Fig 1.
The citrus landscapes for three regions in Texas and California and the survey data for emerging HLB and ACP epidemics in the regions.
(A.C,E) Gridded approximation of the citrus distribution in the (A) Lower Rio Grande Valley, Texas, (C) southern California, and (E) the Central Valley. The regions comprise large commercial citrus groves (high host density) interspersed with dooryard trees in residential areas (low host density). (B,D,F) Geo-coded diagnostic plant samples collected as part of the (B) Texas HLB state-wide survey between December 2011 and October 2018, (D) California state-wide survey between June 2015 and June 2019. We classed samples with Ct value less than 36 (out of 40 qPCR cycles) as positives. (F) Geo-coded ACP samples recorded by California Department of Food and Agriculture independently from the state-wide HLB survey. The samples were found in sticky yellow traps set up near trees in the Central Valley from 2012 to 2017 inclusively. Basemap shapefile for cartographic boundaries reproduced from U.S. Census Bureau under open data use.
Fig 2.
Epidemiological models for ACP and HLB spread.
(A) The stochastic compartment model for HLB epidemiological dynamics in a Texas citrus grid cell and the observation model that matches infection status to survey diagnostic data. An infectious cell can infect other susceptible cells via the movement of local vectors, which were known to have established over the whole region before the emergence of HLB. ϕ, γP, πP denote the transition rates and detection probability and are described in Table 1 and derived in the Methods section. (B) The joint epidemiological model for ACP and HLB dynamics in California where ACP is invading and not yet endemic. The model extends the Texas model and introduces a new epidemic category ‘ACP + HLB infected’ that connects the dynamics of ACP infestation to HLB infection in a grid cell. The transition rates ϕ, ψ, φ, γV, γP and detection probabilities are described in Table 1 and derived in the Methods section.
Table 1.
Principal variables and parameters used in the models, together with estimated values for parameters derived from data augmented Markov chain Monte Carlo inference (see text for details).
Fig 3.
Model selection and goodness-of-fit of the full HLB spread in the Lower Rio Grande Valley, Texas.
(A) Performance of the full model and four model variants in predicting the outcome of survey trials assessed by receiver operating characteristic (ROC) curves. The area under the curve (AUC) of the ROC curve measures the predictive capability of calibrated models when used as binary classifiers to separate positive from negative (1km x 1km) sites. Models were fitted to a randomly sampled portion of survey data up to August 2016, and tested using the remaining portion and data up to August 2017. (B) Performance of the full model in predicting the outcome of survey trials within and beyond the temporal scope of the Texas dataset used for parameter estimation in Table 1. The model was fitted to a randomly sampled portion (80%) of the training data (collected between December 2011 to August 2016) and tested using the remaining part (20%) of the training data and the testing data (collected between September 2016 and October 2018) for the ROC analysis. (C) Temporal progression of the prevalence of three infection categories (Exposed, Infectious, and Detected) for the full model compared with the surveillance data. Besides the medians of 1000 simulation realizations (solid lines), we also show 50%, 75%, and 95% credible intervals (shades of decreasing intensities). The vertical dotted line separates the training dataset (used for parameter estimation) from the test dataset. (D) Spatial autocorrelation scores using Moran’s I of the ‘detected’ categories of the model with the survey data at the end of August 2017: medians of 1000 simulation realizations (red line) with 50%, 75%, and 95% credible intervals (shades of decreasing intensities).
Fig 4.
Spatiotemporal retrospective and prospective prediction of HLB spread in the Lower Rio Grande Valley, Texas.
Spatiotemporal retrospective analysis of the historical spread of training period (Dec 2011 –Aug 2016) and prospective prediction of testing period (Sep 2016 –Oct 2018) and the future where no data were available (Nov 2018 –Dec 2020). We calculated infection risk by averaging over 1000 simulation realizations of the model fitted to the training data. Basemap shapefile for cartographic boundaries reproduced from U.S. Census Bureau under open data use.
Fig 5.
The role of primary and secondary transmission rates and forces of infection on HLB Spread in the Lower Rio Grande Valley, Texas HLB spread.
(A) Posterior distribution of transmission rates via four sources: movement of local vectors from infectious trees in the landscape (β), the arrival of infected vectors over the international border (εB) and from further away (εw)and introduction of infected trees by other human-mediated movement (ε). (B) Contribution of the human-mediated, psyllid-driven primary infection and psyllid-driven secondary transmission sources to the realised infection pressure overall on susceptible trees in the survey period 2012–2018.
Fig 6.
Retrospective analysis of the effectiveness of vector control scenarios in slowing down HLB spread.
The effect of the efficiency of the annual coordinated spraying program on (A) the temporal progression and (B-D) spatial snapshots in October 2014 of the cumulative Exposed area. We considered three different hypothetical area-wide coordinated control scenarios at efficiencies of (B) 20%, (C) 50% and (D) 80%. Basemap shapefile for cartographic boundaries reproduced from U.S. Census Bureau under open data use.
Fig 7.
Spatiotemporal prediction of further HLB spread in southern California.
We used California HLB survey data up to June 2017 to infer the locations of the hidden infectious cells and used the estimations to seed forward simulations. We calculated infection probabilities by averaging outputs from 1000 simulation runs for the prospective spread between June 2017 and December 2021. Basemap shapefile for cartographic boundaries reproduced from U.S. Census Bureau under open data use.
Fig 8.
Evaluation of HLB epidemic progress in southern California vs survey data and the impact of quarantine radius to epidemic outcome.
(A) Temporal progression of the prevalence of three infection categories (Exposed, Infectious, and Detected) in comparison with the training and testing data. We show means of 1000 simulation realizations as solid lines, and 50%, 75%, 95% credible intervals as shades of decreasing intensities. (B) Comparison of the spatial autocorrelation scores of the HLB Detected categories (red line and shades) in southern California with that of the HLB survey data (purple line) at the end of the testing data (June 2019).
Fig 9.
The effect of the radius of the quarantine area centring around HLB newly detected sites on the total Infectious area for southern California.
We started simulations from June 2017 and assessed the total infectious area for December 2021.
Fig 10.
Spatiotemporal prediction of the potential ACP and HLB spread in the Central Valley.
We used Central Valley ACP trapping data to seed the simulations for ACP spread and locations of inconclusive HLB samples (Ct value less than 38 in qPCR diagnostic test) as the initial infected HLB sites. HLB spread can only happen between ACP infested sites. We calculated the ACP infestation and HLB infection probabilities by averaging over 1000 simulation runs for the prospective epidemics from January 2020 to December 2030. Basemap shapefile for cartographic boundaries reproduced from U.S. Census Bureau under open data use.
Fig 11.
Potential ACP and HLB epidemic progress in the Central Valley and the likely effectiveness of control.
(A) Temporal progression of the predicted infestation prevalence in comparison with the trapping data for the data availability period (up to October 2017). (B) Temporal progression of the predicted ACP and HLB epidemics from January 2019 to December 2030. We assumed that as ACP prevalence passes 0.3, ACP the reactive treatment program would have been dropped due to the high cost of maintaining the program and the reduced effectiveness as ACP becomes widespread. (C) Comparison of the spatial autocorrelation scores of the predicted ACP infestation prevalence (red line and shading) in the Central Valley with that of the ACP trapping data (purple line) at the end of the data availability period. (D,E) The effect of varying the efficiency and radius of pesticide treatment upon detection of the vector ACP on the total Infested area in the Central Valley. We started simulations from January 2020 and calculated the total Infested area for December 2021 for (D) control efficiency from 0% to 100% for treated circles of radius 0.4 km and (E) treatment radius from 0.1 km to 2 km assuming spraying efficiency of 80%.