Temperature data.

DDRP requires daily T_min and T_max data in a gridded format for an area of interest in CONUS (Fig 1). For real-time modeling, we have been using daily T_min and T_max data at a 4 km spatial resolution from the PRISM (Parameter-elevation Relationships on Independent Slopes Model) database (available at https://prism.oregonstate.edu) [21]. Daily PRISM data become available ca. 1 day after weather station observations are reported, and are typically updated and improve in quality as more observations are added (see PRISM website for details). The phenology mapping system of the USA National Phenology Network [8] uses Real-Time Mesoscale Analysis (RTMA) weather data at a 2.5 km resolution, which are available within hours after data are observed. The daily T_min and T_max RTMA data set could potentially be used in DDRP; however, the RTMA methodology lacks PRISM’s update and quality control regimes [21]. Another alternative is Daymet v3, which offers daily climate data for North America, Hawaii, and Puerto Rico at a very high spatial resolution (1 km) (https://daymet.ornl.gov) [22]. However, Daymet data are released months after the end of each year, so they would be less useful for within-season modeling and decision support.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic of the DDRP model framework.

1) Input data sets (blue shaded boxes) include a) data on the developmental requirements, climatic tolerances (optional), and emergence times of population cohorts of a species (Table 1), and b) daily minimum and maximum temperature (T_min and T_max, respectively) data. 2) Hollow blue boxes indicate calculations conducted on each daily time step, where a dashed outline represents calculations for climatic suitability. Phenological event map (PEM) calculations for each life stage (E = egg, L = larva, P = pupa, A = adult) are shown in green font. A full generation is counted when adults lay eggs (in red), and the number of generations subsequently increases. 3) After the daily time step completes, DDRP combines the results across all cohorts and exports the model outputs as multi-layer raster (“.tif”) and summary map (“.png”) files (orange shaded boxes). Orange shaded boxes with a dashed line represent model outputs for PEMs and climatic suitability.

https://doi.org/10.1371/journal.pone.0244005.g001

For forecast modeling, DDRP is currently configured to use either monthly-updated, daily-downscaled NMME (North American Multi-Model Ensemble) 7-month forecasts at a 4 km resolution [23], or recent 10-year average PRISM daily data that are calculated on a bi-monthly basis. We consider 10-year average data to be an improvement over 30-year climate normals for producing forecasts because temperatures in CONUS have significantly increased over the past 30 years [24, 25]. The match of mean forecasts of the NMME model’s ensemble to the observed value (i.e. skill) varies both spatially and temporally due to topography, season, and the presence of an El Niño-Southern Oscillation (ENSO) signal [26, 27]. It may therefore be more conservative, and provide more consistent predictions, to use 10-year averages instead of NMME data to avoid potential issues with skill. However, we caution that the 10-year average data do not simulate variation in daily T_min and T_max, which may result in the under-prediction of degree-day accumulation in the spring or fall as daily T_max only slightly exceeds the lower developmental threshold of a species, or for cooler sites that have temperatures that are often near the threshold. We have also prepared and plan to use the National Weather Service gridded National Digital Forecast Database (NDFD) 7-day forecasts (https://www.weather.gov/mdl/ndfd_info) [28] for use in DDRP.

Phenology modeling: Species data and parameters.

The life history and behavior of a target species must be considered for appropriateness to model in DDRP. In its current form, the platform can model four separate life stages (the egg, the larva or nymph, the pupa or pre-oviposition, and the adult) plus a separately parameterized overwintering stage. As movement and migration are not handled by DDRP, it is currently limited in its ability to model migratory species, such as those that may establish in southern areas of their potential range and migrate yearly to more northern areas. Species that lack an overwintering stage, which are common in tropical and subtropical areas, may be difficult to model because the timing of first spring activities and stages present cannot be accurately estimated. Currently DDRP is entirely temperature-driven, so species whose growth and reproduction are strongly influenced by additional environmental factors such as day length or moisture may not be accurately modeled.

DDRP requires data on the developmental temperature thresholds (in either degrees Celsius or Fahrenheit) and durations for each life stage of an insect species in degree-days (Fig 1 and Table 1). These data are typically collected in the laboratory by measuring how temperature influences the rate of development, although data derived from season-long monitoring studies are also used [14, 29]. We round Fahrenheit values of thresholds to the nearest integer in all DDRP models because it allows for simpler communication of models to end-users, and it is a long standing convention for degree-day models in the United States. A different developmental threshold may be assigned to each stage, although we typically solve for a common threshold if differences across the stages are minimal. Additionally, applying common thresholds allows a modified version of the DDRP model to be used by the site-based modeling tool at USPEST.ORG (https://uspest.org/dd/model_app), which requires common thresholds across stages, and to more easily compare the DDRP and site-based model implementations for a given species. Presently, the model depends upon a fixed starting date such as January 1, specified by the user for the entire region of interest. The duration of the overwintering stage represents the number of degree-days that must accumulate from the start of the year for the stage to complete.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Species-specific parameters used in DDRP with corresponding values for Epiphyas postvittana (light brown apple moth) and Neoleucinodes elegantalis (small tomato borer).

https://doi.org/10.1371/journal.pone.0244005.t001

Users must specify the number of degree-days that are required for the overwintering stage to complete development and emerge for the growing season. These data are typically gathered using field monitoring studies, whereby the temporal distribution of emergence times and number of individuals that emerge on a given date is documented [e.g. 32]. Assigning a single value to the overwintering stage duration parameter would assume that an entire population develops simultaneously, which may not be biologically realistic because several intrinsic (with a genetic basis) and extrinsic (e.g. microclimate, nutrition) factors can produce variation in development rates within a species [19, 34]. Indeed, phenology models that incorporate developmental variability in a population may have increased predictive power [17, 19, 35]. DDRP therefore allows the duration of the overwintering stage to vary across a user-defined number of cohorts (groups of individuals in a population). Much of the intrinsic variability in insect development during a generation often occurs in the overwintering stage [36], although developmental variation may occur in any life stage [17, 37, 38]. DDRP uses five parameters to generate a frequency distribution of emergence times: the mean, variance, low bound, and high bound of emergence times, and the shape of the distribution (Gaussian or lognormal; Table 1). The platform uses these data to estimate the relative size of the population represented by each cohort, which initializes the population distribution that is maintained during subsequent stages and generations. Individuals within each cohort develop in synchrony.

Users may specify the timing (in degree-days) of phenological events that are important to their target system to generate phenological event maps in DDRP, which depict the estimated calendar dates of the event over a time frame of interest. We typically generate phenological event maps based on temperature data for an entire year so that events for multiple generations of each of the five life stages are modeled. For example, phenological event maps that depict when the overwintering stage would emerge may be useful for identifying start dates for surveillance operations for a species, whereas maps for subsequent generations could help with planning operations later in the year. The timing of phenological events may be based on life stage durations (e.g. the end of the egg stage signifies egg hatching), or on occurrences within a stage such as the midpoint or peak of oviposition or adult flight. Currently, one user-defined phenological event for each life stage for up to four generations may be modeled, although the platform could be modified to predict multiple events for each stage (e.g. first, midpoint, and end of the stage) for any number of generations.

Climatic suitability modeling: Species data and parameters.

Climatic suitability modeling in DDRP is based on cold and heat stress accumulation and requires data on temperature stress threshold and limits of a species (Fig 1 and Table 1). While parameter values may be estimated from laboratory or field experiments, such data are lacking for most species. Additionally, extrapolating laboratory data to the field to project accumulation of stress is difficult due to oversimplification of the number of variables and the temporal and spatial variation in natural environments [39]. We have been using the CLIMEX software [40] (Hearne Scientific Software, Melbourne, Australia), which is one of the most widely used species distribution modeling tools for agricultural, livestock and forestry pests and non-pests [6, 7], to assist with climatic suitability model parameterization in DDRP. Laboratory collected data may help with parameterizing a CLIMEX model; however, model parameters are fine-tuned and the model is fitted using observations from the species’ known geographical distribution [40, 41].

DDRP was designed to be complementary to CLIMEX in several ways to facilitate climatic suitability model parameterization, but the two programs also differ in several respects (Table 2). Both platforms use a process-based modeling approach in which parameters that describe the response of a species to temperature stress are included in calculations of climatic suitability. Certain model outputs, particularly maps of temperature stress accumulation, are therefore directly comparable. DDRP uses the stress accumulation method of CLIMEX in which cold and heat stress units begin to accumulate when temperatures exceed the cold and heat stress temperature thresholds, respectively. In both platforms, cold stress units are calculated as the difference between T_min and the cold stress temperature threshold, and heat stress units are calculated as the difference between T_max and the heat stress temperature threshold. However, DDRP uses daily temperature data and stress accumulates linearly over time, whereas CLIMEX uses average weekly temperature data and stress accumulates at a weekly rate that becomes exponential over time (each week’s stress in multiplied by the number of weeks since the stress first exceeded zero) [40]. These differences appear to be minor in using CLIMEX outputs and parameters to support calibration of DDRP parameters, as demonstrated in our case studies (see ‘Case Studies’). Similar to CLIMEX, DDRP uses a single cold and heat stress temperature threshold for all life stages, and stress units accumulate across the entire time period of interest (i.e. across all life stages and generations). We apply the cold and heat stress threshold of the life stage that would be most likely to experience the coldest or hottest temperatures of the year, respectively. DDRP assumes that stress could indirectly kill individuals by restricting their activity, or directly cause mortality through extreme cold or heat events such as a hard freeze.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Comparison of the characteristics, parameters, and outputs of climatic suitability models in DDRP and CLIMEX.

https://doi.org/10.1371/journal.pone.0244005.t002

Importantly, DDRP was designed to model climatic suitability based on daily current or forecast temperature data at fine spatial scales (e.g. a single state or region), which would give users insight into the potential risk of establishment or spread during a particular season or year. In contrast, CLIMEX is normally used to estimate a species’ potential distribution using coarse-scale (10ʹ and 30ʹ resolution) global gridded 30-year monthly climate normals centered on 1975 (1961‒1990) or future projections from selected global circulation models (GCMs) [42]. In theory CLIMEX could be used for real-time climatic suitability, but it has no native ability to import and process common gridded formats and is incapable of using daily resolution climate data. Thus, DDRP’s climatic suitability models are intended to improve the efficiency of surveillance and trap deployment at a relatively small focal area for a current or near-future time period, whereas CLIMEX models provide a more general and coarse-scale assessment of suitability based on averaged climate data.

Relying on real-time climatic suitability models for decision support on where to employ pest management and eradication operations for a given year or season is preferable to using models based on 30-year climate normals. A model that uses current climate data is more biologically relevant because the risk of establishment in an area would be affected by the conditions that a species physically experiences, not by averages of historical climate. Additionally, climate in CONUS is changing rapidly, so models based on climate normals may produce unrealistic predictions of present-day climatic suitability. Over the past ca. 30 years, the average annual temperature in CONUS has increased by 1.2°F (0.7°C), the number of freezing days has declined, and extreme temperature events have increased in frequency and intensity [24, 25]. Nonetheless, DDRP is not currently capable of including moisture factors in the modeling process like CLIMEX, so model predictions for moisture-sensitive species in very arid or wet areas should be interpreted with caution. We discuss the potential implications of generating a climatic suitability model based solely on temperature in the ‘Discussion.’

We compare CLIMEX’s predictions of temperature stress accumulation and overall climatic suitability to similar outputs in DDRP to help parameterize a DDRP climatic suitability model. Temperature stress thresholds may be calibrated so that predictions of cold and heat stress accumulation at the end of the year are spatially concordant with CLIMEX’s predictions. Climatic suitability in CLIMEX is estimated with the Ecoclimatic Index (EI), which is scaled from 0 to 100, and integrates the Annual Growth Index and the Annual Stress Index (all climate stress indices) to give an overall measure of favorableness of a location or year for long-term occupation by the target species [40, 41]. An EI ≥ 1 is often used as a threshold for defining whether a location is suitable for long-term survival, although an EI exceeding 20 or 30 (depending on the species) is sometimes used to indicate a highly suitable climate [40, 41]. As discussed in more detail in ‘Case Studies’, temperature stress limits in DDRP can be adjusted so that areas predicted to be suitable by CLIMEX are also included in DDRP’s prediction of the potential distribution.

Comparing DDRP climatic suitability model outputs to those of CLIMEX for model fitting purposes is naturally more appropriate when temperature data are derived from the same time period. We have therefore been using a PRISM T_min and T_max 30-year average data set centered on 1975 (1961‒1990) to match the time-schedule of the CliMond CM10 (also 1961−1990) world climate data set currently supplied with CLIMEX [42]. We temporally downscaled monthly PRISM estimates for 1961‒1990 because DDRP requires daily data and PRISM daily temperature data for years prior to 1980 are not available. For each month of a given year, a bilinear interpolation method was used to assign each day an average temperature value that was iteratively smoothed and then adjusted so that the monthly averages were correct.

Phenology model.

DDRP calculates daily degree-days over the specified time period using developmental temperature threshold information and gridded temperature data that have been cropped to the extent of the focal area (Fig 1). Currently DDRP has two methods to calculate degree-days: the simple average using an upper threshold with a horizontal cutoff, and the single triangle method with upper threshold [43–45]. The single triangle method is also used as a close approximation to the more complex sine-curve calculation method [45].

With the exception of phenological event maps, which are computed only for the last day of the daily time step, DDRP saves the following phenology model results for each sampled day:

1. Accumulated degree-days. While daily degree-days are calculated for each life stage, the cumulative degree-days are summed only for the first cohort of the larval stage, as these degree-day maps are representative for all cohorts and life stages. Accumulated degree-days calculated for larva will be the same for other life stages if common developmental thresholds are used.
2. Life stages. The life stage present (overwintering stage, eggs, larvae, pupae, and adults) for each cohort.
3. Number of generations. The current generation count for each cohort. If the model is run for an entire year, then the output for the last day of the year would represent the potential voltinism of the species. The generation count increases when adults progress to the egg stage (i.e. oviposition occurs).
4. Phenological event maps (optional). The timing of phenological events is estimated by computing daily degree-day totals from the gridded temperature data, and storing the day of year when an event threshold is reached. Event results are generated only on the last day of the daily time step (typically, the last day of the year) because the entire time period must be analyzed for all potential event days to be considered.

Climatic suitability model.

Independently from the phenology model simulations, DDRP calculates daily cold and heat stress accumulations and compares these estimates to user-defined moderate and severe stress limits to delineate the potential distribution of the species (Fig 1). We opted to use moderate and severe stress limits to reflect two distinct themes. First, they may provide a way to depict the potential for short term vs. longer term establishment. For most species, the potential distribution could be represented by areas where cold and heat stress have not exceeded the severe or moderate stress limits, as these should allow for long-term survival. DDRP depicts these areas with maps of cold stress exclusion, heat stress exclusion, and all stress exclusion (cold plus heat stress exclusions; Fig 1). Areas under moderate stress exclusion may represent temporary zones of establishment in which a species establishes only during a favorable season, such as after an annual migration event. Conversely, areas under severe stress exclusion do not allow for even short-term establishment. Typically we visualize exclusion maps calculated for the last day of the year (day 365) under investigation to provide insight into the potential distribution for an entire growing season. For CONUS, the northern range limit is typically delineated by cold stress and the southern range limit, if any, is delineated by heat stress.

Second, using two levels of stress may provide a way to represent uncertainty for estimating the potential distribution. As discussed in more detail in the ‘Discussion,’ several sources of uncertainty and error in the modeling process may bias model predictions, such as applying inappropriate parameter values, using climate data with low skill or poor spatial resolution, ignoring biotic factors such as species interactions, or ignoring non-temperature abiotic factors such as microclimate effects, moisture, and photoperiod [34, 46]. Defining the potential distribution as areas under severe stress only would typically provide a broader estimate than a definition based on both stress levels. While this approach may over-predict the risk of establishment, conducting surveys over too broad an area is probably better than surveying too small of an area, which may allow a new invasive species to establish and spread.

System and software requirements.

DDRP requires the R statistical software platform and can be run from the command line or within RStudio [47]. It takes advantage of functions from several R packages for data manipulation, analysis, and post-model processing. The raster package [48] is used to crop daily temperature rasters to the focal area, store and manipulate daily loop raster results, and process and further analyze results for each cohort. Many non-spatial data manipulations are conducted with functions in the dplyr, tidyr, and stringr packages [49–51]. The ggplot2 package [52] is used to generate and save summary maps of raster outputs, and options from the command line argument are parsed using the optparse package [53].

DDRP capitalizes on the multi-processing capabilities of modern servers and computers to run multiple operations in parallel, which is made possible with the R packages parallel and doParallel [54]. This significantly reduces computation times, particularly in cases where modeling is conducted with multiple cohorts and across large areas. For example, parallel processing is used to crop rasters for multiple dates, run multiple cohorts in the daily time step, and to analyze time step outputs for multiple days or files simultaneously. For very large areas (currently defined as the Eastern United States and CONUS), temperature rasters are split into four tiles and both the tiles and cohorts are run in parallel in the daily time step.

We recommend running DDRP on a server or computer with multicore functionality because certain processes are very memory intensive and may execute slowly or stall without parallel processing. The platform was designed to run on a Linux server, but we have also successfully used it on a Windows OS with eight cores. As an example, a model that was run on a Linux server with 48 cores completed ca. 3.5 times faster than it did on a Windows OS with eight cores.

Climatic suitability, voltinism, and phenological events in Epiphyas postvittana.

The light brown apple moth, E. postvittana (Walker 1863) (Lepidoptera: Tortricidae), is a leafroller pest native to southeastern Australia, including Tasmania [30]. The species invaded Western Australia, New Zealand, New Caledonia, England, and Hawaii more than 100 years ago [55–57], and has been established in California since 2006 [58, 59]. It poses a significant threat to agricultural production in the United States because it feeds on more than 360 host plants, including economically important fruits such as apple, pear, citrus and grapes [30, 31, 60]. For example, an economic risk analysis of E. postvittana to four major fruit crops (apple, grape, orange, and pear) in CONUS estimated an annual mean cost of US$105 million associated with damage to crops and control, quarantine, and research [60]. The CAPS program at APHIS conducts annual surveys for E. postvittana at various counties across CONUS.

A summary of phenology and climatic suitability model parameters used for E. postvittana in DDRP is reported in Table 1. We assigned all life stages a lower developmental threshold of 7.1°C (45°F) [30] and an upper developmental threshold of 31.1°C (88°F). Laboratory studies revealed small differences in the lower developmental threshold (< 1°C) across different life stages [30, 31]. The upper developmental threshold value is based on studies showing that all life stages cease development between 31‒32°C [30–32]. We derived life stage durations (in degree-days Celsius; hereafter, DDC) for E. postvittana based on an analysis of published data [30], which resulted in 127, 408, 128, and 71 DDC for eggs, larvae (females on young apple foliage), pupae, and adults to 50% egg laying, respectively (S1 Appendix).

We set the overwintering stage to larva because the predominant overwintering stage of E. postvittana in the United States are the late larval instars [61, 62]. We applied seven cohorts to approximate a normal distribution of emergence times that spanned 100 to 320 DDC (average = 210 DDC) based on a report that overwintering larvae at four sites in California required between 102 and 318 DDC to finish development [61]. This would correspond to the time required for mid-stage (3rd−5th instars, average 4th instar) female larval feeding on old foliage (0.45 × 494 DDC = 210 DDC), after a January 1 start date. The single triangle method was used to calculate degree-days.

We tested the hypothesis that DDRP can correctly predict the timing of first spring egg laying and generation length for E. postvittana by analyzing three monitoring data sets that were collected in and around the San Francisco Bay Area in California over a 12 year time frame (2008‒2009, 2011‒2014, and 2019‒2020). All three data sets were comprised of moth count data that had been collected via pheromone trap surveys on a bi-weekly or monthly basis by USDA APHIS or the University of California Cooperative Extension. For each data set, we estimated the difference between the date of the peak in first spring flight and DDRP predictions of the average date of first spring egg laying. This event was chosen because peak flight would likely happen at about the same time that peak egg laying is occurring [63]. Additionally, we used DDRP to calculate the number of degree-days that accumulated between the last peak fall flight and first peak spring flight for four winters (2011‒2012, 2012‒2013, 2013‒2014, and 2019‒2020), which should serve as a rough estimate of generation time. These estimates were compared to the generation time that is used by the DDRP model (823 DDC for the overwintering generation; 734 DDC are used for later generations that have young foliage to feed upon [30]). Model runs for each year used PRISM data and applied seven cohorts and a normal distribution of emergence times. Additional details about the data sets and methods for this analysis are presented in S2 Appendix.

We generated a CLIMEX model for E. postvittana using CLIMEX version 4.0 [40] to help parameterize the climatic suitability model in DDRP. The model applied a combination of parameter values (Table 3) derived from two previous CLIMEX studies of this species [64, 65]. However, we used a cold stress threshold (TTCS) of 3°C, which is lower than He et al.’s (2012) value (5°C) [64], and higher than Lozier and Mill’s (2011) value (1.5°C) [65]. We applied a top-up irrigation (additional simulated rainfall) rate of 2.5 mm day⁻¹ for the winter and summer season because irrigation mitigates the hot-dry climate that limits distribution of E. postvittana within CLIMEX. We fit the CLIMEX model using a data set of 393 georeferenced locality records from Australia (N = 317) and New Zealand (N = 76), which were obtained from GBIF.org (18th July 2019; GBIF Occurrence Download https://doi.org/10.15468/dl.a4ucei) and Nick Mills at UC Berkeley (pers. comm.).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Parameter values used to produce a CLIMEX model for Epiphyas postvittana (light brown apple moth) and Neoleucinodes elegantalis (small tomato borer).

https://doi.org/10.1371/journal.pone.0244005.t003

For model validation, we used locality records from 14,949 sites in California where the California Department of Food and Agriculture detected E. postvittana between 2007 and 2013, and from an additional 47 localities in California where it was positively diagnosed between 2011 and 2017. We removed localities that occurred within the same CLIMEX grid cell, which resulted in 140 unique localities. Of these, 95% (133/140) were in areas predicted to have high suitability (EI > 20), 4% (6/140) were in areas predicted to have low suitability (0 < EI < 20), and 0.007% (1/140) were in unsuitable areas (S1 Fig). These findings suggest that the CLIMEX model accurately predicted suitable climates at the majority of localities where the species is known to be established in CONUS, and it is therefore a useful tool for calibrating a climatic suitability model in DDRP.

In DDRP, we generated a climatic suitability model for E. postvittana using the daily downscaled PRISM T_max and T_min estimates for 1961‒1990 and calibrated model parameters in accordance with the CLIMEX model (Fig 3). Specifically, we compared maps of temperature stress accumulation, and adjusted temperature stress limits so that most areas predicted to be under moderate and severe climate stress by DDRP had low (20 > EI > 0) or zero (EI = 0) suitability according to CLIMEX, respectively. Additionally, we modeled climatic suitability for the species in California for each year between 2010 and 2019 and calibrated stress parameters to maximize the fit of the model to 144 georeferenced records that were collected over the same time period (N = 144 from GBIF, N = 77 from Nick Mills). For each modeled time period, we extracted climate stress exclusion values for presence localities in the California Department of Food and Agriculture data set to test our hypothesis that DDRP could correctly predict the potential distribution of the species. We removed localities that occurred within the same 4 km grid cell, which resulted in 872 unique validation localities.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3.

Predictions of cold stress, heat stress, and climatic suitability for Epiphyas postvittana (light brown apple moth) in CONUS produced by CLIMEX (A‒C) and DDRP (D‒F) based on 1961‒1990 climate normals. Climatic suitability is estimated by the Ecoclimatic Index (EI) in CLIMEX, and by combining cold and heat stress exclusions in DDRP. In DDRP, long-term establishment is indicated by areas not under moderate (excl.-moderate) or severe (excl.-severe) climate stress exclusion. Cold and heat stress units in DDRP were scaled from 0 to 1000 to match the scale in CLIMEX. CLIMEX maps were generated for this study based on parameters documented in Table 3 and have not been previously published.

https://doi.org/10.1371/journal.pone.0244005.g003

Finally, we modeled phenology and climatic suitability for E. postvittana in 2018 to provide insight into its potential voltinism, seasonal activities, and risk of invasion in particularly warm temperatures. The summer of 2018 in the United States was the warmest since 2012 and tied for the fourth-warmest on record (NOAA website https://www.noaa.gov/news/summer-2018-ranked-4th-hottest-on-record-for-us; last accessed 11/21/19). We generated a phenological event map that depicted the date of first egg laying by first generation females, because this activity is relevant to monitoring both eggs and the emergence of adults, which typically occurs two to three days prior to egg laying.

Climatic suitability, voltinism, and phenological events in Neoleucinodes elegantalis.

The small tomato borer, N. elegantalis (Guenée) (Lepidoptera: Crambidae), is native to South America and is distributed throughout the Neotropics including in Mexico, Central America, and the Caribbean [66, 67]. A major insect pest of tomato (Solanum lycopersicum), it also attacks fruits of other plants belonging to the family Solanaceae including eggplant, paprika, naranjilla, and green and red pepper [67]. There are at least 1175 recorded interceptions of the species from the United States, where it is considered a serious threat to agricultural biosecurity because it lowers tomato production in South America [68]. The CAPS program has conducted surveillance for N. elegantalis since at least 2011.

A summary of phenology and climatic suitability model parameters used for N. elegantalis in DDRP is reported in Table 1. We re-analyzed data from a laboratory study on the development of N. elegantalis on hybrid tomato (Paronset) at five temperatures [33] to estimate a common lower temperature threshold for all life stages, which involved adding a point to force the x-intercept to an integer value in degrees Fahrenheit. We weighted the analysis to select a common lower threshold for immature stages, which are the longest in duration, because this should produce the lowest error for the overall life cycle. The lower threshold values for immature stages were very similar to the overall egg-to-adult value of 8.89°C (48°F), so we chose 8.89°C as the common threshold instead of a higher one solved for the adult pre-oviposition stage (11.5°C). We estimated the duration for eggs, larvae, pupae, and adults to peak oviposition as 86, 283, 203, and 96 DDC, respectively. This analysis is presented in S3 Appendix.

Neoleucinodes elegantalis has no apparent photoperiodic response, diapause, or specific overwintering stage. In subtropical climates in Brazil, the insect remains active throughout the year if host plants are available [33]. We defined adults as the overwintering stage and used January 1 as the model start date for CONUS because few host plants would be available for immature stages at this time. We assumed that adult feeding and host search activities could begin immediately if temperatures are suitable, and that first egg laying would subsequently occur after the estimated pre-oviposition period of ca. 55 DDC. The durations of later events (first to peak oviposition, immature development, etc.) were estimated from previously published data [33, 69]. We applied seven cohorts to approximate a normal distribution of emergence times that spanned 0 to 111 DDC (average = 50 DDC) because overwintered adults begin finding hosts over this time frame, and we used the single triangle method to calculate degree-days. Unfortunately, we were unable to find monitoring data that were suitable for validating the DDRP phenology model for N. elegantalis. Population monitoring studies of the species have been conducted in Brazil; however, the species at these locations does not have a discrete overwintering stage so peaks in flight are more or less sporadic and not indicative of first spring activity or voltinism.

We used two different approaches to parameterize a DDRP climatic suitability model for N. elegantalis for CONUS. First, we used a similar approach taken for E. postvittana in using a CLIMEX model to help calibrate temperature stress parameters in DDRP. As detailed in S4 Appendix, we re-parameterized a previously published CLIMEX model for N. elegantalis [70] because it appeared to underpredict suitability in warmer areas where the species is known to occur. Briefly, this involved fitting a CLIMEX model using 228 locality records from the Neotropics and validating it using a separate set of 54 localities (S1 Table). CLIMEX correctly predicted suitable conditions (EI ≥ 1) at 96% of the validation localities (51/53) that fell within a unique grid cell (S2 Fig), which suggests that the model is robust at predicting climatic suitability. We therefore compared DDRP maps of temperature stress accumulation to those of CLIMEX, and adjusted temperature stress limits so that most areas predicted to be suitable by CLIMEX were also included in the potential distribution by DDRP (Fig 4). We considered any area that had an EI ≥ 1 to potentially be suitable because some localities occurred in areas that had only marginal suitability (S2 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4.

Predictions of cold stress, heat stress, and climatic suitability for Neoleucinodes elegantalis (small tomato borer) in CONUS produced by CLIMEX (A‒C) and DDRP (D‒F) based on 1961‒1990 climate normals. Climatic suitability is estimated by the Ecoclimatic Index (EI) in CLIMEX, and by combining cold and heat stress exclusions in DDRP. In DDRP, long-term establishment is indicated by areas not under moderate (excl.-moderate) or severe (excl.-severe) climate stress exclusion. Cold and heat stress units in DDRP were scaled from 0 to 1000 to match the scale in CLIMEX. CLIMEX maps were generated for this study based on parameters documented in Table 3 and have not been previously published.

https://doi.org/10.1371/journal.pone.0244005.g004

We further calibrated the DDRP climatic suitability model for N. elegantalis by fitting a model for Brazil using 83 locality records and daily gridded T_min and T_max data for 2005 to 2016 at a 0.25° (ca. 28 km) spatial resolution [71]. To our knowledge, Brazil is the only country within the range of N. elegantalis for which daily gridded meteorological data are available. We did not use temperature data for earlier years (1980 to 2004) because they had lower accuracy [71]. A random subsample of 70% of locality records (S1 Table) were used to fit the DDRP model (N = 58), and the remaining 30% (N = 25) were used for model validation. We modeled climatic suitability for each year between 2005 and 2016 and iteratively adjusted heat stress parameters to maximize model fit. Brazil represents a relatively warm part of the distribution of N. elegantalis, so we were unable to calibrate cold stress parameters. During this process, we also considered how well DDRP predictions of suitability and heat stress aligned with those of CLIMEX (Fig 5). Despite DDRP’s use of climate data for different time periods than CLIMEX for this analysis (i.e. individual recent years vs. 30-year climate normals), its predictions of heat stress accumulations and climate stress exclusions over several years were often spatially concordant with CLIMEX’s estimates (Fig 5). This finding provided some additional reassurance beyond the validation analysis that DDRP could correctly model climatic suitability of the species.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5.

CLIMEX (A) and DDRP (B) predictions of heat stress and climatic suitability for Neoleucinodes elegantalis in Brazil. The CLIMEX model is based on 1961‒1990 climate normals whereas DDRP models are presented for all years between 2011 and 2016. Locality records that were used to fit (blue triangles) and validate (black circles) the DDRP climatic suitability model are depicted. Both models predicted suitable conditions in regions where N. elegantalis is widespread, including in the states of Espírito Santo (ES), Goiás (GA), Minas Gerais (MG), Pernambuco (PE), Rio de Janeiro (RJ), Paraná (PR), Rio Grande do Sul (RS), Santa Catarina (SC), and São Paulo (SP). The pink and blue lines in DDRP heat stress maps depict the moderate and severe temperature stress limits, respectively.

https://doi.org/10.1371/journal.pone.0244005.g005

Finally, we modeled phenology and climatic suitability in DDRP using temperature data for 2018. We generated a phenological event map for the average date of the beginning of egg hatch of the first generation (Fig 2B). Predictions of egg hatch could enhance population control of N. elegantalis because this species is most vulnerable to pesticides before larvae enter the fruit of host plants [72].

Linear (degree-day) modeling.

DDRP uses a relatively simple degree-day modeling approach, whereas some platforms including ILCYM, phenModel, and devRate [87] offer complex functions to model nonlinear responses of insects to temperature. Degree-day models are ideal for multi-species platforms like DDRP because there are sufficient data to parameterize a degree-day model for most insect pest species of economic importance in the United States [18, 88]. Linear degree-day models are also readily calibrated and sometimes constructed entirely using field data, making them more practical for extension and decision support use [14]. Additionally, degree-day models require only daily T_min and T_max data (as opposed to hourly data for most nonlinear models), which are available at a high spatial resolution for CONUS from multiple sources including PRISM and RTMA. Nonetheless, it is important for users to recognize potential sources of error and lack of precision in degree-day models, such as their limited ability to accurately model development at supra-optimal temperatures [13, 14, 89].

Environmental inputs.

DDRP is intentionally parameterized in a simple, conservative manner, which will hopefully achieve the goal of a parsimonious balance of both model simplicity and accuracy [14, 90]. Nonetheless, DDRP is driven entirely by temperature, and therefore ignores other factors that may affect the development and distribution of insects such as photoperiod, moisture, dispersal, resources, disturbance, and biotic interactions [7, 91]. The potential consequences of this limitation will depend on the biology of the organism under study. For example, dry stress is a major factor restricting the current distribution of N. elegantalis in its native range [92–94], and it limits the distribution of E. postvittana both in its native range [30, 95] and in Southern California and Arizona [96]. Our CLIMEX model for N. elegantalis predicted higher dry stress in arid parts of the Southwest than the rest of CONUS (S4 Fig), which suggests that an absence of moisture factors in DDRP may result in over-predictions of climatic suitability in this region. However, this conservative-leaning error may in fact better reflect human manipulation of the landscape (e.g. greenhouse and irrigation usage) that may allow the species to exist in such regions. Future versions of DDRP that can process gridded moisture data and incorporate moisture stress factors into climatic suitability models may help overcome our current limitations in matching CLIMEX models, and may improve predictions for moisture-sensitive species such as N. elegantalis. Additionally, we are developing a version of DDRP that incorporates photoperiodically induced life history events such as winter diapause and summer aestivation, which builds on earlier phenology modeling work that estimated voltinism of photoperiod-sensitive insects [97].

Presumptive models.

Uncertainties regarding the accuracy of temporal or spatial predictions of invasive species that are not yet established is inevitable, in part because no validation data are yet available, and species interceptions do not imply establishment [7, 98]. DDRP models for species for which only presumptive models exist should therefore be used conservatively. For example, surveillance or management actions could be implemented in advance of predicted phenological events as a precautionary measure (e.g. installing traps even earlier than estimates for the earliest date of overwintering adult emergence). To potentially avoid under-predicting the risk of establishment, the potential distribution could be defined as areas not under severe climate stress as opposed to defining it using both stress levels. Additionally, climatic suitability models generated by DDRP could be combined with those produced using different modeling methods (e.g. correlative, semi-mechanistic, or mechanistic) to create a “hybrid” model, which may increase the reliability of predictions [7, 91].

Web platforms that support sharing of pest observations from the United States will be valuable resources for validating and increasing the predictive performance of DDRP models for species that are already established in CONUS. For example, the iPiPE and its sister platforms (http://www.ipipe.org, https://ipmpipe.org) have created a national information technology infrastructure for sharing pest observations in near real-time and contributing them to a national repository [99]. Similarly, the USA National Phenology Network provides a repository of plant and insect phenology observations contributed by citizen scientists [16]. The National Agricultural Pest Information System (NAPIS; https://napis.ceris.purdue.edu/home) currently has over 5.17 million records from pest detection surveys, and is another potential source of validation data.

Geographic variation.

Populations of an invading species may exhibit geographic variation in temperature dependent development if genetically divergent individuals are introduced to different areas or rapid evolutionary changes occur in new environments [100, 101]. If the geographic distribution of variation in a relevant thermal trait is well understood, then model accuracy may be improved by building separate models for each genotype. For example, an egg hatch phenology model for a subspecies of the Asian gypsy moth, Lymantria dispar asiatica (Vnukovskij), had reduced error compared to a similar model constructed for the European subspecies that has invaded North America, Lymantria dispar dispar (Linnaeus), which has a markedly different predominant phenotype [102]. An alternative approach may be to run several models, each with a different value for the parameter of interest, and present a range of model predictions. Conversely, DDRP could be modified to accept a grid of parameter values so that geographic variation would be accounted for in a single model run.

A lack of knowledge on how early-season environmental conditions or events that initiate the first spring activity of a species (biofix) vary geographically may be a source of error because the model start date affects all downstream predictions. For example, how does first spring activity vary across the wide range of warming conditions possibly encountered for a large region such as CONUS? As a case in point, our phenology model for N. elegantalis assumes that moths have only 55 DDC before egg laying behaviors may occur. This assumption may not be valid for sub-tropical zones of the United States, where flight and reproduction could occur even earlier. Conversely, a much longer spring warm-up may be needed in temperate zones because commercial tomatoes are transplanted much later in the year. Studying how first spring activity (adult flight) in N. elegantalis potentially varies geographically in Central or South America would help to refine a range of model start times. The phenology model for DDRP could then be parameterized using a necessarily conservative selection of start dates or by inputting a grid of start dates. Using a broad distribution of emergence times to initiate the cohorts could be another approach to accommodate uncertainty in first spring activity.

Distributed delay.

There is currently no distributed delay function in DDRP, meaning that the overlap in generations and life stages of cohorts does not increase over multiple generations. This is particularly an issue for species that have significant overlap in generations because they continue to develop throughout winter months or lack a temperature or photoperiodic event that synchronize populations, such as E. postvittana [31, 57, 61]. DDRP may accurately predict peak events in each generation for these species, but inaccurately predict the first appearance of one or more life stages after the first or second generations because of increasing overlap in generational cohorts. Consequently, phenological event maps should be most reliable for the first few generations. This will be among the high priority issues in development of future versions of the platform.