In 2018, three centers collaborated in the setting up and the realization of the calibration experiments (Table 1). Symbiotic inoculation was performed using a Micosat F ® bio-fertilizer, as coating (1 kg ha^-1) or granulate (10 kg^-1). A total of 52 plots were compared to establish the results on yield. Because some of the litter-bags and leaf samples referred to several plots, 13 Symbiotic inoculated (S) vs. Control (non-inoculated; C) pairwise comparisons were made after a grouping operation to develop the symbiotic model.

Two centers collaborated in the 2019 validation experiments. The inoculation was performed using a Micosat F ® bio-fertilizer and four AFM types as coatings (1 kg ha^-1).

¹Biota composition: finely ground cultivated Sorghum sudanensis roots, containing spores and ifae of Funneliformiscoronatus GO01 and GU53, F. caledonium GM24, F. intraradices GB67 and GG32, F.mosseaeGP11 and GC11, F. viscosum GC41; saprotrophic fungi: Streptomyces spp. ST60, Streptomyces spp. SB14, Streptomyces spp. SA51, Beauveria spp. BB48, Trichoderma viride, Trichoderma harzianum TH01, Trichoderma atroviride TA28, Trichoderma spp.; rhizosphere bacteria: Bacillus subtilis BA41, Pseudomonas fluorescens PN53, Pseudomonas spp. PT65 and Pochoniachlamidosporia, in a relative percentage of 40% crude inoculum and 21.6% bacteria and saprotrophic fungi; ²Rhizophagus intraradicesCA502;³Gigaspora rosea NY328A;⁴Sclerocystis sinuosa MD126; ⁵Claroideoglomus claroideum ON393.

As previously described 18, hay litter-bags were buried along the plant row. After about sixty days, the probes were opened and dried at a temperature of ~ 65 ° C for 36 h and the residues were examined by means of a smart NIR-SCiO^TM (Consumer-Physics, Tel Aviv) operating in the 740-1070 nm NIR band; two scans were carried out on each side of the litter-bags. The chemometric elaborations were carried out by means of the SCiO^TM-Lab software, which operates by means of AKA (As Known As) matrices and provides a percentage recognition of the classification matrix cells. The algorithm used for the classification was the random forest algorithm. The percentages of fingerprinting for the Control (CC) and the Symbiotic (SS) classes were analyzed using an online MedCalc software, and one proportion was tested.

An NIR-SCiO^TM equation, taken from unpublished results of an experimental trial on tomato plants, was used to obtain an indirect estimate of the soil induced respiration (SIR) capacity, which was measured according to Anderson and Domsch23. After having added sugar to the soil sample, it was possible to measure the aerobic activity, stimulated by the microbial consortium, using the infrared meter. The correlation between the estimated and measured data resulted to be sufficiently high (R²_{cross-validated} 0.86) to be considered reliable under comparable conditions.

The NIR spectrum of the leaves was detected using the smart NIR-SCiO^TM in the lower leaf page, in duplicate. Coding samples were obtained and a chemometric elaboration was carried out, as described for the litter-bags.

The raw foliar pH measurements were carried out, according to the acquisitions of previous works192021, on cut leaves, using a BORMAC "XS pH 70" pH meter (www.giorgiobormac.com), with a pH range of 0 ÷ 14, two decimals, and supplied with a Hamilton Peek Double-PoreF, / Knick combined 35 x 6 (LxØ) glass-plastic electrode. The measurement was performed on 15 leaves per sub-group, in the central rib in the lower page, using a screw to slightly dent the rib in order to insert the electrode. Other types of electrodes have been found to not be suitable for this kind of measurement. The pH operation was conducted on fresh or conserved leaves and chained with an NIR scan.

Individual data of the raw foliar pH and of the soil respiratory capacity were analyzed by means of mixed models 24 with the bio-fertilizer and C-S pairwise comparisons as the fixed factors, while the experiment was considered as random.

Yield data from the Control and Symbiotic sub-plots and their effects were analyzed using the Friedman test for paired comparisons 25.

In order to model the dependent variable, as an effect- size, the yield was expressed as the Ln of the response ratio (S/C), where the mean of the inoculated treatment (S) was divided by the mean of the non-inoculated control (C), that is, d_Y = Ln(S/C). This mode of expression is arithmetically equivalent to calculating the relative prevalence of S over C (d _Y= S/C -1), with the result being given in percentage. Seven independent variables, including the squared values of the litter-bag fingerprints, were fitted to the response ratio 25 in a multiple linear regression model. In order to make the relationship positive, the average pH values were transformed into H⁺ as (10^-pH*10^6).

The average results of the six independent parameters issued from the two validation experiments were applied in the model and the dependent predictions were compared to the realized S/C yield deviation. A positive-negative quadrant classification was performed for the more extensive DISAFA 2018-2 experiment.

The average yield measured in the fifty sub-plots of the calibration set was 14.47 t ha^-1 for C vs. 15.53 for S (Table 2 and Figure 1). The relative difference spanned from +33.6% to -15.6%, with a prevalence of +7.3 %, which is globally significant for the Friedman test (P=0.03). In fact, the responses to inoculation were different in the three experiments; the inoculation was not positive in CREA-CI or Maïsadour, while in DISAFA-1, the average yield increased by +11.3% (P=0.01). In the first 2019 validation experiment (Table 3), the yield response was positive (+4%), but the average response to the five types of bio-fertilizers was negative (-3.2%; P 0.11) in the DISAFA-2 experiment and nearly significant for the Shaniya cultivar (P 0.07). Table 4.

The average mixed model solutions were 5.34 for C vs. 5.14 for S (P<0.0001, Table 5) for all the single measurements, and an acidification of 3.7% units of pH (P<0.0001) was observed after inoculation with the BF. Table 6 reports the thirteen pairwise plot yields and the pertinent individual average values of foliar acidity and soil respiration. As far as leaf acidity is concerned,the mean increase as a result of the S treatment was 13.8% (P=0.0066), in terms of H⁺, a result that was derived from four significant increases and one significant decrease (in pairwise 12). The soil respiration activity was significantly increased in two pairs, but it was also decreased in three cases, and on average it was not relevant.

In the 2018 calibration experiments (Figure 2), the regression was only negative for the S plots. As highlighted in Figure 3, which is scaled in the same way as Figure 2, the pH range was more limited in the 2019 DISAFA-2 validation experiment, but a positive regression basically appeared for both the C and the S plots.

The increase in yield was favored where the respiration of the soil increased , as can be seen in Figure 4. This result was clear after the exclusion of two outliers (in red), and a parabolic positive trend was noted.

On average, the leaves of the C and S groupes were classified at a similar sensitivity level, that is, at 73.6 and 72.9%, respectively, for the C and S categories (Table 7). When considering the NIR-Tomoscopy and the yield (Figure 5), the linear regression in the fingerprint % of the leaves of the Control was slightly positive, while that of the Symbiotic was slightly negative. No deviation from linearity was apparent.

A high sensitivity was reached for the AKA classification of the litter-bags (86.0 and 87.6%, respectively, Table 7). When the NIR fingerprint of the litter-bags and the yield were considered together (Figure 6), the regressions in the fingerprints of the Control and of the Symbiotic probes were markedly negative, but a parabolic trend enhanced a maximum of the yield response by around 80-85% for the fingerprinting, and this was further confirmed for the CC litter-bags.

After excluding one outlier, the d_Y model reached an adjusted R2_adj. of 0.78 (P=0.01), with a standard error of ±4% (Table 6; Figure 7). The relationship with the yield increase was driven by the fingerprint in the Control litter-bags (std coef. = 11.85) and by its squared term (-11.77). Table 8

The first experiment in 2019-1, which was based on 75 pH and a leaf examination plus 40 NIR SCiO scans of 20 litterbags, has shown a good agreement between the yield predicted by the model (7.0%) and the yield realized in the field (4.2%). Both values are in the same quadrant of Figure 7 near the diagonal.

In the 2019-2 DISAFA-2 experiment, where a relevant body of measurements were conducted, that is, nearly two thousand (730 leaves for pH and NIR-Tomoscopy plus 350 scans of litter-bags), the average regression response was null (Figure 8), but the classification in the quadrants (Table 9) showed that the general negative response to bio-fertilizers was correctly predicted by 84% in the “bad” quadrant (P 0.0012) and the general means collimated (-2.66% predicted vs. -3.02% realized).

A significant, albeit highly variable positive response of maize yield to bio-fertilizer was observed (+7.3%) in the calibration, but the response to bio-fertilizers was disappointing in the validation experiments. In literature, maize-grain responses in yield were   positive in some cases (+6.4% 26; +4÷30% 3; +10% 19; +18% 27; +20% 28) or null 29. Mazzinelli 30 observed a +5% increase from a Mycorrhizal consortium in three experiments at a farm level in 2018. Thanks to the great variability of our 2018 cornucopia experiment, an attempt may now be made to explain some of the involved factors by means of the new methods that have been adopted here.