INTRODUCTION
Evaluating nitrogen (N) nutritional status is a key issue for analysing, monitoring and managing cropping systems (e.g. Naylor & Stephen Reference Naylor and Stephen1993; Senanayake et al. Reference Senanayake, Naylor and DE Datta1996; Jeuffroy et al. Reference Jeuffroy, Ney and Ourry2002; Jaggard et al. Reference Jaggard, Qi and Armstrong2009). A reliable estimation of crop requirement allows optimization of both qualitative and quantitative aspects of production (Ghosh et al. Reference Ghosh, Mandal, Mandal, Lodh and Dash2004), and increases N use efficiency, therefore reducing the impacts on the environment. For these reasons, different typologies of tools for supporting N management have been developed. They differ in their scope (i.e. supporting fertilization, evaluating losses and identifying scenarios suitable for specific contexts), in the reference spatial scale (e.g. field and region), in the economic and time resources they demand and in the skills required by the user. Some of these tools (e.g. leaf colour charts (LCC); Alam et al. Reference Alam, Ladha, Khan, Foyjunessa, Harun-ur-rarun-ur-rashid, Khan and Buresh2005) are very simple and cheap, being conceived to support farmers directly for N-management at field level. Others (i.e. SPAD-502; Konica Minolta Inc., Tokyo, Japan) are more complex and use plant chlorophyll content – estimated through the absorbance of the leaf in the red and near-infrared regions (Peng et al. Reference Peng, Garcia, Laza, Sanico, Visperas and Cassman1996) – as a proxy for plant N concentration (PNC) (Clark et al. Reference Clark, Gowing, Lark, Leeds-Harrison, Miller, Wells, Whalley and Whitmore2005). Stroppiana et al. (Reference Stroppiana, Boschetti, Brivio and Bocchi2009) estimated rice PNC from hyperspectral radiometry, finding significant correlation between PNC and a vegetation index calculated using reflectance values measured for wavelengths of 483 and 503 nm. Among the available tools for evaluating crop nutritional status, simulation models play a major role in a variety of applications (Stöckle & Debaeke Reference Stöckle and Debaeke1997; Meynard et al. Reference Meynard, Cerf, Guichard, Jeuffroy, Makowsky and Recous2001), because of their flexibility, ability to work at different spatial scale and suitability for scenario analysis. In most of the cases, they quantify crop N nutritional status through the definition of a critical N threshold (Ncrit, mg/g). When PNC (mg/g) is lower than Ncrit, the crop is growing under N-limiting conditions, otherwise consumption is not limiting.
Although some of the proposed crop models relate Ncrit to crop development (e.g. EPIC, Williams et al. Reference Williams, Jones, Kiniry and Spanel1989; DAISY, Hansen et al. Reference Hansen, Jensen, Nielsen and Svendsen1991), the approach most commonly used is that based on the N dilution law (Salette & Lemaire Reference Salette and Lemaire1981; Justes et al. Reference Justes, Mary, Meynard, Machet and Thelier-Huche1994). According to this law, an allometric function is used to derive decreasing values of Ncrit for increasing values of above-ground biomass (AGB, t/ha). The theoretical explanation for the dilution law refers to the representation of the plant as composed of two N pools (Caloin & Yu Reference Caloin and Yu1984): one is involved in physiological activities such as photosynthesis and nutrient uptake, the other consists of mechanical and vascular structures (Seginer Reference Seginer2004). The two N pools change their relative weight with crop ageing because of (i) the self-shading of leaves and (ii) the allocation of resources to plant components with different N concentrations: mainly to leaf blades (high N concentration) during early growth stages and to stems (low N concentration) in later stages. Self-shading of leaves results in the lower canopy layers experiencing a decrease in the red/far-red ratio because of a preferential absorbance of red photons by chlorophylls in the upper layers. Changes in this ratio trigger senescence processes aimed at recycling N-rich compounds, which are no longer needed in the shaded leaves. These are reallocated to younger leaves, thus reducing the global need for N by the whole plant (Seginer Reference Seginer2004).
Although the N dilution law is adopted by some of the most widely used crop models (e.g. CropSyst; Stöckle et al. Reference Stöckle, Donatelli and Nelson2003) and parameterizations for a variety of crops are available (e.g. Greenwood et al. Reference Greenwood, Lemaire, Gosse, Cruz, Draycott and Neeteson1990), the estimation of its empirical coefficients is quite complex, requiring various experiments with at least 4–5N application rates (e.g. Justes et al. Reference Justes, Mary, Meynard, Machet and Thelier-Huche1994). Moreover, it is characterized by a level of empiricism that makes too implicit the representation of the two processes used to give theoretical support to the law itself (i.e. self-shading and allocation of photosynthates to organs with different N concentration).
The current paper presents the results of the project Modelling critical N concentration for agronomic management (MAZINGA), carried out with students of the Cropping Systems Masters course of the University of Milan with the aim of developing new approaches for the simulation of rice Ncrit. The proposed model is evaluated using field data collected in Northern Italy and compared with other existing approaches.
MATERIALS AND METHODS
Approaches compared in this study
The dilution law (Salette & Lemaire Reference Salette and Lemaire1981) is based on the following equation:
where a and b are empirical coefficients derived from field experiments carried out under different N fertilization treatments, as proposed by Justes et al. (Reference Justes, Mary, Meynard, Machet and Thelier-Huche1994). The a and b values used in the present work (5·18 and 0·52, respectively) are those proposed by Sheehy et al. (Reference Sheehy, Dionora, Mitchell, Peng, Cassman, Lemaire and Williams1998) and are considered reference values for rice.
The approach implemented in the EPIC model (Williams et al. Reference Williams, Jones, Kiniry and Spanel1989) derives Ncrit using the following equation:
where HUI (0·0–1·0) is a heat unit index, equal to the ratio of the current thermal sums to the thermal sums to reach maturity, defined as Zadoks Growth Stage (ZGS) 92 (Zadoks et al. Reference Zadoks, Chang and Konzak1974); b1, b2 and b3 are coefficients (0·005, 0·045 and 2·197 in the present study) derived by forcing Eqn (2) to return species-specific tabulated PNCs for HUI values of 0·0, 0·5 and 1·0 (Williams, personal communication).
The DAISY model (Hansen et al. Reference Hansen, Jensen, Nielsen and Svendsen1991) forces an allometric relationship between Ncrit and growing degree days (GDD; °C-day) to return the PNC values of 30 and 10 mg/g, respectively, for 100°C-day and GDD to reach maturity (ZGS=92).
The approach proposed in the present study (MAZINGA) reproduces the effect of leaf self-shading in remobilizing (recycling) N from senescent leaves through an inverse of the fraction of radiation being intercepted by the canopy:
where Nmat (mg/g) is a parameter indicating the critical N concentration at physiological maturity (ZGS=92), 1−e−k× LAI is the fraction of solar radiation intercepted by the canopy, k (−) is the extinction coefficient for solar radiation, and LAI (m2/m2) is the total leaf area index. Values of 10 mg/g and 0·5 were used for Nmat and k (Confalonieri et al. Reference Confalonieri, Acutis, Bellocchi and Donatelli2009), respectively.
Since all the approaches would lead to positive infinity for values of their driving variable (AGB, HUI, GDD and LAI) approximating zero, the minimum between their Ncrit values and 36 mg/g was used (Confalonieri & Bocchi Reference Confalonieri and Bocchi2005).
Experimental data
Data used to evaluate the different approaches for identifying the critical N curves were collected in four experiments (Table 1). These experiments were not designed to collect data to determine the parameters of the curves (in this case, at least five different N application rates would have been needed to parameterize the dilution law curve with sufficient accuracy), but simply to test and compare the different models for the estimation of Ncrit.
Table 1. Datasets used for the evaluation of the models under test
The experiments are representative of the conditions experienced by rice growing in the eastern part of the north-west Italian rice district (Piedmont–Lombardy), responsible for more than 0·40 of the total European rice production. Indica-type cultivars were used in all experiments (Table 1). Rice was scatter seeded and grown under continuous flooding conditions in Vignate, Milan and Opera during 2004 (Expts 1, 3 and 4, respectively; see Table 1). For Expt 2 (Opera, 2002), rice was row seeded and the field was flooded at the 3rd leaf stage (ZGS=13).
During the 2002 and 2006 experiments, three levels of N fertilizer (Table 1) were used in a completely randomized block design with three replicates. The same scheme (completely randomized block) was used for the 2004 experiment, but with two levels and four replicates. Where more than two N levels were set, only the maximum and the minimum were used in the present study (Table 1). Nitrogen fertilizer was always distributed as urea by splitting the total amount (Table 1) in three events (pre-sowing, mid-tillering, ZGS 22 and panicle initiation) in the 2002 experiments, and in two (a single top-dressing fertilization at panicle initiation) events in the others.
For all the experiments, the soil presented a medium organic N content (1·3 mg/g in Opera 2004 and Milan 2006, 1·5 and 1·2 mg/g for Vignate 2002 and Opera 2002, respectively).
Possible differences among observers were reduced (i) by always using the same sampling protocol: sample size determined according to Confalonieri et al. (Reference Confalonieri, Stroppiana, Boschetti, Gusberti, Bocchi and Acutis2006), an aggregated sample of three randomly selected sub-samples for each plot, harvesting the whole plant and division (at the collar) of the above- and below-ground parts of the plants in laboratory and (ii) by supervising all the sampling activities. All the analyses were carried out at the laboratory of the Department of Plant Production of the University of Milan.
In all the experimental fields, diseases and weeds were controlled by two treatments with tricyclazole (at ZGS 31 and 45) and with propanil (at ZGS 12 and 22; Imazamox for Expt 4); treatments with alphacipermetrine prevented pest damage.
Total plant and soil N contents were estimated with an elemental analyser (NA 1500 series 2 NC, Carlo Erba, Italy), whereas soil N-HH4+ and N-NO3− were measured with a continuous flow analyser (Flow Comp 1500, Carlo Erba, Italy).
Evaluation of the different approaches
The evaluation of the different approaches for Ncrit was carried out using two methods. The first was the graphical method proposed by Stöckle & Debaeke (Reference Stöckle and Debaeke1997). Ncrit curves were plotted on charts together with PNC values collected from plants growing under N-limiting conditions and on fully fertilized plants. The ideal condition was considered the one where PNC values from plants growing under N-limited (white symbols in Fig. 1) and N-unlimited (grey symbols) conditions were below and above the Ncrit curve, respectively. Different symbol shapes in Fig. 1 refer to different experiments (Table 1). The second evaluation method aimed at quantifying different typologies of error and it was based on the following assumptions: (i) errors are calculated on PNC values staying on the ‘wrong side’ of the Ncrit curves and, for these PNCs, (ii) the ‘true’ values are the corresponding points on the Ncrit curves. Error was quantified using relative root mean square error (RRMSE; minimum and optimum=0; maximum +∞), modelling efficiency (EF; −∞ to 1; optimum=1; if positive, indicates that the model is a better predictor than the average of measured values) and coefficient of residual mass (CRM; −∞ to 1; optimum=0; if positive indicates model underestimation).
Fig. 1. Performance of the model MAZINGA (see Eqn 3) in estimating critical N curve. Comparison with the dilution law, EPIC and DAISY approaches. Plant N concentration (PNC; mg/g) on the y-axis. Grey and white symbols refer to PNCs determined, respectively, on plants grown under non-limited (fully fertilized) and N-limited conditions. Different symbols refer to different experiments (Table 1): circles, rhombus, triangles and squares refer to Expts 1, 2, 3 and 4, respectively.
RESULTS
Figure 1 shows a considerable over-estimation of the Ncrit curve for the dilution law. Although less important, the EPIC approach also showed some over-estimation. On the other hand, DAISY under-estimated the Ncrit curve, although the situation appeared to slightly improve in the last part of the crop cycle. Under the conditions explored, the approach proposed in the present study was the one showing the best performance, with a balanced distribution of the PNC values above and below the Ncrit curve.
Table 2 confirms the improvement of the MAZINGA approach, which presented the lowest number of points (17) on the ‘wrong’ side of the curve (i.e. those appearing above the curve that were expected to appear below it, and vice versa) and a balanced distribution of over-estimated (expected below the curve but above) and under-estimated values (0·53 and 0·47, respectively). The tendency of the dilution law and of the EPIC approach in over-estimating the critical N concentration already discussed (Fig. 1) is confirmed by the proportions of PNC values on the wrong side of the curves. The values from fully fertilized plants that are below the critical N curves are 0·95 and 0·85 for the dilution law and EPIC. The DAISY model appeared less unbalanced than either the dilution law or the EPIC approach, with 0·70 of the values from unfertilized plants above the critical N curve. However, the DAISY approach was the one with the largest number of values on the wrong side of the curve (23).
Table 2. Evaluation metrics calculated for the PNC data staying on the ‘wrong side’ of the critical N (Ncrit) curves (see Fig. 1); their projections on the Ncrit curves were used as simulated data to calculate the indices of agreement
* PNC (mg/g).
† PNC values from unfertilized plants: they should be below the curves.
‡ PNC values from fully fertilized plants: they should be above the curves.
§ RRMSE; 0 to +∞; the closer values are to 0, the better the model.
** Modelling efficiency; −∞ to 1; optimum=1.
†† CRM; −∞ to 1; optimum=0; if positive indicates model underestimation.
DISCUSSION
The dilution law and the DAISY approach were less accurate in the first part of the crop cycle, when top-dressing fertilization (usually at the panicle initiation, ZGS=31, sometimes at tillering, ZGS=20) is carried out. This is important since inaccuracies in the estimation of N requirements at these times could nullify the usefulness of models as diagnostic tools for supporting N management in cropping systems. Moreover, most of the total N present in the plant at maturity is taken up before flowering (Ntanos & Koutroubas Reference Ntanos and Koutroubas2002; Lin et al. Reference Lin, Zhou, Zhu, Chen and Zhang2006; Zhang et al. Reference Zhang, Fan, Zhang, Wang, Huang and Sheng2007; Acreche & Slafer Reference Acreche and Slafer2009). Thus, errors in the first part of the crop cycle could have a considerable impact also on the accuracy of the estimates of the N budget terms. EPIC and MAZINGA appeared to have a more constant behaviour along the crop cycle.
Although the two approaches deriving Ncrit from development, i.e. EPIC and DAISY, presented the highest number of values on the wrong side of the curve (20 and 23, respectively), measured PNC values appeared less spread around their curves (Fig. 1) compared to the other two approaches. This is the reason why they produced the best values of RRMSE and EF, since the larger number of points on the wrong side of the curves is counterbalanced by a general closeness of the points to the curve.
In spite of its higher accuracy compared to the other tested approaches, the model MAZINGA does not include any empirical coefficient, being based only on canopy extinction coefficient for solar radiation and leaf area index, both easily estimable using simple instrumentation (Stroppiana et al. Reference Stroppiana, Boschetti, Brivio and Bocchi2009). This makes the model MAZINGA easy to use, both within crop models and as diagnostic tool for supporting N fertilization. The results achieved in the present work may be applicable to other herbaceous species.
The didactic experience of carrying out a (small) modelling research project with students can be considered fully satisfying (a new approach was proposed), and follows some similar activities recently carried out (Antunes et al. Reference Antunes, Brejda, Chen, Leavitt, Tsvetsinskaya, Weiss and Arkebauer1998; Graves et al. Reference Graves, Hess, Matthews, Stephens and Mason2002), with models used as tools in education.
