Spotted lanternfly predicted to establish in California by 2033 without preventative management

Species distribution (niche) models have been widely used to estimate future distributions of invasive species based on climatic suitability in introduced ranges, but these approaches are not temporally dynamic and generally do not incorporate species biology, limiting their ability to predict timing or routes of spread. Spatio-temporal, process-based models can simulate dispersal and establishment dynamics, offering forecasts of when and where invasions may occur to support proactive management. The spotted lanternfly (SLF, Lycorma delicatula), first detected in Pennsylvania in 2014, has spread across the Mid-Atlantic and threatens high-value crops including grapes, fruit trees, walnuts, and hops. While niche modeling has identified the grape-growing regions of California and Washington as climatically suitable, the timing of SLF arrival to the western US has been unknown. Given the substantial economic value of US grape production ($6.5 billion; >1 million acres, with California producing 82% of the US crop), decision-makers require forecasts of SLF spread without control to benchmark potential management strategies. This study uses a process-based model (PoPS) to simulate the spatio-temporal spread of SLF from its current range across the contiguous US, forecasting establishment timing in vulnerable regions and comparing results with a niche model, while quantifying annual economic risk to major crops.

Prior research has predominantly used niche modeling (e.g., MaxEnt, CLIMEX) to estimate potential ranges of SLF and other invasive species, identifying climatically suitable areas in the US and globally. Studies have indicated high suitability for SLF in the western US grape-growing regions, including California and Washington. However, niche models lack temporal dynamics and typically exclude explicit dispersal and demographic processes, limiting their utility for predicting arrival times and informing management planning. Process-based, spatially explicit models have been applied to other invasive pests and pathogens (e.g., sudden oak death, hemlock woolly adelgid), demonstrating the value of integrating dispersal, host availability, and environmental constraints to forecast spread over time. Evidence also highlights human-mediated transport, particularly along rail networks, as a key driver of long-distance SLF dispersal, underscoring the need to incorporate transportation pathways in spread modeling.

Model framework: The PoPS (Pest or Pathogen Spread) Forecasting System v2.0.0 was used to simulate SLF spread at yearly time steps on a 5-km grid across the contiguous US, starting from the infestation status as of January 1, 2020. The model simulates groups of SLF at the grid-cell level. Reproduction follows a Poisson process with mean β, modified by local environmental conditions (seasonality and temperature indices). Dispersal comprises two components: (i) a natural-distance Cauchy kernel with scale parameter α1 and (ii) a custom network dispersal kernel along railroads to represent accidental human-mediated transport. For each dispersing group in a rail-containing cell, a Bernoulli draw with probability γ determined whether dispersal occurred via the natural kernel or along the rail network. The network kernel allows disembarkation at any point along rail edges and includes parameters for minimum (dmin) and maximum (dmax) travel distances. Calibration and validation: Model parameters (β, natural dispersal distance at, percent natural dispersal γ, dmin, dmax) were calibrated via Approximate Bayesian Computation (ABC) with sequential Monte Carlo and a multivariate normal perturbation kernel, using 2015–2019 SLF survey data (>300,000 observations including positive and negative) from USDA APHIS and state partners (PA, NJ, DE, MD, VA, WV). Parameter sets were retained if accuracy, precision, recall, and specificity each exceeded 65%, iterated over seven generations to 10,000 retained sets per generation with progressively stricter thresholds. Out-of-sample validation using 2020–2021 observations (>100,000) yielded accuracy 84.4%, precision 79.7%, recall 91.55%, and specificity 77.6%. A comparative PoPS run using the prior long-distance kernel (α2) instead of the rail network kernel performed worse (accuracy 76.5%, precision 68.1%, recall 92.68%, specificity 57.2%). Parameter posterior distributions are presented for β, at, γ, dmin, and dmax. Environmental data: Daily Daymet temperature (1980–2019) was converted to monthly indices (0–1) reflecting suitability for survival and reproduction, aggregated to 5 km. Future climates were sampled by random draws from historical years to represent uncertainty in weather. Host distribution: Tree of heaven (Ailanthus altissima), presumed primary host crucial for SLF spread, was modeled via MaxEnt using 19,282 presence records from BIEN and EDDMaps. Predictor variables included mean annual temperature, precipitation of the coldest quarter, precipitation of the driest quarter (WorldClim), and distances to primary roads and railroads (US Census TIGER) to capture disturbed habitats and urbanization. The 1-km MaxEnt output was thresholded (≤0.2 set to 0), rescaled 0–1, validated with an independent 10% subset (95% of validation sites had probability >0.65), and aggregated to 5 km for PoPS. Post-simulation, probabilities were downscaled to 1 km and set to zero where tree of heaven was absent. Simulations and outputs: The model ran 10,000 stochastic iterations for 2020–2050 under a no-management scenario for both SLF and tree of heaven. For each year, cell-level probabilities of occurrence were computed as the fraction of simulations with infestation. County-level probabilities were derived using (1) the maximum pixel probability per county and (2) the mean of pixel probabilities per county. Crop risk analyses used USDA NASS 2017 county-level production data for grapes, almonds, apples, walnuts, cherries, hops, peaches, plums, and apricots to estimate annual risk exposure. Model output for 2050 was compared to the SLF suitability map of Wakie et al. (MaxEnt), using the same categorical thresholds for unsuitable, low, medium, and high risk.

• Timing to California: SLF has a low probability of first reaching grape-producing counties in California by 2027 and a high probability in some California counties by 2033, with likely spread throughout the grape-producing region by 2034. • Crops at risk: Beyond grapes (Vitis), at-risk commodities include almonds (Prunus subgenus Amygdalus), apples (Malus), walnuts (Juglans), cherries (Prunus subgenus Cerasus), hops (Humulus), peaches (Prunus subgenus Amygdalus), plums (Prunus subgenus Prunus), and apricots (Prunus subgenus Prunus). US grape production is valued at $6.5 billion, with more than one million acres in production and California producing 82% of the US crop. • National spread: If unmitigated, SLF is expected to establish across much of the contiguous US by 2037; rail networks represent a high-risk pathway for long-distance dispersal. • Model performance: Against 2020–2021 observations, the calibrated PoPS model achieved accuracy 84.4%, precision 79.7%, recall 91.55%, and specificity 77.6%. Using the previous long-distance kernel (without explicit rail network) reduced performance (accuracy 76.5%, precision 68.1%, recall 92.68%, specificity 57.2%). • Comparison with niche model: In 2050, PoPS and the MaxEnt model (Wakie et al.) agreed that SLF would be unlikely in 47.3% of pixels and that SLF would have some probability of occurring in 32.4% of pixels. In 15.6% of pixels, MaxEnt predicted presence while PoPS did not (of these, 72.6% were low risk, 22.0% medium, 5.4% high per MaxEnt). In 4.7% of pixels, PoPS predicted presence but MaxEnt did not (of these, 41.6% low, 44.3% medium, 14.1% high per PoPS).

Spatio-temporal process-based modeling (PoPS) complements and extends niche modeling by providing not only spatial suitability but also projected timing of establishment, which is essential for management planning. The forecasts indicate that, without intervention, SLF is likely to reach California’s grape regions with nontrivial probability by 2027 and high probability by 2033, threatening high-value perennial crops. Human-mediated transport, especially along rail networks, is a dominant pathway for long-distance spread, justifying targeted surveillance and outreach around rail hubs. Agreement between PoPS and MaxEnt on large portions of the landscape supports the spatial plausibility of modeled spread, while discrepancies highlight the added constraints imposed by dispersal processes and host availability in PoPS. The temporal forecasts provide actionable lead times for surveillance, host management (e.g., tree of heaven removal), and strategic planning by agencies and growers. Scenario modeling in collaboration with state and federal partners can evaluate the potential effectiveness and timing of different management strategies, including eradication, containment, and slow-the-spread regimes.

This study presents a national, process-based, spatio-temporal forecast of SLF spread using PoPS that predicts a low probability of arrival to California grape counties by 2027 and a high probability by 2033 under a no-management scenario. The model integrates host distribution, environmental suitability, and both natural and rail-mediated dispersal, offering a baseline for comparing management strategies and prioritizing surveillance in high-risk areas. The results underscore the value of incorporating transportation pathways into spread models and provide early warning for stakeholders in the western US. Future work should tailor model scenarios to jurisdiction-specific budgets and strategies, incorporate dynamic management actions (e.g., host removal, chemical or biological control), update forecasts with new surveillance data, and evaluate outcomes under alternative climate trajectories.

Forecasts assume no management actions to contain or eradicate SLF or tree of heaven; actual interventions could delay or reduce spread. The model represents populations as groups on a 5-km grid, which may miss fine-scale dynamics. Long-distance dispersal is modeled primarily via the rail network; other pathways (e.g., road freight, air cargo) are not explicitly represented. Future weather was sampled from historical variability rather than climate projections, potentially underrepresenting climate change effects. Tree of heaven was used as the primary host; while supported by evidence, SLF utilizes many hosts and host importance may vary regionally. Impacts on some hosts remain uncertain. Access to comprehensive occurrence data is constrained by confidentiality, though calibration and validation used extensive multi-agency datasets.