al., 2011). These modeling approaches enable study of the interactions between upland and bottomland fields, groundwater table fluctuation, and the nitrogen cycle in the soil–plant system. They are especially relevant for localizing the sources and sinks of nitrogen within landscapes – areas prone to nitrogen leaching versus areas favorable to nitrogen retention – that are dynamically changing depending on the cropping patterns and hydrological conditions. The spatial resolution of the simulated processes is linked to the resolution of the available input data (land use, soil, aquifer, and topographic maps). High-resolution data may eventually be required to accurately assess the impact of agricultural practices on water resources. (Perrin et al., 2012) used the SWAT model to simulate groundwater storage under intense agricultural pumping rates in South India. They used high-resolution optical satellite images (between 5 and 10 m) to derive the spatial groundwater extraction from the extent of the irrigated area. This high spatial resolution of pumping rates coupled with hydrogeological setting maps are used within SWAT to identify areas prone to the exhaustion of groundwater resources under current usage for present and future climates (Ferrant et al., 2014).
In complex distributed agro-hydrological models, which simulate numerous processes with numerous parameters to represent spatially the temporal dynamics of water, nutrient cycle, and crop growth, conventional stream-flow calibration may lead to equifinality problems, e.g., more than one parameter leading to similar results (Beven, 2001) or compensation between processes leading to similar stream water fluxes (Ferrant et al., 2011). Uncertainties raised by these modeling approaches at the watershed level are mainly related to (1) the lack of agronomic observations corresponding to all the soil–climatic situations encountered within the catchment, i.e., crop biomass production and the partition between export by harvest and incorporation within soil organic matter by straw burial; and (2) the lack of a priori spatial knowledge, such as the soil's organic matter transformations, saturated conditions within slopes, and feedback on crop productivity. The calibration process is limited to optimizing integrative variables at the watershed scale: discharge and nutrient fluxes at the outlet, occasionally average crop yield (Ferrant, 2009; Ferrant et al., 2011; Moreau, 2012), or, more rarely, aquifer recharge (Perrin et al., 2012). Another important aspect of the uncertainty raised by these modeling approaches is that agricultural operations are imperfectly known. Hutchings (2012) has demonstrated the importance of the timing of field operations in complex dynamic carbon and nitrogen models; for instance, winter crop growth in Europe is highly sensitive to the time of the first fertilization as well as the seeding date.
The above description suggests that spatially explicit process modeling
requires a better spatial and temporal calibration in order to strengthen
the spatial representation of the C, N, and water cycles at the catchment
scale. Products derived from remote sensing (RS) are promising tools for
better constraining and spatially calibrating the agro-hydrological models.
Land cover and sometimes land use (temporal cropping patterns) derived from
RS are generally introduced as input variables. However, RS products have rarely been used in calibration processes. Wagner et al. (2009) have
reviewed the RS techniques used in hydrological models to force RS-derived
variables such as soil moisture, evaporation, snow cover, vegetation
structure, and hydrodynamic roughness. Many of these studies used low
spatial resolution imagery such as the Moderate-resolution Imaging
Spectroradiometer (MODIS), scatterometer data, or microwave and radiometer
data (Brocca et al., 2009, 2012;
Laguardia and Niemeyer, 2008; Liu et al., 2009). Nagler (2011) has
reviewed the recent advances in our knowledge of evaporation on an
environmental scale over recent decades by using remote sensing. For
instance, Chen et al. (2005) have calibrated a TOPMODEL-derived
(Beven, 1997) hydrological model in a small forested catchment using RS
leaf area index (LAI: area of vegetation cover in m
More specifically, some studies have demonstrated the potential interest of using RS-derived AET and LAI in agro-hydrological models to quantify the water balance components in irrigated areas (Taghvaeian and Neale, 2011). AET derived from satellite products has been used to spatially calibrate SWAT in short-period studies (Cheema et al., 2014; Immerzeel and Droogers, 2008; Immerzeel et al., 2008). Cheema et al. (2014) combined global extraterrestrial radiation with atmospheric transmissivity derived from 1 km pixel resolution MODIS data to compute a local net radiation at the scale of the Indus catchment. The latter is used to compute the evapotranspiration with the Penman–Monteith algorithm. The SWAT model is then calibrated against this spatial representation of evapotranspiration fluxes for all the hydrological response units. The spatial calibration method presented in this recent study is still limited by the resolution gap between evapotranspiration products at a moderate resolution and the patchy pattern of irrigated areas that need to be described at a high spatial resolution. Another promising example of RS products used in crop model calibration is reported by Jégo et al. (2012). These authors used LAI retrieved from RS data to reset selected crop management input parameters (seeding date and density) and soil input parameters (field capacity) in the functional crop model STICS (Brisson, 1998). They demonstrated that the predicted yield and biomass were improved, especially in the case of water-stress conditions.
Turning to the distributed agro-hydrological model TNT2, which is based on
the
STICS model spatially coupled with a hydrological model TNT derived from the
TOPMODEL hypothesis, a calibration of crop input parameters could be
performed by matching simulated and observed LAIs at the crop-field level.
The question is whether the spatial calibration of the LAI dynamics using an
LAI map series derived from high-resolution RS data may have a positive
impact on the calculation of water and nutrient fluxes compared with a
standard calibration using discharge. This calibration method would require
high spatial resolution images with a 4 to 5 day revisiting period, which will
be provided by two satellite missions: Ven
The aim of the present study is therefore to explore the advantage of using leaf area index map series, derived from high-resolution RS products, for the spatial representation of the water and nutrient fluxes in an agro-hydrological model. The study focuses on an experimental catchment where intensive monitoring of stream water discharge and nitrate concentration has already been used to calibrate a distributed agro-hydrological model (TNT2) for the period 1985–2001 (Ferrant et al., 2011) by taking into account climatic variables, crop rotation, and agricultural practices. From this starting point, the calibrated model TNT2 was run on a new agricultural and climatic data set for the 2005–2010 period. A set of 105 LAI maps derived from Formosat-2 images (8 m resolution) has been used to optimize LAI temporal growth by iteratively resetting the seeding dates at the crop-field level. Since this input is commonly not reported, missing values were estimated using existing records of seeding dates. Resetting the seeding date is a way to shift crop growth in time. We explore the impact of this spatial optimization using LAI maps derived from optical RS in terms of the water and nitrogen budgets at the catchment level.
Location of the catchment area studied. The Auradé catchment comprises 101 cultivated crop fields. The cropping pattern is a rotation of winter wheat and sunflower. The Formosat-2 series ground coverage is representative of the cropland area characterizing the region around Toulouse. Atmospheric turbulent fluxes, ground vegetation dynamics, and agro-meteorological measurements have been performed in the experimental crop field near the study site since 2005. A detail of the LAI map derived from the Formosat-2 image for 12 July 2009 shows the high variability of the LAI within the sunflower plots (still active at that period of the year), whereas other areas are close to zero, corresponding to winter wheat having reached the senescence stage.
The Montoussé catchment at Auradé (Gers, France) is an experimental
research site monitored since 1983 to investigate the impact of fertilizers
on stream-water quality. In 1985 the fertilizer manufacturer GPN-TOTAL began
nitrate measurements in the stream in order to assess the impacts of
agricultural practices and landscape management on nitrate concentrations in
stream water. This catchment was selected for intensive survey because of
its rapid hydrological response in an intensive agricultural context. The
crop rotation system consists primarily of a sunflower and winter wheat
rotation, fertilized only with mineral fertilizers. Figure 1 illustrates the
agronomical and hydrological situation of the study site. As a tributary
channel of the Save River, itself a left tributary of the Garonne, the
catchment area is representative of a wider agricultural area embedded
within the Gascogne region in southwestern France, where a number of similar
agricultural and geomorphologic settings are found (Ferrant, 2009). This
small catchment (3.35 km
The catchment's soils were mapped in 2006 by Sol-Conseil and EcoLab; the map is presented in Ferrant et al. (2011). 12 soil types were identified along a topographic sequence, from deepest soil (around 2 m) in the bottomland to shallowest soils from middle slope to the top of slope (30 cm to 1 m). These agricultural soils exhibit low organic carbon (from 1.1 to 2 % in the first centimeter to 0.4 % in deep horizons) and high clay contents (25 to 40 % in the first cm to 50 % in deep horizons). Each soil map unit represents an area in which a specific soil type is dominant. Although the delineations are based on only 200 auger boreholes in 325 ha, this map is nevertheless a reliable proxy for the fine variability of soil characteristics observed in the field.
The climate is influenced by both the Oceanic and Mediterranean climates.
Mean annual rainfall recorded on the study site for the 1985–2001 period was
656 mm, with a minimum of 399 and a maximum of 844
A hydrochemical database containing daily discharges and high-frequency nitrate concentration measurements was created and maintained by the AZF company from 1985 to 2001 and has been used to study nitrate contamination of the stream water at the catchment scale (Ferrant et al., 2011, 2013). Using this nitrate-oriented monitoring protocol, many more recent systematic observations and measurements were implemented to improve our understanding of the main processes that drive water, nutrient, and carbon fluxes in the agro-ecosystem and that are likely to be impacted by global changes.
Stream-water nitrate concentrations and discharge were continuously monitored at the outlet of the catchment during the 2005–2012 period (measurement protocol and data are fully described in Ferrant et al. (2012). From the continuous recorded signal, nitrate and water fluxes at the outlet of the catchment are aggregated to a daily time step to match the modeling time step.
Annual inquiries about agricultural land cover and practices are collected from volunteer farmers within the framework of the farmers' association Association des agriculteurs d'Auradé. Seeding dates, tillage operations, fertilizer applications, crop harvest dates, and the amount of fertilizer applied constitute the basic agricultural practices reported by the farmers for each crop field. This cooperative survey never reaches 100 % participation, so many crop-field operations remain unknown. For a given year, the missing seeding dates, fertilization amounts, and fertilization dates are deduced from existing recorded practices. A priori seeding dates are selected on the basis of the farmers' annual reports. Only crop fields owned by a member of this association and located within the area of the municipality are included. Yields are also collected but frequently correspond to an average yield from several unidentified crop fields. This database is not exhaustive: for example, in 2006 only a third of the seeding dates are recorded for the whole municipal area, but none of the corresponding crop fields are included in the experimental catchment. In 2007, the seeding dates of only 18 crop fields among the hundred composing the catchment area are recorded. Expert opinion rules are used to fill the gaps in the database. For a given year, each missing seeding date is estimated by using the average seeding date recorded for the crop fields owned by a farmer. If no seeding date is recorded for a crop field belonging to the farmer, the average of recorded seeding dates, computed for the crop type (wheat or sunflower) and for the year, is used. In this area, recorded winter wheat seeding dates may vary from the beginning of September to the end of November and sometimes even into December. Sunflower seeding dates vary from the middle of March to the end of April. This data reconstruction based on expert opinion rules is designed to find appropriate seeding dates based on farmer behavior and climatic years.
On the other hand, the crop rotation is known for the entire area during the study period. We compare the land cover information contained in the Registre Parcellaire Graphique (RPG) database with crop cover mapping using supervised classification of Formosat-2 and SPOT images. The RPG is based on annual farmer declarations of the land cover for crop-field blocks, a statement which is mandated by the European Common Agricultural Policy (CAP). However, both sources of information give the crop type (wheat, sunflower, rapeseed, barley) but no indication of the cultivar used. The main uncertainty in this agricultural database is linked to the seeding and fertilization dates, as well as to the amounts of fertilization. We refer to these agricultural practices data as “a priori” because they were compiled using non-exhaustive enquiries and used for a first run of the TNT2 model.
Atmospheric flux instruments were set up in March 2005, located in an
experimental crop plot 800 m beyond the eastern margin of the
catchment (Fig. 1). Turbulent fluxes of CO
Destructive measurements of vegetation dynamics were carried out on the
experimental plot during each crop season of the study period. They
consisted of estimating LAI and green area index (GAI)
from aerial biomass measurements at the main development stages
(Béziat et al., 2009). 10 and 30 plants were collected on two
diagonals across the fields for wheat and sunflower, respectively. Sampling
frequency was adapted to the vegetation development, from 1 month during
the slow vegetation development period to 2 weeks during the fast
development period. LAI and GAI were measured by means of a LI-COR
planimeter (LI3100, LI-COR, Nebraska, USA). Between each destructive
measurement date, several randomly distributed hemispherical photographs
were taken to capture the leaf development dynamics. The camera used for
these measurements, a Nikon COOLPIX 8400 equipped with an FC-E8 fisheye
lens, was placed on top of a pole to keep the viewing direction
(downward-looking) and canopy-to-sensor distance (1.5 m) constant throughout
the growing season. The hemispherical photographs were processed using
CAN-EYE V5 (
We used optical remote sensing data from Formosat-2 (F2; Chern et al., 2006) to estimate the LAI for each pixel of the ground coverage area (see Fig. 1). F2 is a high spatial (8 m) and temporal (daily revisit time) resolution satellite with four spectral bands (488, 555, 650, and 830 nm) and a swath of 24 km. For a given site, F2 data can be acquired every day with a constant viewing angle. This characteristic was used to perform accurate atmospheric corrections by estimating the aerosol optical thickness using a multi-temporal method (Hagolle et al., 2008). All F2 images were first pre-processed for geometric, radiometric, and atmospheric corrections, as well as cloud and cloud–shadow filtering (Hagolle et al., 2010).
105 LAI maps at 8 m resolution encompassing the whole catchment (ground coverage shown in Fig. 1) were derived from 105 Formosat-2 images over 5 years (2006–2010) using the BV-NNET tool (biophysical variable neural network; Baret et al. (2007). BV-NNET is based on the inversion of a radiative transfer model (PROSAIL; Jacquemoud et al. (2009) using artificial neural networks. The LAI retrieval method is fully described in (Claverie, 2012, 2013). A main advantage of this method is that it does not require any prior calibration against in situ measurements.
Process of optimization of the seeding date by matching
the early variations of simulated LAI with the interpolated LAI derived from
F2 image series at crop-field scale. The interpolated LAIs are obtained by
fitting a double logistic equation against discrete satellite-derived LAI at
the crop-field scale. The equation describing the growth of the LAI
depends on the cumulative daily temperature
The land cover within the experimental catchment was derived from field survey and F2 images by supervised classification at the crop-field level. These map series were used to explore the spatial and temporal heterogeneity in terms of crop growth at the pixel and crop-field level. Daily values of LAI as a function of cumulative daily temperature were obtained by fitting a double logistic equation against discrete satellite-derived LAI (see equation in Fig. 2) at both crop-field and pixel levels. The results at the pixel level are used to discuss the spatial variability of the crop development observed within slopes and fields, whereas the results at the crop-field level are used in the optimization procedure described in Fig. 4.
TNT2 is a process-based, spatially distributed model developed to study N
fluxes and water cycles in small agricultural catchments (< 50 km
The TNT2 model has been successfully calibrated on the Auradé experimental catchment for the water and nitrogen fluxes at the outlet for a long period of time (1985–2001, Ferrant et al. (2011). TNT2 inputs and parameters include four types of spatial information: (i) a landscape pattern delineating the agricultural plots, roads, hydrological network, and landscape features (wetlands, hedgerows, etc.); (ii) a soil map; (iii) a climate map of climate gradients within the catchment; and (iv) agricultural practices associated with a crop sequence for each agricultural plot during the simulation period.
The TNT2 agronomical module is based on a STICS modeling approach (Brisson, 1998), a generic model that simulates crop growth at the plot scale using the input of agricultural practices: seeding date, crop cultivar characteristics, and mineral and organic fertilization. The crop plant is described by its shoot dry biomass (carbon and N), LAI, and the biomass of harvested crop organs. The cumulative air temperature is the main input variable driving crop growth: crop temperature is used to calculate the sum of degree days by phenological stage. Seeding date and first phenological stage lengths have a great impact on the crop emergence date and the entire LAI profile. Phenological stages are calibrated for each cultivar. One cultivar of wheat (Biensur) was selected (Brisson et al., 2002; Brisson, 1998). One cultivar of late sunflower was calibrated for STICS (Brisson et al., 2003). Water and nutrient stress indices are associated with limitations regarding leaf growth and the net photosynthesis of plants. The soil water and nitrogen contents simulated at a daily time step are combined with the daily crop requirements to compute the transpiration fluxes and nitrogen assimilation within crop biomass.
The water and N cycling in the soils is explicitly detailed by simulating evaporation (maximized by PET derived from a Penman–Monteith methodology) and transpiration, percolation to deep layers and lateral flows, organic matter mineralization, mineral nitrogen denitrification (NEMIS model; (Henault and Germon, 2000; Oehler et al., 2009), and leaching into the hydrological network. The agricultural practices inputs are supplied at the crop-field level: seeding date, fertilization date and amount, straw management, and harvesting date.
The TNT2 hydrological module is a fully distributed hydrological model, adapted to a topography-based shallow aquifer. It is based on the assumptions of the hydrological model TOPMODEL (Beven, 1997): water fluxes are assumed to follow Darcy's law with a constant hydraulic gradient. The hydraulic transmissivity depends on the soil water deficit of saturation. The main differences between TNT and TOPMODEL lie in the distribution of the recharge and the deficit of soil water saturation. TOPMODEL computes water fluxes at the outlet and an average deficit of saturation for the whole catchment, which can be distributed to each point of the basin according to a topographic index. In TNT, calculations are performed following an explicit cell-to-cell routing. The catchment is represented by a cluster of columns. Each top-of-column surface corresponds to a pixel in the digital elevation model (DEM). Each column height is divided into two soil layers corresponding to a root growth zone and a shallow aquifer layer. The soil and aquifer porosity is described as a dual porosity: the retention (micro) and drainage (macro) porosities. The porosity volume must be set up for each layer and for each soil type spatially delineated by the soil raster map. The water's flow paths follow a multi-directional scheme (a pixel may flow into several other pixels), which depends directly on the surface topography calculated from the DEM. Water percolation and nitrogen leaching are computed using cascading horizontal layers similar to Burns' model (Burns, 1974), according to soil porosity characteristics. Both the spatial soil characteristics and the multi-directional scheme derived from DEM define a spatially explicit distribution of recharge and deficit of soil water saturation. In addition to that, the cropping pattern and associated agricultural practices add spatial heterogeneity to this theoretical scheme in terms of water and nutrient transfers.
The model runs on a daily time step. Water balance and N transformations are computed for each cell of the raster grid of the DEM, from upstream to downstream, by following the cell-to-cell drainage routing. Daily discharge and nitrogen fluxes are computed at the outlet from the catchment.
The model was calibrated for the period 1985–2001 firstly by optimization of
the daily discharge using both hydrological parameters To and
Using the same set of parameters as in Ferrant et al. (2011, 2013), we evaluated the simulations for the period 2005–2010 in terms of hydrological and nitrogen fluxes, as well as the evapotranspiration and LAI/biomass data that were measured in the experimental crop field (Fig. 1). We then used the F2 LAI data from 2006 to 2010 to perform the optimization process of the LAI.
An algorithm designed to minimize temporal shifts between simulated LAI
profiles and interpolated LAI profiles based on satellite images at the
crop-field level was implemented (Fig. 2) to reassess seeding dates at the
crop-field level. A first LAI profile is simulated for each crop field.
Since the cumulative air temperature is the main input variable driving crop
growth, the temporal shift (
Drainage and nitrogen fluxes simulated for the whole catchment are compared
to the measurements at the outlet. For this study, the hydrological
calibration of input parameters presented in Ferrant
et al. (2011) is not modified. Similar performances are found for daily
discharges (Nash–Sutcliffe coefficient
Yearly water and N balance simulated in TNT2 model for a priori and re-set seeding date.
Observed in-stream nitrogen fluxes from January 2007 to December 2010 are
close to 7 kg N ha
Above: maps of maximum LAI observed for each year and each crop mask. Maxima of winter wheat LAI were not observed during spring 2008 owing to heavy cloud cover throughout the area. Each date of image acquisition constituting the F2 series is indicated by a triangle in the timeline. Below: spatial variability of the LAI as a function of the time between (inter-sigma) and within (intra-sigma) crop-field measurements for sunflower.
Figure 3a shows the maps of maximal LAI for each pixel, year, and crop. Figure 3b shows the LAI spatial variability observed for a sunflower crop field as a function of time: the spatial variability increases concomitantly with crop growth. This variability, expressed as the standard deviation (sigma), is of the same order of magnitude when considering variability between crop fields and within crop fields. The processes driving this spatial variability are mainly related to soil patterns, localization within slope, or aspect of the slope. The absolute value of LAI retrieval is compared with field measurements. Figure 4 compares two measurements of LAI: (1) RS LAI retrieved from satellite or hemispherical photographs, and (2) direct measurement by the destructive method. Error bars represent plus or minus one standard deviation of the median of the samples collected for the destructive method. The variability of the result is associated with both the spatial variability of LAI and biomass encountered throughout the crop field and an imprecision attributed to the measurement method itself. The LAI estimated from hemispherical photographs is an average estimate for the whole area covered by the camera lens; error bars represent a fixed uncertainty related to this measurement method (Demarez et al., 2008). The satellite-derived LAI estimates for the crop-field level are represented by the median, plus or minus one standard deviation of the LAI value of each pixel located within the crop field. The error bar represents the spatial variability detected by remote sensing.
Leaf area index derived from satellite F2 images, hemispherical photographs (LAI effective CanEye), and direct field measurement (LAI destructive) in the experimental crop field located near the Auradé catchment (see location in Fig. 1). The standard deviation represents the spatial variability within the crop field (LAI satellite), spatial variability and associated sampling error (LAI destructive), and uncertainty concerning the photo interpretation (LAI effective).
The 44 cloud-free Formosat-2 images acquired in 2006 ensure a fine-grained
description of the winter wheat development. The intra-field LAI spatial
variability obtained from the satellite retrievals is close to 1 m
Figure 5 shows the results of optimizing the temporal dynamics of the LAI
average over the 101 crop fields. Reinitialization of the seeding dates
decreases the temporal shift (
Seeding date trajectories for each crop field as a function of iteration number.
The trajectories of seeding date solutions as a function of the iteration number (Fig. 6) show a rapid convergence after five iterations. There are few crop fields for which no realistic solutions were reached. For the sunflower in 2007, four crop fields converged on an early seeding date in October to December. This exclusively concerns the sunflower in certain small crop fields (several hectares) for which average LAI remains low (< 1). In these cases, the maximum of observed LAI is too low or the proportion of mixed pixels (at the crop-field border) is too high, thereby leading to unrealistic interpolations of the LAI profile at the crop-field level. The annual seeding dates estimated by this method constitute a long period for the winter wheat and a short one for the sunflower. These ranges are from September to the beginning of November for winter wheat and between January and April (highly dependent on the climatic year) for the sunflower. In 2007, 2008, and 2010 the seeding dates for winter wheat and sunflower crop fields were recorded within the experimental catchment. The average differences between estimated and actual seeding date in 2007 were 20 and 8 days for the wheat and sunflower crops, respectively. These figures rise from 1 day to 1 month and from 1 to 17 days for wheat and sunflower, respectively. Three factors are responsible for these heterogeneous differences: inappropriate cultivar growth parameters, inaccurate detection of emergence period by biased LAI interpolation from remote sensing, and uncertainties in farmer statements (completed at the end of each year).
Table 1 presents the annual water and nitrogen fluxes computed for the
entire simulation period (2006–2010). The changes in crop development
induced by the reinitialization of input parameters have a small effect on
the discharge and AET (around 1 to 2 mm yr
LAI and biomass simulated for 4 years in the crop within the catchment that exhibits a cropping pattern comparable to the experimental crop field (except in 2009) where the ground measurements are carried out. Rows stand, respectively, for winter wheat 2006, sunflower 2007, winter wheat 2008, and winter wheat 2010. LAI in the first column: the red curve is the interpolated LAI profile from the F2-derived values (red circles) with the spatial variability represented by the bars. The black diamonds represent the F2–LAI values for the experimental crop field located outside the catchment. Black solid and dotted lines are the average LAI after and before seeding date modification, respectively; bars represent the standard deviation of simulated LAI within the crop field. Biomass in the second column is represented by black diamonds for the measurements, with the measurement variability associated with the spatial variability and accuracy of the measurement method. Black solid and dotted lines are the average biomass after and before seeding date modification, respectively; bars represent the standard deviation of simulated LAI within the crop field.
Figure 7 shows the results of seeding date reinitialization on the LAI and biomass estimates, respectively. We compare two crop fields: one is located within the catchment where TNT2 simulations are performed (crop field 8) and the other is the experimental crop field where measurements of turbulent atmospheric fluxes are carried out (see Fig. 1). The crop fields are close to each other and comparable in terms of slopes and crop rotation, except in 2009 when rapeseed crop was grown in the experimental field and sunflower was sown in crop field 8, located within the catchment. Remotely sensed LAI values for both crop fields are compared to illustrate the differences observed between the crops in terms of vegetation dynamics. The interpolated daily LAI is presented in Fig. 7 for the crop within the catchment, and the simulated LAI profiles before and after the optimization are plotted. The simulated biomasses before and after optimization in crop field 8 are compared to the measured biomass within the experimental crop field. The spatiotemporal variability of this variable is close to the measurements for the 4-year period, except in 2008 when no optimization could be performed because of heavy cloud cover. The seeding date modifications have a substantial impact on biomass production and clearly improve the biomass predictions for 2010.
Left: measured versus simulated daily actual
evapotranspiration from the experimental crop field and crop field 8,
respectively. Measured AETs are given with the uncertainty envelope
associated with the eddy covariance measurement precision (Béziat
et al., 2009). The Nash–Sutcliffe coefficient, correlation coefficient
(without units), and RMSE (mm day
Figure 8 compares the daily simulated and measured evapotranspiration fluxes
at the crop field 8 and experimental crop-field levels. Each series is
strongly correlated in time (
Field measurements of LAI or AET are expensive, time-consuming, and limited to local evaluations of the crop cover. The satellite observations are thus essential for monitoring the crop-cover dynamics at crop-field scales. In the context of the present study, leaf area index and biomass are highly variable in space and time and within crop field. The high spatial resolution (around 10 to 20 m) is sufficient to capture the spatial variability of crop productivity. The large number of images, provided by the high frequency of satellite revisit, makes it possible to describe the temporal crop development and productivity at pixel and crop-field levels by describing the LAI profile retrieved from F2 images by means of a physically based double logistic descriptive equation (Fig. 2). This temporal information has been used at the crop-field level to optimize the simulated LAI of the process-based model STICS, coupled with a hydrological model that aims to reproduce the varying local situations created by hydrological conditions within the catchment. The objective of minimizing the temporal shift between measured and observed LAI by re-initializing the seeding date in TNT2 is satisfactorily fulfilled: there is a rapid convergence of the optimization process, with temporal shifts being generally minimized with a realistic seeding date solution. The improvement achieved from the a priori situation constructed from the local database would have been made more evident by constructing a seeding-date scenario based on regional recommendations. This could be done in future applications at larger scale, e.g., by considering ground coverage of complete Formosat-2 scenes.
Nevertheless, although seeding-date values are a good numerical solution for phasing simulated and RS-retrieved LAI profiles, final seeding-date values mainly depend on the cultivar parameters, such as length of the early development and vernalization stages. For instance, the duration of winter wheat vernalization, corresponding to the low temperature periods required to hasten plant development, will depend on the number of vernalizing days defined for each wheat cultivar (JVC parameter) and the crop temperature computed from climate input data. The mild winter conditions in the study area make the LAI profile insensitive to the seeding date for high values of JVC (> 8). We have therefore set the JVC parameter to 6 days for the winter wheat cultivar used in this study. This shows that the variety of wheat sown is crucial information for a better estimation of the true seeding date, crop-growth dynamic, and yield. More generally, crop variety is not recorded in an agricultural database. In this specific study site, several varieties were recorded which were not pre-calibrated in the STICS model. The estimation of a “true seeding date” at catchment scale is accordingly not possible at present.
Soil and crop-field map used in TNT2 (top). Spatial NUEs for the years 2006 and 2007 (bottom left and right, respectively). The higher the value, the more efficiently the fertilizer is used by the plant. A low fertilizer amount with weak biomass production could lead to high NUE. Mineralization of soil organic matter creates a source of mineral nitrogen that leads to NUEs higher than unity.
Jégo et al. (2012) used LAI data retrieved from satellite images to better constrain input parameters for the STICS crop model. By reinitializing the seeding date, they greatly improved the model's predictions in terms of biomass and yield. The optimization method is based on the simplex algorithm to minimize the weighted sum of squared differences between RS-retrieved and simulated LAI series. A run of the crop model is carried out for 1 crop and 1 year and takes less than a second. This optimization method is appropriate since it tests several input parameter couples in order to converge quickly on an optimal solution in terms of the chosen estimator. In the case of TNT2, simulations are sequentially executed: each pixel calculation depends on the previous and simulated neighborhood conditions. A single run corresponds to the simulation of water and nutrient fluxes in 134 013 modeling units, covering 101 crop fields for 5 years. It thus requires much more computation time (around 2 h for the Montoussé river catchment). The hydrological interactions between modeling units in space and time imply that changes in seeding dates are interdependent. The optimization method described in this paper was chosen because it is based on a quantitative (rather than a statistical) estimator of the temporal shifts, which is used to quantitatively correct the input parameter (in this case the seeding date) based on the model's functioning. The temporal delay between RS-retrieved and simulated LAI series is evaluated as a physical variable: the cumulative daily air temperature difference. The results of this optimization show a rapid convergence after five to eight iterations.
The STICS crop model (the agronomical portion of the TNT2 model) is a
process-based model, i.e., it is able to scale up the results of local
experiments. It extrapolates the crop-growth variables from analogous
situations described by input data (soil, climatic, and cropping management)
without the need for new testing. The coupling of this process-based model
with a hydrological model seeks to simulate the varying local situations
described by hydrological conditions within the catchment: saturated zones
and
soil water content as a function of the situation within a slope. The
hydrological variables – evapotranspiration and discharge – are not heavily
impacted by this change in crop-cover dynamics. The difference obtained for
AET, i.e., 2
Other input parameters than the seeding date should be considered for further optimizations. Jégo et al. (2012) have identified a second input parameter known to have a great impact on crop productivity within the STICS crop model: the soil's water-retention capacity. In the TNT2 model, the soil map defines homogeneous zones where 21 soil parameters are defined. The sensitivity of the spatial pattern of soil input parameters within agro-hydrological models has not yet been deeply explored. Figure 8 shows the spatial variability of the F2-derived and TNT2-simulated LAI at the pixel level for two dates. Two covariates seem to drive the spatial variability of the LAI variations simulated by TNT2: the soil map and the location of the drainage network. Three main situations are simulated: (1) systematic saturated conditions, which limit LAI development in the drainage network location; (2) low soil water deficit, which enhances LAI development; and (3) intermediate or low soil water content, which limits LAI development. There is an excellent potential for agronomical calibration of agro-hydrological models by reinitializing soil input parameters and refining local situations at the pixel scale, using these new LAI map series derived from optical RS with high revisit frequency. Considering only the hydrological variables, Moreau et al. (2013) tested the sensitivity of the TNT2 model's response to spatial soil input parameters for both water and nitrogen-related parameters. They analyzed the output's sensitivity in terms of in-stream water and nitrogen fluxes at the outlet and concluded that sensitivity to the spatial distribution of soil input factors is low. Looking ahead, we consider that the sensitivity of spatial soil input parameters is high for crop variables and would impact the spatial representation of the N cycle within slopes. Reinitialization of physical soil parameters in the TNT2 model will be proposed in a forthcoming study at the pixel level using the same F2 data set. The control of these parameters versus other physical catchment parameters (aspect, slopes, etc.) on the spatial and temporal variability of the crop growth will be explored.
The present study has evaluated the potential of remote sensing data series for the spatial and temporal calibration of a distributed agro-hydrological model over a 5-year period (2006–2010). The use of a process-based crop model (STICS) coupled with a simplified hydrological model (TNT) provided the means to simulate the water and nitrogen budgets as well as the yields of a soil–plant system at the catchment scale, taking climatic and agricultural variables into account. The lack of spatial and temporal calibration of soil–crop situations is assessed in light of the additional spatiotemporal information derived from RS images. The spatial calibration of model input parameters by using LAI derived from RS image series, previously confined to a priori values, opens new opportunities for constraining spatial and temporal crop development at the catchment scale. In this example, we satisfactorily constrained the temporal LAI development at the crop-field level by reinitializing the seeding dates. This calibration step adds value to the conventional calibration process usually employed in agro-hydrological models. The improved representation of crop-cover growth has no noticeable impact on the water budget at the catchment scale (around 1 %), but had substantial impacts on the nitrogen cycle in terms of crop uptake and biomass, as well as on nitrate leaching and in-stream losses. The optimization process using RS-derived LAI profiles has enabled an increase in nitrogen uptake by the crop and in biomass production for winter wheat, leading to a significant drop in the simulated in-stream nitrogen losses of around 12 %. This result indicates that a spatial calibration of the crops' biophysical variables such as LAI changes the nitrogen-use efficiency (NUE) at the crop-field level, which impacts the nitrogen cycle at the catchment scale.
This study demonstrates the contribution of high spatial resolution optical
satellite images with frequent systematic observations to the spatial
calibration of agro-hydrological models. This type of spatial calibration
greatly improves the capacity of agro-hydrological modeling to explain,
reproduce, and predict spatial crop growth by constraining the spatial water
and nutrient fluxes within a hydrological catchment. Massive systematic
satellite observations will soon become widely available thanks to the
forthcoming satellite missions Ven
We would like to thank the Association des Agriculteurs d'Auradé (today Groupement des Agriculteurs de la Gascogne Toulousaine) for their cooperation and the people who are behind the large amount of data presented in this manuscript: Nicole Ferroni, Bernard Marciel, Pascal Keravec, Hervé Gibrin, Tiphaine Tallec, Pierre Béziat, Pierre Adrien Solignac, Aurore Brut, Jean-François Dejoux, Claire Marais-Sicre, Jérôme Cros, Olivier Hagolle and Mireille Huc. Nitrate concentration and stream discharge were recorded within the framework of a GPN-ECOLAB convention on the experimental catchment of Auradé. The BVEA (Bassin Versant Expérimental d'Auradé) is a regional platform of research and innovation in Midi-Pyrénées and innovation and which is involved in the French SOERE Network RBV (Experimental catchment network) and in the international Critical Zone Exploratory Network. Sylvain Ferrant was the recipient of a CNES (Centre National d'Etudes Spatiales) post doctoral research grant. Edited by: V. Andréassian