The PA experience started in 1998 with a harvester which was equipped with a yield meter, and then with a yield monitor, a satellite receiver and a light bar to assist manual drive. The farm now spreads fertilizers using a VRT technology, following prescription maps or real time signals of vigor sensors using an OptRx by AgLeader (Figure 3), which reads both NDVI (wavelengths of 670 nm and 780 nm) and NDRE (730 nm and 780 nm).

Before N-pi administration, a drone was used to assess the mean and probable range of the NDRE readings of the sensor, and to fix the thresholds. Following an RTK signal, herbicides are sprayed by a VRT sprayer, which controls the opening and closing of the nozzles lead from a GPS signal, and yield maps are obtained by an harvester equipped with an AgLeader yield meter. All the technology had to be adapted to the soil (Table 1). The name of the farm, Palestro (GPS data of the farm buildings: WGS 84 zone 32 N Lat 45,371488 - Lon 8,298165), comes from the Latin Palustris , which means Swampy. All the soil, that is, 249 ha, of which 230 ha are cultivated, is silty-clay, and is flooded from April to August. Moreover, a rice monoculture has been grown on this farm since 1965.

Yield data from each 19x19m2 mesh plot have been elaborated for the 2012-17 interval (No. 1165 ha^-1). Yield repeatability was calculated as a Pearson correlation of the plot-yields within field between consecutive years. Several mixed models (PROC MIXED by SAS 9.0), concerning the yield and NDRE, were applied to the statistical parameters of the fields, whereby the effects of the field or cultivar were considered fixed and the year was taken as the random factor in order to estimate the linear trend. Conversely, when the years were considered fixed, and the fields were considered as the random factor, a least squares solution was found. The covariation of the variability between the different traits was enhanced by fitting a linear or non-linear regression of the estimates. A Friedman s test for paired samples was used to compare the yield averages of the farm vs. the regional means 7. In a sample field, over a period of two years, the total N (N-tot), and N divided into N-pre-pi and N-pi, was related to the yield by developing a partial correlation with NDRE.

Place: r12; r13 and r23 then r12.3 = (r12 - r13 * r23) / ((1-r13^2)^.5* (1-r23^2)^.5)).

Place: regression b1/2 = covariance12 / stdev2 ^2; regression b2/1 = covariance 12 / stdev1 ^2 then (b1 * b2) = r12^2.

The first lesson learned from the farm was that there was a low repeatability of parcel yields in subsequent years. Since this parameter is independent of the operators, it is generally neglected. The PA capabilities can now be used to monitor the parcel responses at a capillary level, thus parcel repeatability could be considered by means of algorithms, especially considering that its positive evolution seems to follow the increase in variability. In fact, a perfect PA aims to a strong homogeneity in parcel production, ergo a zero-correlation between the O/I framework. The germinability of the seeds for flooded rice production is a very critical point, and a variable steam density may affect the yield 4,5

The second lesson pertained to the increase in variability of the yield and, to a lesser extent, in NDRE after that a strong negative correlation (< -0.90) was imposed on the N-pi variable fertilization, as inversely proportional to the vegetative vigour attained near PI. Cordero et al 8 in a two-year study on experimental plots fertilized at 0÷300 N-tot, have recently proposed a more refined but similar inverse function to establish the acceptable reduction relative to the maximum obtainable yield. In a study on sugarcane 9, the biomass was the main crop parameter that influenced canopy sensor readings, because biomass readings generally result in stalk yields. Thus, canopy sensor readings can efficiently predict relative sugarcane yields when data that are normalized to the mean of the field are considered. Hence, an algorithm that considers this relationship was established, based on an r -0.68 correlation between NDRE and its coefficient of variation, which was lower than the -0.33 of the present work. However, the NDRE in the 16 sugarcane plots was positively correlated with the yield (+0.82), so that the N topdressing was calibrated in positive vs. NDRE, and not in negative as in the present work. Thus an increase in NDRE should reduce the yield variability, but this has not yet been observed in real fields.

The improvement of the mean and maximum yields, and not of the minimum, depends on several plots where the limiting factor cannot be referred to the N availability. A next challenge will be to check the repeatability for yield in those critical plots and to find which is the limiting factor, or, more simply, to reduce the N rate, avoiding a waste.

The third lesson concerned the negative correlation between N-tot and yield, which resulted even after absorbing the N-pi in the N-tot. In the experimental plot studies, the correlations were generally positive, and even very high, until the maximum yield was reached 5,8 but results from the real fields are much less optimistic: i.e. Zavattaro et al 10 observed a sparse regression (+0.0125 t kgN-tot^-1) of the yield (avg. 6.55 t ha^-1) on the N-tot (121.6 kg ha^-1) for seven variety groups from 67 land units in NW Italy, where the correlation coefficient was recalculated to r 0.26; the average NUE, compared to the present work, was 54 kg kg^-1,that is,an increased value of 15%, while the N-tot and yield appeared reduced by 27% and 16%, respectively.

The fourth lesson is related the uncertainty of the reference parcels which identified at zero- input 11, or at well-fertilized input5 or by a corrective sufficiency index 8,5, which are necessary to set up the production models and the PA. As shown in Figure 9, the NDRE response over different N-pre-pi was highly variable, that weakens the right point of reference.

The NDRE variability also depends to a great extent on the cultivar, and on the evolution of the temperature and light intensity of the growing season, and it is therefore not automatic to apply the correlation between NDRE and the N rate for PI fertilization of the previous years. The empirical method applied till now on the farm is: i) to consider the prescription threshold table applied in the same cultivar in the previous years; ii) to detect by a drone the gross average NDRE for this cultivar in the current year, and iii) to adapt the prescription threshold table to this parameter. It could also be useful to adapt the N rates to the weather forecast for the following 40 days, if it could be considered reliable.