Development of canopy vigour maps using UAV for site-specific management during vineyard spraying process

Site-specific management of crops represents an important improvement in terms of efficiency and efficacy of the different labours, and its implementation has experienced a large development in the last decades, especially for field crops. The particular case of the spray application process for what are called “specialty crops” (vineyard, orchard fruits, citrus, olive trees, etc.) represents one of the most controversial and influential actions directly related with economical, technical, and environmental aspects. This study was conducted with the main objective to find possible correlations between data obtained from remote sensing technology and the actual canopy characteristics. The potential correlation will be the starting point to develop a variable rate application technology based on prescription maps previously developed. An unmanned aerial vehicle (UAV) equipped with a multispectral camera was used to obtain data to build a canopy vigour map of an entire parcel. By applying the specific software DOSAVIÑA®, the canopy map was then transformed into a practical prescription map, which was uploaded into the dedicated software embedded in the sprayer. Adding to this information precise georeferenced placement of the sprayer, the system was able to modify the working parameters (pressure) in real time in order to follow the prescription map. The results indicate that site-specific management for spray application in vineyards result in a 45% reduction of application rate when compared with conventional spray application. This fact leads to a equivalent reduction of the amount of pesticide when concentration is maintained constant, showing once more that new technologies can help to achieve the goal of the European legislative network of safe use of pesticides.

was conducted with the main objective to find possible correlations between data obtained 26 from remote sensing technology and the actual canopy characteristics. The potential 27 correlation will be the starting point to develop a variable rate application technology 28 based on prescription maps previously developed. An unmanned aerial vehicle (UAV) 29 equipped with a hyperspectral camera was used to obtain data to build a canopy vigour 30 map of an entire parcel. By applying the specific software DOSAVIÑA®, the canopy 31 map was then transformed into a practical prescription map, which was uploaded into the 32 dedicated software embedded in the sprayer. Adding to this information precise 33 georeferenced placement of the sprayer, the system was able to modify the working 34 parameters (pressure) in real time in order to follow the prescription map. The results 35 indicate that site-specific management for spray application in vineyards result in a 45% 36 reduction of application rate when compared with conventional spray application. This  77 Target detection has been developed either by using advanced techniques, such as vision 78 systems and laser scanning, or by ultrasonic and spectral systems. Gil et al. (2007) 79 obtained a significant reduction in the total amount of applied volume (57%) using a 80 sprayer prototype with ultrasonic sensors able to measure the crop width variations and 81 to apply a variable dose rate according to the instantaneous measured vine row volume 82 (VRV), in comparison with a conventional and constant application volume rate.  However, even if the canopy characteristics can be defined using the methods outlined, it 104 is also clear that a certain amount of variability can be assumed to exist inside the parcel.

105
When a uniform canopy structure is assumed for the whole parcel, differences in the total 106 amount of pesticide arriving at the canopy can occur, which reduces the effectiveness of 107     The overall objective of this paper is to find a good correlation between data obtained 166 from remote sensing technologies and canopy characteristics. The hypothesis is that 167 NDVI is a good indicator of canopy vigor and consequently application volume can be      vegetation pixels could be masked out of the image. The NDVI threshold to create the 296 vineyard mask (Fig. 4b)  The spraying process began when all the parameters and information (canopy vigour map, 328 prescription map, and working conditions) were uploaded into the embedded controller.

329
During the spraying process the system recorded, information concerning the sprayer 330 position in the parcel, the applied volume rate, and the adjusted working pressure.

331
In order to simplify the process, the system was programmed to apply the same amount 332 of liquid on the two simultaneously sprayed rows, avoiding differences between left and 333 right side of the sprayer during the circulation over the rows.
where r is the expected value; and p is the obtained value (actual). were assigned (each representing an increase of 5% compared with the previous one). 352 The defined intervals ranged from zero to 50% deviation. Each point was compared and 353 quantified for its coincidence between p and r values. In addition, a determination was    This subsection presents and discusses the maps generated during the process. The 398 sequence of the obtained maps was as follows: (1) NDVI map, (2) canopy vigour map, 399 (3) prescription map, and (4) actual application map (Fig. 7).  Data obtained from the multispectral camera embedded in the UAV was used to generate 407 the NDVI map (Fig. 7a). This map shows how the intensity of colour was captured by the 408 camera, being the first step for determining the different canopy vigour zones.

410
Once the NDVI map was developed, all the data were appropriately managed and 412 classified in order to distinguish the three clearly different zones in the parcel. The three 413 zones were plotted on the map (Fig. 7b) (Fig 7b).  From that, the intended prescription map was generated (Fig. 7c). In this case, the 435 corresponding obtained values were 150 l·ha -1 for low canopy vigour, 206 for medium 436 canopy vigour, and 260 l·ha -1 for high canopy vigour.

437
Actual variable application map 438 Once the prescription map was embedded into the controller installed on the sprayer, the 439 spray application process started. During the process, data associated with the georeferenced position of the sprayer, actual working pressure, and forward speed were 441 automatically recorded and saved in the dedicated software. Following further processing 442 of the saved data the actual application map was generated (Fig. 7d)  The white lines observed in the actual application map (Fig. 7d) correspond to internal 451 roads in the parcel. As the spraying process was continuous, in those zones without the 452 presence of canopy, the spraying process was automatically turned-off according to the 453 signal detected by the ultrasonic sensors installed on both sides of the sprayer.

454
Quantification of the accuracy of the system 455 Following generation of the actual application map, the mathematical procedure outlined 456 below was used to evaluate and quantify the accuracy of the process. As will be seen, the 457 results obtained indicated that the developed system had exceptionally good accuracy, 458 quantified by the comparison between the actual spray application rate and the intended 459 application rate.

460
The obtained value of RMSE for the whole group of 100,000 sample points was 24.4. of 150 l·ha -1 (low case), the areas were the actual spray application rate ranged from 75 471 to 225 l·ha -1 were counted; for medium application rate (206 l·ha -1 ), the counted range 472 was from 103 l·ha -1 to 309 l·ha -1 ; and, for the highest intended spray application rate (260 473 l·ha -1 ) the measured range was from 130 l·ha -1 to 390 l·ha -1 . Table 4 shows the complete 474 range of thresholds applied during the accuracy evaluation process.    Figure 8 shows the spatial distribution of the accuracy in the parcel, classified according 486 the established threshold level, ranging from 0% to 50%. The dark zones on the maps 487 indicate the areas where the accuracy of the system exceeded the established thresholds.

488
The main percentage of dark zones corresponds to transition zones, where the variable 489 application sprayer was forced to modify the working parameters (working pressure) 490 while maintaining the forward speed. During the data processing, some outsider cases 491 were also detected. In a small number of points, differences greater than 50% between 492 intended and actual spray application rate were detected. A small percentage of the total 493 measured area (1.2%) was identified as the worst cases. Those zones correspond to values 494 were the spray application rate (table 3) fell lower than 75 l·ha -1 (less than 50% of the 495 lower recommended application rate of 150 l·ha -1 ) and were higher than 390 l·ha -1 (50%   The actual spraying application map obtained following the variable rate application 510 procedure, was compared with the standard application map based on a constant volume 511 rate of 325 l·ha -1 , the normal volume rate selected by the farmer for conventional spray 512 application. For those two scenarios, the total time for the spray process, the amount of 513 water, and the number of tanks to be filled were calculated, and the hypothetical amount  Table 5 shows the absolute and relative values for the following cases: conventional 519 spray application, variable rate spray application, and variable rate spray application with 520 ultrasonic sensors. In this last case, savings were also calculated for the specific zones 521 where the sprayer was turned-off (internal rows in the parcel) according the received 522 signal from the sensors.

523
The results clearly show the positive effect of the variable rate application process. The 524 total amount of liquid applied in the 5 ha parcel was reduced by 44.3% and 47.3% using 525 the developed site-specific management sprayer, without and with US sensors, 526 respectively. The corresponding saving in terms of time was approximately 45 min for 527 both cases, equivalent to circa 9 min·ha -1 . Finally, the potential savings on active 528 ingredient were 3.1 Kg and 2.9 Kg, with and without ultrasonic sensors, respectively. The results obtained in this study indicate that a bright future is ahead with the application

550
• The canopy vigour map was easily transformed into a prescription map by using the 551 dedicated decision support system DOSAVIÑA® 552 • It was possible to develop a specific software application to upload the prescription 553 map for a certain parcel of vineyard into a modified sprayer for the variable 554 application process. This will enable improved spray application for field crops, 555 which are widely disseminated, and is a novelty for 3D crops such as vineyard crops.

556
• Excellent accuracy was obtained with the system (demonstrated by comparing the 557 intended and actual application maps), with assumed tolerances of around 10% 558 deviation.

559
• The proposed method for accuracy quantification resulted in an objective, practical, 560 and useful procedure for those types of data.

561
• Savings on water and pesticide of over 40% were quantified. However, the saving 562 concerning the total amount of pesticide can be expected only for the cases where 563 dose recommendation on the pesticide label is based on concentration.

564
Overall, this study demonstrated that improvements arise from the combination of canopy 565 characteristics, intra-parcel variability, new technologies for variable application rate, and 566 the latest developments linked with the use of UAV in agriculture.