Development and evaluation of a decision prediction tool for the reduction of fungicide applications for the control of strawberry powdery mildew epidemics

Strawberry powdery mildew (Podosphaera aphanis) causes serious losses in UK crops, potentially reducing yields by as much as 70%. Consequently, conventional fungicide application programmes tend to recommend a prophylactic approach using insurance sprays, risking the development of fungicide insensitivity and requiring careful management relative to harvest periods to avoid residual fungicides on harvested fruit. This paper describes the development of a prediction system to guide the control of P. aphanis by the application of fungicides only when pathogen infection and disease progression are likely. The system was developed over a 15-year period on commercial farms starting with its establishment, validation and then deployment to strawberry growers. This involved three stages: 1. Identification and validation of parameters for inclusion in the prediction system (2004-2008). 2. Development of the prediction system in compact disc format (2009-2015). 3. Development and validation of the prediction system in a web-based format and cost-benefit analysis (2016-2020). The prediction system was based on the temporal accumulation of conditions (temperature and relative humidity) conducive to the development of P. aphanis, which sporulates at 144 accumulated disease-conducive hours. Sensitivity analysis was performed to refine the prediction system parameters. Field validation of the results demonstrated that to effectively control disease, the application of fungicides was best done between 125 and 144 accumulated hours of disease-conducive conditions. A cost-benefit analysis indicated that, by comparison with the number and timing of fungicide applications in conventional insurance disease control programmes, the prediction system enabled good disease control with significantly fewer fungicide applications (between one and four sprays less) (df=7, t=7.6, p=0.001) and reduced costs (savings between £35-£493/hectare) (df=7, t=4.0, p=0.01) for the growers.


Introduction
The yield of the strawberry crop in the UK has more than trebled from the same area in the period 46 1995 to 2019, with the overall yield increasing from 41,600 tonnes in 1995 to 141,600 tonnes in 2019, 47 giving a market value of £347.8 million [1]. This has been achieved through judicious use of short-day 48 (June bearers) and day-neutral (everbearers) cultivars, polythene tunnels (S1 Fig can range from 20% to 70%; where a 20% loss at 2019 prices would have given a market value loss of 55 £69.6 million [1,2]. Cultivar choice by growers is influenced by a number of agronomic qualities, with 56 resistance or tolerance to diseases being a low priority compared to aesthetic appeal, flavour and taste.

57
As the potential loss from strawberry powdery mildew is great, growers generally resort to insurance (or 58 prophylactic) spraying by applying a fungicide at one to two weekly intervals from April to 59 September/October (this applies for both June bearers, which may have two plantings a year, and 60 everbearers, with one planting a year), resulting in up to 24 fungicide applications per growing season 61 and hence a higher incidence of fungicide residues compared to insecticide residues [4]. Insurance sprays 62 cost money and can result in adverse environmental impacts, such as fungicide residues entering the soil 63 or waterways due to spray drift or run-off [5].

64
Strawberry plants are supplied to growers by specialist propagators. In the UK, most strawberry 65 crops are now grown in substrate (coir) on raised beds (S1 Fig a) or tabletops (S1 Fig d) or directly in the 66 soil on raised beds. The crop is irrigated and fertilised via drippers inserted into the substrate (fertigation).

67
There is a trend for annual planting to start in February, and there can be a very low incidence of P.
68 aphanis on these plants when they are delivered to the grower by the propagator. Crops for early harvest 69 are often covered with horticultural fleece to increase the temperature and prevent frost damage (S1 Fig   70  b); this also increases the relative humidity under the fleece. This intensive production process has both extended the length of the UK growing season (harvest from late May to September/October) and 72 increased the yield per hectare from 12.6 tonnes in 1999 to 29.7 tonnes in 2019 [2]. The use of polythene 73 tunnels and other agronomic practices aim to provide optimal conditions for the growth of the crop, but 74 it also provides ideal conditions for the growth of P. aphanis, risking epidemic development. Growers 75 therefore need a reliable and efficient prediction system as a decision support system (DSS) to warn them 76 when disease is most likely to occur and hence when to apply fungicides.

114
Within each field site, in-crop environmental conditions were recorded hourly using TinyTag Plus data loggers (Table 2) located within the crop canopy, with Tiny Tag Plus TGP-1500 for temperature and   116  relative humidity and TinyTag Plus TGP-0903 for leaf wetness. 117 Assessment methods used for strawberry powdery mildew 143 In stage 1, strawberry crops were assessed at the field sites for the progression of the incidence 144 of strawberry powdery mildew, caused by P. aphanis. Incidence (% plants affected) assessment was based 145 on the presence of leaf upward cupping, visible fungal mycelium and/or red blotches on leaves (Fig 2).

146
In spring (March-May) the fleece was removed from the strawberry beds and polythene covers were 147 replaced onto the tunnel frames, having been removed during the winter to mitigate potential storm 148 damage. Once this was completed, the regular assessment of plants began, until the symptoms of P.
149 aphanis were fully developed throughout the untreated polythene tunnel (no applications of fungicides), 150 so the incidence of strawberry powdery mildew was not inhibited by fungicides. Sensitivity analysis 200 To determine which of the new field parameters had the greatest and smallest influences on the 201 outputs of the strawberry powdery mildew prediction system, sensitivity analysis was done as previously 202 described [23,24]. In this case, the outputs of the strawberry powdery mildew prediction system were the 203 number of disease-conducive hours, as determined by a rule-based system. Sensitivity analysis required 204 keeping all the parameters constant, apart from the one being tested, which was altered in small 205 increments of 0.5°C or 5% (RH) over a range of values (Table 4). This was to identify the parameter(s) 206 that had the least impact on the prediction system, so that they could then be removed to simplify the 207 rules of the prediction system. The data from these analyses were used to create the rules that determine 208 the number of accumulated disease-conducive hours as the output of the prediction system. Table 4. Sensitivity analysis of the new field parameters (Table 3)  At the end of a given growing season, the number of high-risk periods when a strawberry grower 215 was advised to use a control product (e.g. fungicide) for P. aphanis according to the prediction system 216 was compared to the actual number of fungicide applications made according to the commercial routine 217 spray programmes. To achieve this, environmental parameters were recorded hourly with TinyTag data 218 loggers and downloaded to a computer for the duration of the growing season to provide input for the 219 prediction system. The actual number of insurance applications made to crops of strawberries were 220 obtained from the growers. At the field sites from stage 1 (Table 2), the growers managed disease control 221 according to accepted commercial practice which meant that they applied a fungicide when, based on 222 experience, they perceived the risk of P. aphanis to be high. By default, the strawberry growers followed 223 a customary routine spray programme and applied what they considered to be insurance sprays [25] 224 (spraying every 7-14 days regardless of disease severity) to mitigate the potential for P. aphanis to 225 develop in their polythene tunnels.

Development and validation of the prediction system as a decision
228 support tool (stages 2-3) 229 The development and validation of the rule-based prediction system to support on-farm decision 230 making involved two stages: the use of computer-based software with off-line infrastructure (stage 2) and 231 the development of an online, fully connected real-time web-based system (stage 3). A summary of the 232 validation process is presented in Table 5, with comparisons between these two stages. Table 5. Development and comparison of the prediction system over stage 2 (offline computer-based software) and stage 3 (online real-time web-based system).  (Table 5). By viewing the prediction system 249 output directly on-line, the grower could track the accumulation of disease-conducive hours in real time.

250
Initial validations of the real-time prediction system were done on commercial farms (Table 2)  to be suitable for the growth of P. aphanis when using the initial parameters (Table 3). If the initial 300 parameters had provided a prediction that was a good fit to disease development in the crop, then the 301 accumulation of disease-conducive conditions should have reached 100% just as the disease 302 development was at the start of the exponential phase. However, this did not happen, because strawberry 303 powdery mildew symptoms were observed before disease-conducive hours were predicted by the 304 system using the initial parameters (Fig 3, line with open squares). Therefore, field observations were 305 used to modify the initial parameters to give the new parameters (Table 3). These new parameters were 306 then used to calculate the accumulation of disease-conducive conditions for the development of P.

307
aphanis in the crop. In Fig 3, the line representing the new parameters (solid squares) reached 100% 308 just as the first visible symptoms of strawberry powdery mildew were observed on the strawberry plants.
309 As the prediction system is based on only those hours that are suitable for the initiation of germination 310 and growth of P. aphanis, the prediction system must necessarily depict the accumulation of disease-311 conducive hours as defined by the new parameters. 318 aphanis asexual cycle (i.e. the time it would take for a conidium to germinate, colonise and produce 319 more conidia, which is usually 144 hours of suitable conditions) was calculated when using the initial 320 and new parameters (Table 3) for the strawberry powdery mildew prediction system. The initial 321 parameters did not accurately predict the start of the epidemic; the disease progress (indicated by ) 322 was in the exponential phase and symptoms had already appeared when the cumulative percentage was 323 starting to increase. Therefore, using the initial parameters had not accurately predicted when to apply  (Fig 3), a sensitivity analysis was done. This determined which 359 of the parameters had the greatest and smallest influences on the total number of predicted completed 360 asexual life cycles (referred to as high-risk periods) of P. aphanis in the crop. As shown in Table 6, leaf   361 wetness had very little effect on the number of high-risk periods, when considered as a parameter on its 362 own or in combination with each of the other parameters. The parameter that had the greatest influence 363 on the number of high-risk periods was temperature. Following the sensitivity analysis, the parameters 364 were refined again by removing leaf wetness from the prediction system, since it had a limited impact 365 on the output. Table 6. Results of sensitivity analysis on new parameters presented in Table 3.

389
The initial phase of developing the strawberry powdery mildew prediction system, using Excel, 390 was done to demonstrate that it had the potential to reduce the number of applications of fungicides to 391 control P. aphanis. Once this had been done, it was possible to relate the prediction system with stages 392 of the asexual life cycle of P. aphanis to visualise the steps in the process (Fig 4).

Validation of the prediction system (stages 2-3)
395 Results of the validation of the prediction system over stage 2 (offline computer-based) and stage 3 396 (online real-time web-based) are presented in Tables 7 and 8. 397 Computer-based software 398 The validation results in stage 2 showed that crops from the prediction system field plots had 399 less disease (i.e. smaller AUDPC values) compared to those from the insurance spray programme 400 (commercial) plots and untreated control (Table 7). Growers reported a low incidence of disease in the 401 crop over the duration of the season; the prediction system gave good control (commercially acceptable) 402 of strawberry powdery mildew, as an epidemic did not develop. The prediction plot in 2014 received seven sprays compared to 10 in the commercial practice 410 plot (Table 7). In 2015, the prediction plot was sprayed only three times using two fungicide active 411 ingredients, compared to seven sprays with a total use of eight active ingredients in the commercial 412 practice plot. The prediction system offered good control with fewer sprays compared to commercial 413 plots, which also had higher levels of disease even though they were sprayed more frequently. Manual 414 downloading of the data from the Davis Pro2 TM sensors via the Weatherlink™ software and then 415 uploading it onto the prediction system software was required several times per week to obtain an 416 accurate prediction of the high-risk periods, and this was time-consuming for the grower. Importantly,

417
it has been successfully demonstrated that the prediction system had potential for commercial use if the 418 data transfer and integration could be made less cumbersome, making the process more user-friendly.

420
Real-time web-based software 421 The initial validation in stage 3 showed that responding to the monitoring of the disease-422 conducive hours using the prediction system, gave longer intervals (>14 days) between fungicide applications at the start of the season, enabling the grower to save sprays. Overall, cost savings 424 averaging £250 ha -1 were achieved at the sites where growers followed the prediction system (Table 8).
425 In the final validation, the prediction system was used on eight sites (six farms) resulting in 431 varying levels of success, with a control site (CB24) where it was not used to support the application of 432 fungicides but instead, the system collected the disease conducive hours in the polythene tunnels and 433 the grower recorded when fungicides were applied using their insurance spray programme (Table 9).

434
Some growers used the prediction system well (e.g. site HR8), whereas some did not use it to its full 435 potential as indicated by the reduction in spray applications as well as the overall cost savings, ranging 436 from £35/ha up to £493/ha. The results showed a significant difference (p<0.05) for both the number 437 of sprays (df=7, t=7.6, p=0.001) and overall costs (df=7, t=4.0, p=0.01) between the prediction system 438 and insurance spray programme. Every site, except CB24, was able to save at least one spray application 439 by following the prediction system, and the most efficient (HR8), was able to reduce applications by as 440 much as 50% compared to their insurance programme. The number of insurance sprays more or less 441 correlated with the level of risk perception by the grower and data clearly showed that a large number 442 of spray applications (e.g. sites CB24, ST18 and HR9) do not necessarily result in good disease control. e Poor control -disease present with visible symptoms. 450 f Fungicide applications were made between 115 and 144 accumulated hours of disease-conducive conditions. 451 g Fungicide applications were made at below 115 but above 50 accumulated hours. 452 h Some fungicide applications were made at or below 50 accumulated hours. 453 i The grower was given access to the system, but chose not to use, and followed his own insurance programme instead. Readouts from the prediction system gave good indications of how growers used the system as 455 well as their willingness and/or confidence to rely on its warnings to spray. Fig 5 shows that the best 456 grower (site HR8) recorded all fungicide applications made, by resetting the system to zero (Fig 5,   457 vertical lines) and three out of four sprays (Fig 5, Jul 20, Aug 5 and Aug 31) were performed between 458 115 and 144 accumulated hours (medium and high risk), while one fungicide (Jun 30) was applied at 459 80 accumulated hours (low risk). Using the prediction system extended the interval between the four 460 fungicide applications by 20 days, 16 days and 26 days, respectively (Fig 5), thus reducing the number 461 of applications made, compared to the insurance programme (7-14 days intervals). At site PH12, the grower used the predictions system reasonably well, but was not willing to 479 allow the system to run beyond 115 accumulated hours before applying a fungicide (Fig 6). In this case, 480 two fungicides (Fig 6, Jun 21 and Jul 10) were applied before 50 accumulated hours (low risk) and four fungicides (Fig 6, Jul 2, Jul 23, Aug 2 and Aug 15) were applied between 50 and 115 accumulated hours 482 (low to medium risk). This grower managed to extend the interval between applications by three days 483 throughout the season by using the prediction system, when compared to their insurance spray 484 programme of 10-day intervals.  was able to show that if the initial clean-up spray was neglected, as was the case with the grower at site 547 CB24 (Fig 8), visible disease will occur, and the prediction system becomes ineffective for disease 548 control. With correct use of the prediction system, as by the grower at site HR8 (Fig 5), we clearly 549 demonstrated that epidemics of P. aphanis could be controlled with fewer fungicide applications.
550 Without the prediction system to guide them, or where growers were not willing to completely 551 rely on the system, there was a tendency to overestimate how quickly P. aphanis was developing in 552 their polythene tunnels, as they were concerned about the economic impact of severe disease, which 553 caused them to use insurance sprays [25]. This behaviour was self-evident in the readout traces from 554 the system (e.g. Fig 7, site HR9), indicating the application of fungicides when the system was showing 555 a low risk of disease. This risk-averse attitude is also largely reflected in the higher frequency of 556 fungicide applications recommended by insurance programmes for disease control.

557
To help growers gain confidence in using our system, and to manage the aversion to risk, we 558 adopted a traffic light system to visually display the accumulation of disease-conducive hours with low 559 (green), medium (amber) and high (red) risk periods clearly identified (e.g. Fig 5). This made it easy 560 for growers to understand the data outputs and the practical implications. The use of the environmental parameters that had been developed and validated on commercial farms, rather than only in a laboratory, 562 not only improved the precision of the system but also increased the confidence of growers in using it 563 in commercial farm settings. The number of participating sites in the final validation (stage 3) increased 564 from three in 2018 to nine in 2019 across England and Scotland. For those who followed the system 565 well, good disease control was achieved. For example, site PE14 (Table 7) saw a lower disease 566 incidence (a reduction of up to 45% in the AUDPC value) for both everbearer and June bearer crops in 567 the prediction system plot compared with the untreated control. The expression of risk as disease-568 conducive hours enabled growers to understand that, for example, it was not necessary to frequently 569 apply fungicides according to an insurance programme when only 50 out of 144 disease-conducive 570 hours had been accumulated. In addition, the slope of the line in the graphical display output enabled 571 the growers to follow the rate at which the disease risk was changing from medium to high risk, which 572 allowed them to decide when to schedule a fungicide application and/or harvest window. 573 The prediction system has been tested on a range of cultivars (with different levels of disease 574 susceptibility), with a selection of growing methods, at a range of geographical locations in both 575 England and Scotland, using different types of weather sensors (Table 2), and proven to be effective in 576 supporting growers to prevent epidemic build up and to control disease ( . Initially (stage 1), encouraged by the increase in availability of personal computers 591 on farms, we devised a process that required manual retrieval of relative humidity and temperature data 592 from the data loggers and transfer to an Excel spreadsheet, to compute the disease risk. Growers found 593 this to be too cumbersome and time-consuming. They needed to be able to access the data quickly, 594 easily and frequently and at that time, this was addressed by using WeatherLink TM enabled Davis 595 weather stations for automated temperature and relative humidity data acquisition, and this was 596 combined with the prediction system software on a local CD (stage 2). Unfortunately, the manual data 597 download from the WeatherLink TM cloud computers and upload to the CD format did not sufficiently 598 improve usability of the prediction system. The move to online real-time, wi-fi enabled technology in 599 stage 3 rendered data acquisition and processing fully automatic and greatly improved the usability of 600 the system. It enabled growers to quickly access the system online with relative ease and observe how 601 disease-conducive hours were accumulating in real-time.

602
One of many factors that influences the adoption of a DSS is the growers' willingness to take The work reported here shows that a strawberry powdery mildew prediction system has been 640 developed in the UK and works well on commercial farms. The system has been proven to be reliable, 641 simple and easy to access. It serves as a decision support tool for informing decision making, and has 642 been shown to be effective in guiding precise timings of fungicide application (i.e. with reduced