A field monitoring of three clusters of different vigor, assessed in the summer of 2019, was conducted in the summer of 2020 in commercial vineyards with the aim of searching for sky-earth correlations to advance precision agriculture in vineyards. The Litterbag-NIRS method was applied, in parallel without soil chemical analyses, to correlate the yield of the identified production clusters. Moreover, by assessing a fingerprint in the electromagnetic spectra that was available from previous templates, it was possible to unravel some of the microbiological activities and soil traits that are in general favorable for the yield.

The trial was conducted in the western Po valley in three Erbaluce (CS, CS and EV) and two Nebbiolo (NG, NV) vineyards where three clusters had been identified, and five vines were observed in each cell (Table 1).

All the vineyards are on a hilly area of glacial morainic origin, with soils that show a tendentially acid / sub-acid reaction and an important presence of skeletons that greatly limit the workability of the soils. It is therefore normal to observe perennially grassed vineyards in this area.

The three Erbaluce vineyards were pruned following the traditional system, which involves pergola training with long and rich pruning, characterized by 2/3 fruit heads per plant and the same number renewal spurs in the following years. The CS and DS vineyards are located on a slope with South West exposure and have plant densities of 1.111 and 1.850 plants per hectare, respectively. Both are farmed organically with fertilization being carried out every two years with organic fertilizers based on poultry manure (N, P and K contents of 6 - 8 and 15 %, respectively) and with pest management based exclusively on copper and sulfur salts and mechanized grass management. The EV vineyard is flat and it is managed in a conventional way, with mineral fertilization based on 250 kg ha^-1 of potassium and magnesium sulfate (K, Mg compositions of 30 and 10 %, respectively) spread in autumn 2019 and 166 kg ha^-1 of organic/mineral fertilizer in spring 2020 (N, P, K compositions of 10 – 5 and 14,5 %, respectively). Moreover, an integrated pest and soil management, which includes three mechanical grass mulchings between the rows and two chemical weedings of eache row, is adopted during the vegetative period. The three vineyards are adults with vines of between 15 and 25 years.

The two Nebbiolo vineyards are pruned using the mono lateral guyot system with espalier training and a density of 3.460 plants per hectare. The NG vineyard is flat, with an average age of the vines of less than 10 years, while the NV vineyard is ion a steep slope and is West oriented; the plants are adult. In both cases, the vineyard management is conventional, following an integrated pest management protocol and had not been fertilized in the three years before the research. During the vegetative period, the soil management involves three grass mulchings between the rows and three mechanical processings under the rows.

A herbicide trial was performed in the conventional EV vineyard using a Glyphosate-based product.

The determinations were performed following the phenological growth stages and BBCH identification keys of grapevine (Table 1)14.

In order to characterize the structure of the canopy of the plants under study, the leaf layers were determined using the “point quadrat” method 15. Some vigor and foliage density indices were also determined, albeit only for the Nebbiolo plots (Table 1), using an application developed by the University of Adelaide. “VitiCanopy” (https://www.plantransig.com/techniques/viticanopy/) is a free tool that growers, irrigation practitioners and scientists can use to quickly and reliably assess spatial and temporal growth and canopy architecture dynamics. This can then be associated with the final yield and grape quality. The app was released in September 2015, and was funded by the Australian Grape and Wine Authority.

This application, which can be used on an Iphone, was designed to obtain digital images of the canopy, which are then processed through a special algorithm that relates the full spaces of the wall to the empty ones. The application returns a series of indices that are closely related to the plant vigor 16. However, this app cannot be used for horizontal training systems (as for the pergola system) and for this reason the Erbaluce plants were excluded.

A total of 100 berries were collected randomly from each plant during the harvest and subsequently taken to the laboratory.

A total of 20 berries were taken for each sample and were used to determine the average weight of the berries.

The grape sample was then manually pressed, filtered, and brought to a temperature of 20 °C. The degree brix and the pH were determined on the liquid fraction using a bench refractometer and pH meter.

On 20/05/2020, 5 litterbags per cell were buried at a depth of 5-10 cm. Each litterbag was filled with hay for small animals (“Vita Verde Small Animal Alpine Hay”, produced by Vitakraft pet care GmbH & Co. KG, Bremen, Germany), ground to 3 mm. About 2 g of hay was packed into half empty 5x10cm square polypropylene nets (1.5 mm mesh), which were resealed using 4 staples, and a plastic label was applied for identification and for easiness of finding purposes. The litterbags were explanted after 60 days, sun dried, gently cleaned and preserved at room temperature until delivery. The brushed litterbags were opened and the surfaces of both sides were examined, in reflectance mode, protected by a magnetic spacer capsule, measuring 9*40 mm, of a smart miniaturized NIRS web-based wireless spectrophotometer (SCiO v. 1.2, Consumer Physics, Tel Aviv, Israel) 17,operating in the 740-1070 nm range. Three spectra were scanned from each litterbag. The NIR spectra were downloaded from the SCiO^TM repository, and then imported in WinISI II v1.04 (FOSS NIRSystem/Tecator, Infrasoft International, LLC) software compatible format. The spectra of each cluster were averaged, and the set was then subdivided into Erbaluce, with 8 clusters (the litterbags for one cluster of Erbaluce were missed) and Nebbiolo, with 6 clusters. The average yield results from each available cluster were then used separately to calibrate the average NIR SCiO spectra for the two vines, using the modified partial least square procedure; cross-validation was applied and elimination of the outliers was permitted with one passage, with threshold t =2. The original reflectance values (331 points) were used without any math pretreatment.

The chemical composition of the litterbag residues pertaining to each spectrum was predicted using templates assessed under WinISI format in an experiment on biofertilized tomato (9Table 2). The predicted values were averaged per vineyard and cluster (14 cells) and a partial least squares analysis was then carried out on the averages to identify the variables with the maximum response – positive or negative – to yield variations. The XLStat 2019.4.1 (Addinsoft) package was used for this purpose.

On June 20, samples of 10 leaves, randomly chosen from each cluster, were analyzed for petiole pH using a Hamilton Peek Double-Pore F, / Knick combined 35 x 6 (LxØ) glass-plastic electrode, two decimals, and an NIR-SCiO^TM smart device, as described in Masoero et al.18, with two replicates for each leaf. The foliar pH data were analyzed, according to a bifactorial linear model, with vine and vineyards within vine as the fixed effects. Moreover, since the pH value of each leaf was known, a calibration of the single spectra was attempted. As explained above for the litterbags, the composition of the fresh vine leaves was predicted from NIR spectra using PLS models derived from an experiment with mycorrhizae in Sorghumsudanensis19and in a study on the green pruning residues of grapevine 20. The predicted average values were elaborated in the same way as before for the litterbags to establish which properties were connected the most with the realized yield. The information from the foliar spectra was synthesized in four bands, equivalent to those available from the Sentinel-2 satellite remote sensing services, namely b6(750nm), c7(793nm), c8a(875nm) and d9(955nm).

Chemometrics of the 331-point NIR spectra was performed using the SCiO^TM Lab proprietary software, by means of a classification procedure based on a random forest algorithm. The reflectance spectra were mathematically transformed as standard normal variates, Log and 1^st derivate, and the classification then produced an AKA (as known as) confusion matrix on the basis of belonging to a Herbicide or Not-herbicide class within one of the Erbaluce vineyards. The calibration model assessed in that vineyard was then validated in the other two all-organic Non-herbicide Erbaluce vineyards and two similar organic Nebbiolo vineyards. The probability vs. threshold of 50% was tested using the online version of Med-Calc for percentage comparisons.

As expected , Erbaluce produced 112% more than Nebbiolo, but the foliar pH was unusually 10% higher (3.23 vs. 2.93), which, however, is a large amount.

Means with different letters are different at P <0.05; * Two outliers found for the within-vine group linear regression of the yield on pH (Figure 1); § one outlier found for the PLS regression of the yield on the Sentinel-2-like bands

The yield and foliar pH were correlated, but the trends were different for the between or within vine regressions. In fact, when considering all 15 cells (Figure 1), the general regression of the yield on foliar pH was positive (red line), but after excluding two intermediate cells (Table 1), the within vine regressions were clearly negative and of a similar amount (-5.15 kg/pH in Erbaluce and – 5.63 kg/pH in Nebbiolo).

The Nebbiolo leaves were 8.7% more reflective than the Erbaluce ones (Figure 2).

The prediction of the yield from the single foliar NIR spectra, by means of the PLS model, reached an R-square cross-validated (R²cv) value of 0.66 for the overall cases (Table 2), with a closer fitting for Erbaluce than for Nebbiolo (Figure 3).

Several other traits were positively correlated with the foliar NIR spectra (Table 3), inprimis, the Leaf Area Index (0.85), the Canopy cover (0.93), the clumping index (0.66) and the crown porosity (0.84) for Nebbiolo. Some vegetational traits were also positively correlated concerning the gems (0.64 in all for Nebbiolo), inflorescences (0.67 in all) and fertility for Erbaluce (0.68).

The spectral information of the leaves, as available in the four Sentinel-2-like bands, or implemented in the PLS models with the foliar pH, is presented in Table 4 for all the vines together and separately. Considering all the vines, the models were able to attain an R²cv value of 0.77 using the five predictors and 0.73 with only the Sentinel-2 variables (Table 4, Figure 4).

The prediction of yield from the Litterbag-NIR spectra, through the PLS model, reached an R-square cross-validated (R²cv) value of 0.72 for the overall cases (Table 5), albeit with a lower fitting for Erbaluce (0.67, Figure 5) than for Nebbiolo (0.91; Figure 6).

The pruning wood was also closely correlated to the litterbag spectra (Table 4, Figure 7), with R²cv 0.83.

Some positive relationships emerged for the vegetational and fruiting characteristics of Erbaluce, for the leaf layer (0.94) and inflorescences (0.65), especially for the Nebbiolo vine, and for the canopy cover (0.72), the crown porosity (0.65) and the gems (0.93 and 0.77), especially for the pH of the grape (0.91), but less for the Brix (0.61).

Among the 20 variables predicted by means of the litterbag-NIRS method described in a previous paper 9, only a few (four) were clearly responsible for the yield variation (Table 6), and the same four variables emerged for both vines. The favorable correlation with yield depended above all on the soil Substrate Induced Respiration (SIR) expression of the C microbic, and then on the protein residual in the litterbags. On the other hand, the yield was favored when the litterbags showed low amounts of Acid Detergent Fiber (ADF) and the soil showed a lower presence of NO₃^--N.

The variables of the decomposing hay and of the soil, as predicted by the Litterbag-NIRS method, adopting the templates presented in Baldi et al.9 , are shown in Figure 6, together with the subjects of the study, namely the five vineyards with either three or two clusters. The first principal component separates the young Nebbiolo (ng) from the Erbaluce (cs, cd, ev) and old Nebbiolo (nv). The second principal component instead divides the rapid variables, that is, those that decrease in percentage over time and which are colored in red and placed mainly above the horizontal axis, from the resistant variables, that is, those that increase in percentage over time, which are colored in black and placed below the horizontal axis. The rapid r-strategist microorganisms in the scheme of Figure 8 are in the upper part, while the slower k-strategists are in the lower part.

The variables linked to a more intense Mycorrizhal status (Myc-type) are associated with the young Nebbiolo, while the yield is obviously oriented in the Erbaluce clusters.

A relevant spectral signature of the weed treatment appeared in the Litterbag-NIRS spectra for the Erbaluce vineyard (Table 7), with a classification of 94% for the fingerprinting of the litterbag derived soils without Herbicide, but 62% (not significant) for the litterbags derived from soil with Herbicide.

This Non-Herbicide model was then validated in the other organic (Non-Herbicide) vineyards. The results were statistically significant in two out of four cases, equally divided between Erbaluce and Nebbiolo.

The Herbicide in the framework of the weed treatments tended to elicit the total digestibility of the litterbag residues, while it apparently tended to depress the microbic C and its respiration activity in the soil (Table 8).