Predicting Helicoverpa zea (Lepidoptera: Noctuidae) populations in the southern United States

Abstract

Forecasting insect pest populations before crops are planted can help improve management and reduce pesticide use. Pests with long dispersal potentials and wide host ranges are difficult to predict but often cause losses in crops across broad spatial scales. Here, we use corn earworm, Helicoverpa zea Boddie (Lepidoptera: Noctuidae), as a case study of a regional crop pest to develop a forecasting framework. Models used historical H. zea trap data to predict populations across 12 states in the southern United States. Three regression models and one machine learning model were used to evaluate predictive performance in either space or time (ie year over year): linear regression model with fixed-effects only, Bayesian spatial-intercept model with and without fixed effects, and random forest. Results of random forest model provided the most useful information about H. zea population dynamics. We found that the location and year of trap location was the most meaningful predictor of H. zea abundance. The importance of local host crop abundance was less meaningful than the prior predictors. We also tested the importance of trap density to recover forecasting signals in North Carolina. Results showed that the trap density could be reduced by approximately 25% while still recovering reasonable predictions of H. zea density. However, it is clear from prior work that some trap locations are more important for prediction, so further assessment of the specific roles individual trap nodes present would be useful. Together this study highlights opportunities to improve annual H. zea forecasting across the US Cotton Belt.

Introduction

Predicting when and where problematic insect pest populations will arise is a major challenge confronting agriculture. Often aspects of pest biology, such as dispersal capability and host plant breadth, drive the prevalence of pests across multiple spatial scales (Kennedy and Storer 2000). Quantifying this population variation is an important step to inform field-based scouting practices and planning for management interventions. To understand variability, area-wide monitoring has been a common approach to track pest activity in space and time, often relying on informal networks of volunteers to count pests and maintain traps. Farmers use this information to estimate infestation risk based on their experiences for the suite of host crops that might be sensitive to a given pest. Many of these coordinated monitoring networks have been maintained for decades and, as a result, have compiled significant quantities of data that capture pest activity trends that can be useful for modeling (Bell et al. 2015, Storkey et al. 2016, Chappell et al. 2020, Lagos-Kutz et al. 2020, 2024, Emery et al. 2024, Ewald et al. 2024).

Observational networks used to develop pest forecasts are costly, both in terms of logistics for collection and resources required to maintain infrastructure over time. Often these costs are distributed through volunteer effort and piecemeal funding sources to maintain data collection infrastructure. As a result, there is a need to measure network efficiency in space and time to help prioritize locations where data collection provides the greatest return on investment (Paredes et al. 2022, Yanchenko et al. 2024). In plant disease management, spatial allocation of resources is a frequent challenge that is driven by the cost of establishing and monitoring sentinel plantings to track disease onset at broad scales (Ojiambo et al. 2015, Garrett et al. 2018, Ojwang et al. 2024). Similar optimization approaches have been used in entomology, often through field experiments (Caton et al. 2025a), simulation (Guimapi et al 2020, Palma et al. 2023, Caton et al. 2025b), or analysis of historical data (Yanchenko et al. 2024, Slone et al. 2025). Optimization of network configuration or temporal coverage can improve systems while providing important savings that prolong network sustainability over time (Ojwang et al. 2024, Yanchenko et al. 2024). To understand tradeoffs between network saturation and minimal location counts, contemporary analytical approaches using historical trap data could provide insight into which sites are most informative and the duration they need to be operated to inform management.

Corn earworm, Helicoverpa zea (Boddie), is a common moth pest that attacks many economically important crops in North America such as cotton, corn, soybeans, and high-value vegetables (Hardwick 1965). Because H. zea caterpillar feeding reduces crop yields, farmers and researchers have established coordinated monitoring programs for decades (Lawton et al. 2022, Huseth et al. 2024). These programs use black lights or sex pheromone-baited traps to measure activity patterns of adult H. zea moths in space and time (Hodgson et al. 2023). Large adult moth populations often correspond with egg laying in sensitive host crops, which leads to larval feeding. As a result, farmers use moth activity data to estimate larval infestation risk across their fields, which can vary considerably at fine scales (Yanchenko et al. 2024).

These data are an untapped resource for H. zea prediction when considered in aggregate. Previous efforts to combine disparate trap datasets collected across the United States generated new insights into the overwintering range of H. zea and the potential for ranges to expand as the climate warms (Lawton et al. 2022). This descriptive approach provides useful information about long-term trends in H. zea populations but does little to help farmers plan for their severity before the growing season. A recent study provided evidence that H. zea could be predicted at the scale of North Carolina, a state where infestations of this pest are often severe (Wallace et al. 2025). The extent of the study omits a large swath of intensive agriculture in the southern United States, commonly called the Cotton Belt, where H. zea is a chronic pest annually. Expanding the prediction framework to other major agricultural areas in the region would benefit multiple stakeholders that intensively manage crops to alleviate risk from this pest on an annual basis.

The present study used H. zea activity data collected by Lawton et al. (2022) to explore population prediction across 12 states in the US Cotton Belt where sensitive host crops are grown. Our goal was to better understand spatial and agricultural factors that can be used to forecast population dynamics before the growing season begins. Specifically, this analysis was developed to predict pest abundance for unobserved years and locations. We hypothesized that historical trap captures at specific locations would be informative for forecasting the current year H. zea counts. We know that some locations have chronic issues with H. zea in some locations, but little is known about the consistency of populations across the region. We also used spatial components of trap location to predict populations. Anecdotal evidence suggests that the severity of this pest varies markedly across the extent of the study, but these relationships have not been extensively explored for the purpose of pest prediction. Together, improving understanding of H. zea population prediction can accelerate forecasting efforts that specifically target farmers before the growing season begins. This knowledge will help in their ability to plan for severe outbreaks while minimizing crop damage risk on an annual basis.

Materials and Methods

Pre-Processing and Exploratory Data Analysis

We used the weekly aggregated H. zea counts for individual traps reported in the Lawton et al. (2022) dataset. In total, there are $n=\mathrm{98,023}$ entries from $L=\mathrm{2,252}$ locations from 1981 to 2021 ($T=41$year period). We filtered the data to include only the southeastern United states and drop observations from 2021 as well as missing values. Observations from 2021 were discounted due to low representation in the Lawton et al. (2022) dataset. The final dataset included $n=\mathrm{22,890}$ data points over a $T=32$ year period (1989 to 2020) from $L=537$ locations (Fig. 1).

Fig. 1.

Data pre-processing schematic for Helicoverpa zea data collected from 1989 to 2021. Note that 2021 was discounted due to an incomplete number of site locations at the time of dataset assembly by Lawton et al. (2022).

To visualize temporal activity trends, we plotted the number of observations per year, as well as the average week of peak capture per year (Fig. 2A and B). The peak week is the week with the largest number of counts for each year-location combination. We dropped locations with fewer than 5 observations for a given year, removing about 10% of the data (Fig. 1), and also (temporarily) did not include locations with zero counts for all weeks since this implies no peak week. The number of observations greatly increased in 2007, and there is noticeable variability after this point. This increase is due to the availability of digitally curated data. The average peak week appears to remain roughly the same across years, with an average peak in the beginning of August across the southern United States (Fig. 2B). Differences in peak week were observed in some years (eg 1995, 2009, 2010, and 2019).

Fig. 2.

Number of Helicoverpa zea trap observations per year (A) and the average week of peak trap capture by year (B).

We also compared the number (years) of observations per site. The mean and median are 3.3 and 2, respectively, with a maximum of 32 years. We plotted the number of observations per location on a histogram and map (Fig. 3A and B). The majority of sites with 10+ years of observations are located in eastern North Carolina and western Tennessee. Conversely, most of the sites in Mississippi, Texas, and Florida have only 1 year’s worth of data (Fig. 3B). Indeed, the histogram shows that the plurality of sites have only 1 year’s worth of observations.

Fig. 3.

Histogram of number of observations per location (A) and number of observations per location (B).

To predict the abundance of pests during the growing season, we defined $Y_{lt}^{\prime }$ as the average number of pests for location $l$ from June to September of year $t$ and had $n=1829$ data points at $L=475$ locations. We log transformed the data such that the response variable for modeling is $Y_{lt}=\mathit{log}(Y_{lt}\prime +1).$ A complete diagram of the data pre-processing is in Fig. 1.

External Predictor Variables

For each observation, $Y_{lt}$, we included $p=6$ covariates, scaled to have a mean of zero and variance of one. First, we recorded the location (latitude and longitude) as well as the year the data was collected. To potentially improve predictions, we also incorporated county-level crop data from the United States National Agricultural Statistics Service (USDA-NASS 2023): the number of acres of corn, soybeans, and cotton planted in the previous year. These covariates are considered because they are the primary crop hosts at the regional scale and their abundance affects interannual populations (Jackson et al. 2008). Preliminary analyses also considered large scale environmental conditions (eg multivariate ENSO index, southern oscillation index). The resolution of these datasets, however, exceeds the study extent, so they were excluded from the predictive modeling. Additionally, we found that their inclusion did not noticeably change predictive performance.

Because interannual population size is driven by the abundance of host crops in the preceding year, we used the crop data from year $t-1$ to predict the number of pests for year $t$. This step also enables forecasting without additional within-season crop information. For counties with missing data for certain years, we imputed the mean crop values from the observed years for that county. If a county is completely missing, we used the mean of all locations. It is important to note that these crops are intensively managed for H. zea using either genetically engineered Bacillus thuringiensis toxins expressed in the plant (ie cotton and corn) or foliar insecticides (soybean), which could affect interannual densities depending on the intensity of local management. Despite management inputs, H. zea commonly infests these 3 crops, which can be a major source of moths in the landscape (Huseth et al. 2024) and are sensible predictors of interannual moth activity (Jackson et al. 2008, Arends et al. 2021, 2022, Dorman et al. 2021a).

Model

The primary goal of our analysis is to predict the pest prevalence at each location at the start of the growing season. A secondary goal is to understand which variables affect these predictions. Thus, we modeled our outcome using a random forest regression model (Breiman 2001). Random Forests are an ensemble machine learning algorithm based on decision trees. To prevent overfitting, each individual decision tree is constructed using a random sample of the data, as well as a random sample of the covariates. Thus, there are 3 key hyperparameters that were tuned: the number of trees (ntree), number of predictor variables used in each tree (mtry), and the minimum number of terminal nodes in a tree (nodesize). To choose the optimal values, we performed a grid search over ntree$\in \left\{\mathrm{500,1000}\right\},$  mtry$\in \left\{\mathrm{1,2},\dots ,5\right\},$ and nodesize$\in \{\mathrm{2,5},10\}$ using a 75/25 train/test split of the data. This splitting was repeated 5 times and the parameters corresponding to the lowest average MSE on the test data were used for all subsequent analyses. We found that ntree = 1000, mtry = 4, and nodesize = 5 yielded the best results.

The Random Forest model also yielded a measure of variable importance. For each tree, the root mean squared error (MSE) was recorded on the hold-out test data. Then the value of a covariate was shuffled and the MSE is again computed. We found the difference between the original and shuffled MSE averaged over all trees. If a covariate is a meaningful predictor of the response, then the MSE of the shuffled data was larger than the original MSE. On the other hand, if the predictor variable did not relate to the outcome, then shuffling had little effect on the predictions and the MSE was be similar. Thus, the percent increase in MSE served as a measure of feature importance for each covariate. This only yields a measure of the magnitude of the covariate’s effect, however, not its direction.

For comparison, we also fit a standard linear regression model using the (non-location) covariates, and a Bayesian spatial-intercept linear regression model with covariates (hereafter Full Bayesian) and without covariates (Spatial intercept only). More details for these model fitting methods can be found in the Supplementary Materials.

Model Metrics

We quantified the model fit using the predictive R² and MSE. If $Y$ is the vector of observations and $\hat{Y}$ is the vector of in-sample predictions, then the predictive R² is a pseudo-measure of R², measuring the squared correlation between the observed and fitted values, ie $R^2=\mathit{Cor}{\left(Y,\hat{Y}\right)}^2$. A larger value of this metric indicates a better fit. Since the predictive coefficient of determination is a pseudo-R², it cannot be interpreted as the proportion of variation explained by the model, but rather as a general measure of model fit. We also computed the MSE between the observed and fitted values as

$${\left\{\frac{1}{\tilde{n}}\sum _{lt}{\left(Y_{lt}-{\hat{Y}}_{lt}\right)}^2\right\}}^{1/2},$$

where $\tilde{n}$ is the number of terms in the sum, and smaller values indicated a better fit.

Results

In-Sample Model Fit

We fitted the RF and other models on the entire dataset with $n=1829$ observations and $p=6$ predictor variables. In Table 1, we report the summary measures of model fit. The Random Forest and Full Bayesian models have the largest predictive R² and lowest MSE, respectively. The fixed effects-only model provides the worst in-sample fit based on both R² and MSE. We also report the measure of variable importance for the Random Forest (Fig. 4). Location (latitude and longitude) and year were the most important variables, while the crop data (corn, cotton, soybean) was less informative. The MSE percent increase for longitude was greater than 100%, meaning that if the longitude values are shuffled randomly, then the MSE more than doubles.

Table 1.

In-sample measures of model fit.

Modeling method	R2	Mean squared error
Random Forest	0.67	0.90
Full Bayesian	0.63	0.73
Spatial intercept-only	0.61	0.76
Fixed effects-only	0.15	1.18

The best values are in bold.

Fig. 4.

Random forest parameter importance. The percent increase in hold-out mean squared error if the corresponding covariate values are shuffled randomly.

Prediction

While the previous subsection reported in-sample measures of model fit, the main goal of this work was predicting pest abundance for unobserved years and locations. Towards this end, we trained our models on the data from years $<t$, and predicted average H. zea activity for year $t$. We computed out-of-sample predictive R² and MSE and repeated this process for a sequence of years $t\in \{\mathrm{2010,2011},\dots ,2020\}$. For all results in this subsection, we used the hyperparameters selected from the model development section in the Methods, even though these values were selected by potentially using some of the current test data. The results are in Fig. 5, and the average across all years is in Table 2. The Full Bayesian, intercept-only model, and Random Forest all performed similarly, but, on average, Random Forest provided slightly lower MSE and larger R². Clearly, the fixed effects-only model performed the worst with largest MSE and lowest R². We also noticed significant year-to-year MSE and R² variability in the results (Fig. 5), potentially because the number of locations and observations vary each year.

Fig. 5.

(A) Mean squared error for sequential model fits over time. (B) Predictive fit for 1-year-ahead.

Table 2.

Average predictive R² and MSE for hold-out years from 2010 to 2020.

Modeling method	R2	Mean squared error
Random Forest	0.50	1.15
Full Bayesian	0.45	1.20
Spatial intercept-only	0.46	1.24
Fixed effects-only	0.09	1.51

The best values are in bold.

We also predicted H. zea activity for missing locations, rather than future years. In particular, we wanted to know if all the sampling locations are needed, or if comparable prediction accuracy can be achieved using fewer sites. Indeed, there are many trapping locations in eastern North Carolina (Fig. 3B), so we held out the observations across all years from $q$ proportion of the North Carolina sites to train the model and predict on these hold-out sites. We repeated this process 100 times for each $q\in \left\{\mathrm{0.05,0.10},\dots ,0.95\right\}.$ Random Forest clearly performed the best with uniformly lowest MSE (save $q=0.95$) and largest R² (Fig. 6A and B). The 2 Bayesian models are comparable while the fixed effects-only model had the poorest performance. Additionally, for the Random Forest and Bayesian models, the MSE was similar for $q\le 0.25$ and $q\le 0.50$, respectively.

Fig. 6.

(A) Mean squared error for held out North Carolina locations. (B) Predictive fit for held out North Carolina locations.

Discussion

The results from the in-sample model fitting yielded some key insights on H. zea abundance. First, for the Random Forest and Full Bayesian models, the R² and MSE values had a reasonable model fit over time (Fig. 5A and B). The MSE values are also difficult to interpret since the response is on the log-scale, but they show that Random Forest and the Full Bayesian models yield the closest fit. Additionally, in both the Random Forest and Full Bayesian models, the location of the data was the most informative predictor of H. zea abundance. Indeed, Fig. 4 shows that longitude was significantly more important than all the other variables for Random Forest. Moreover, the relatively small difference between the prediction metrics of the Full Bayesian model and spatial intercept-only model, in addition to the poor performance of the fixed effects-only model, suggests that the model performance is largely being driven by the spatial locations. In other words, the majority of the variation in H. zea populations can be explained by location alone using these models. While it is difficult to visualize the spatial effect in the Random Forest model, we do plot it for the Full Bayesian model (Fig. 7). Specifically, we visualized the posterior mean of the spatial intercept for each location and interpolate between locations using methods described in the Supplemental Materials. The figure reveals certain “hotspots” with larger mean pest counts (in red) such as the Carolinas and eastern Texas. These results should be interpreted cautiously, however, as many of these locations have data for only a single year. Despite these issues, understanding where problematic populations occur in the landscape would benefit growers, particularly when insecticide resistance events complicate management. using forecast information, extension specialists could more effectively tailor annual recommendations for stakeholders that are likely to experience higher populations of H. zea that may not respond to specific insecticides or Bt traits.

Fig. 7.

Spatial effect plot from full Bayesian model.

While the Random Forest and Full Bayesian models yielded similar inferential results, Random Forest clearly outperformed the Bayesian models on both temporal and spatial prediction tasks (Figs 5B and 6B). For the North Carolina subset, the comparable MSE values for $q\le 0.25$ also mean that roughly 25% of the sampling locations in trap network could be removed and the Random Forest model would still obtain similar predictive accuracy, a promising sign for improving sampling efficiency. While the removed sites were randomly chosen, selecting the optimal sites to sample from is also an important problem that can be addressed with historical H. zea activity data (eg Yanchenko et al. 2024).

Overall, these results agree with a similar analysis of H. zea trap captures in North Carolina by Wallace et al. (2025). The authors used a random forest model to forecast H. zea abundance with accumulated growing degree days, previous year’s crop composition in the landscape, and the abundance of H. zea observed in the trap the prior year (Wallace et al. 2025). Results of this more complex model showed that trap capture in the previous year and spatial location (latitude only) were the 2 most meaningful predictors of H. zea abundance. Here, our simplified model containing fewer predictor variables found a similar outcome with location and trap year being the most important factors. Although not tested in our study, the abundance of moths in the prior year is likely to be strongly associated with locations tested (Wallace et al. 2025). Because many locations only had a small number of years observed (Fig. 3A), a more robust assessment of prior year moth abundance was not possible. Despite this shortcoming, our results highlight the chronic nature of H. zea in some locations, which can vary markedly at fine spatial scales (Yanchenko et al. 2024). Understanding where these insects are consistently abundant is a step toward recommending more intensive scouting and management of sensitive host crops where economic injury is possible (eg soybeans, cotton, sweet corn). Several previous H. zea studies have assessed drivers of crop injury (Dively et al. 2021, 2023, Dorman et al. 2021b) and H. zea abundance in the southern row crop region (Allen and Luttrell 2011, Dorman et al. 2021b, Santos et al. 2023, Wallace et al. 2025). Future studies could harmonize both crop injury and forecasting models to assess the relationship between predicted moth counts and resultant crop injury.

Forecasting unobserved locations could also benefit Insect Resistance Management (IRM) for genetically engineered corn and cotton that express Bt toxins by distinguishing areas where high H. zea pressure is anticipated (Reisig et al. 2022). For decades, the academic and regulatory communities have largely focused on responsive approaches to unexpected injury reports from H. zea feeding as Bt resistance events unfold (Reisig et al. 2018), often relying on injury evidence from agricultural stakeholders. Establishing forecasting capability could improve timeliness of response to emerging resistance issues. During initial phases of resistance onset, areas where small pest populations occur with limited gene flow intersect with strong selection pressure are thought to accelerate resistance development (Taylor et al. 1983, Caprio and Tabashnik 1992, Lenormand and Raymond 1998). Forecasts might be of limited utility during this phase of resistance events. However, once the frequency of resistant individuals within pest populations increases, the potential for unexpected crop injury due to field-evolved resistance grows (Tabashnik et al. 2009, 2013). During the resistance outbreak phase, the likelihood of unexpected injury can be difficult to predict due to variability in H. zea population density and other associated factors such as host crop abundance. Utilizing predictive models that describe location-wise population densities, like those presented in this study, could provide guidance for agricultural scientists to (1) incentivize scouting practices in forecasted H. zea hotspots, (2) more efficiently search for crop injury, and (3) promote remedial IRM actions where large populations are forecasted.

One surprising outcome of this study was the limited predictive power of host crop abundance in the previous season. Other recent studies have found associations between crop composition in the previous year and H. zea abundance or crop injury in the study region (Dorman et al. 2021a, 2021b, Arends et al. 2022a, 2022b, Wallace et al. 2025). Importantly, these studies utilized remotely sensed crop data at fine spatial extents from focal fields (ie 1 km). This distance threshold was determined by mark-recapture literature that estimated the local flight distance of H. zea (Graham et al. 1978). Based on relationships found in these studies, local crop composition can have moderate to strong effects on the abundance of H. zea detected in traps. In the present study, remotely sensed crop data did not exist for many years in the dataset, which required using coarse county-level data collected by USDA surveys. We do know that that southern H. zea populations can be localized with large differences in abundance across short distances (Yanchenko et al. 2024), which may explain why county-level crop data did not have a strong signal in our analysis. Accounting for these population differences reinforces the importance of more refined modeling approaches to develop more accurate forecasting models for H. zea in the southern United States.

More broadly, the ecological concepts undergirding pest population dynamics and outbreaks have been an interest for agricultural practitioners for more than a century (Chapman 1939), but the recent development of data analytics has enabled researchers to use vast ecological and environmental data to describe variation. Observational datasets derived from trap networks are often the foundation for models associating abiotic and biotic factors to pest activity (Crossley et al. 2021, Hemming et al. 2022, Lagos-Kutz et al. 2024). Goals of the models vary from predicting pest outbreaks during the growing season (Wiman et al. 2014, Hemming et al. 2022, Rincon et al. 2024, Broadhead et al. 2025) to tracking longer term population fluctuations that may inform general infestation risk between seasons (Zalucki and Furlong 2005, Yamamura et al. 2006, Lawton et al. 2022). Extending models to predict pest populations before they compromise crops is a common objective that could improve the sustainability of crop systems by increasing scouting efficiency and reducing unnecessary insecticide applications. In the future, forecasting could be improved by integrating complex interrelationships with their surrounding ecosystems to increase model accuracy in space and time (Bahlai 2023). Moreover, accounting for changing climatic conditions will also be a key step toward accurately modeling interannual phenology and abundance changes over time (Hodgson et al. 2011, Forrest 2016). These complicating factors reinforce the importance of long-term observational data to compliment emerging modeling approaches. Trap networks will continue to be an important resource to provide near-term assessments of pest activity, as well as data to develop a deeper understanding of pest biology and crop interactions through ecoinformatic approaches (Rosenheim and Gratton 2017). Together, these networks coupled with innovative analytical approaches will improve pest forecasting capabilities to inform stakeholders and address their goals for agricultural sustainability.

Author Contributions

Eric Yanchenko (Conceptualization [equal], Data curation [lead], Formal analysis [lead], Investigation [lead], Methodology [lead], Visualization [lead], Writing—original draft [lead], Writing—review & editing [equal], Brian J. Reich (Conceptualization [equal], Formal analysis [supporting], Funding acquisition [equal], Writing—original draft [supporting], Writing—review & editing [equal]), George Kennedy (Conceptualization [supporting], Writing—original draft [supporting], Writing—review & editing [equal]), and Anders S. Huseth (Conceptualization [equal], Formal analysis [supporting], Funding acquisition [equal], Project administration [lead], Visualization [supporting], Writing—original draft [supporting], Writing—review & editing [equal])

Supplementary Material

Supplementary material is available at Environmental ­Entomology Journal online.

Funding

This research is supported by the Biotechnology Risk Assessment Grant 2020-33522-32272 and Specialty Crop Research Initiative, project award no. 2023-51181-4115 from the U.S. Department of Agriculture’s National Institute of Food and Agriculture. Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the authors and should not be construed to represent any official USDA or U.S. Government determination or policy.

Conflicts of Interest

The authors have no relevant financial or non-financial interests to disclose.

Data Availability

All data used in this study are available via Dryad: Lawton, D, Huseth AS, Kennedy GG, et al. (2022). Pest population dynamics are related to a continental overwintering gradient [Dataset]. Dryad. https://doi.org/10.5061/dryad.m0cfxpp5x