Item No. 1 of 1

INVESTIGATOR: Adiga, A.

PERFORMING INSTITUTION:
UNIVERSITY OF VIRGINIA
CHARLOTTESVILLE, VIRGINIA 22901

FACT: NETWORK MODELS OF FOOD SYSTEMS AND THEIR APPLICATION TO INVASIVE SPECIES SPREAD

NON-TECHNICAL SUMMARY: Agricultural commodity flow networks are a critical component of modern food systems. They also serve as conduits for pest, pathogen and contaminant dispersal. Understanding these food flows and their role ininvasive species spread is essential for food security, and preserving biodiversity, health and economic stability. This project seeks to develop (i) novel network representations and analytics to understand domestic agricultural commodity flows in the United States (ii) pest spread and impact models that account for natural and human-mediated pathways of spread. We apply our models to the study of Tuta absoluta, a devastating pest of the tomato crop.The project will employ state-of-the-art statistical and machine-learning techniques for data integration and network construction. We will develop methods for structural and dynamical analysis of these networks in a novel context of directed and time-varying networks. Agent-based epidemiological models from the infectious disease literature will be adapted for the pest spread model with implementation of various types of interventions. Partial equilibrium models will be used for economic impact analysis.The project will contribute novel network-based approaches for data integration, data analytics and computational modeling. In the context of invasive species, the developed tools will provide policy makers with guidance and support to identify vulnerabilities in the food system, inform monitoring efforts and assess various intervention strategies. These analyses will be particularly valuable and timely to address the imminent threat of T. absoluta invasion. The project will nurture graduate, undergraduate and K-12 programsthroughinterdisciplinary researchand team science.

OBJECTIVES: Goal 1: Data exploration, curation and preliminary analysisWe will be analyzing and incorporating datasets from multiple domains, which include but are not limited to commodity flows, production, demographics, operations, pest interception, climate, etc.Each dataset, depending on the type will be standardized and stored along with metadata in a database. Also, we will store metadata on reports and research articles containing relevant qualitative data. We will perform analysis of these datasets individually to identify inconsistencies and anomalies if any.Goal 2: Modeling and analysis of commodity flows.We will develop methods to construct high-resolution attributed network representations of commodity-specific domestic flows for the US market. Node attributes correspond to production/consumption at the location, GDP and population, type of operation, to name a few.b. Network constructionHere, the objective is to create spatiotemporal commodity flow networks. We will explore methods to construct networks at multiple scales along three axes: (i) spatial: state-FAF region-district (resolution from low to high); (ii) time: annual-seasonal-monthly and (iii) commodity: vegetables-specific vegetable. The specific vegetables we will consider are the primary host crops of T. absoluta. It is expected that the challenges increase with increase in resolution. Broadly, there are two steps involved in the process: (a) estimation of node attribute values, and (b) estimation of flows.c. Structural and dynamical network analysisGiven the constructed networks, we will study their local properties such as total outflow, total inflow, degree, clustering coefficients and betweenness centralities, to name a few. We will also develop algorithms to discover higher-order structures such as clusters (well-connected subgraphs, relatively less connected with the rest of the graph) and network motifs. In particular, we will develop methods to discover properties that are important for spread dynamics, i.e., properties that are important with respect to propagation models (some examples arenetwork reliability andspectral properties).Goal 3: Multi-pathway models for invasive species spread.A multi-pathway modeling framework that couples the above commodity flow models with ecological and bioeconomic models to assess the spread and impact of non-indigenous invasive species.a. Model design and implementationWe will develop a stochastic network-based propagation model to simulate the multi-pathway spread of T. absoluta. Short-distance dispersal captures the spread through natural means. Long-distance human-mediated dispersal corresponds to spread through trade between localities (large urban areas with high production and/or consumption).b. Analyze spread under various hypothetical scenariosWe will apply the multi-pathway model to analyze the possible spread patterns under various scenarios of introduction, monitoring and control strategies. A key subtask here will be to analyze the PestID database and identify locations which have historically been susceptible to pest introductions. This analysis will provide candidate routes of entry. Through simulations with the starting conditions informed by the study, we will analyze the patterns and timeline of spread. This analysis will help identify high risk locations and hubs that facilitate large scale dispersal.c. Economic impact assessmentThe assessment of the economic impact resulting from invasion requires an integration of information on: 1) the biology, ecology and damage caused; 2) its entry; 3) establishment; 4) spread; 5) valuation of assets at risk; and 6) market consequences. To compute the impact, we will explore two different measures: 1) the direct impact and, 2) the total impact. The direct impact measures the direct revenue loss from the crop as the sum of loss encountered by each county or administrative/geographic unit considered, which in turn depends on proportional loss in the affected area, yield per unit of land in the district before being affected, tomato production area in the district, and the proportion of area affected by the pest which is informed by the spread model.Goal 4: Uncertainty quantification and sensitivity analysisGiven that the models are complex and data for validation is generally inadequate (both for network construction as well as multi-pathway model), uncertainty quantification and sensitivity analysis are crucial to evaluate the robustness of the models and credibility of the results. For the commodity flow networks, we will consider parameter variability (resulting from approximating flows), structural uncertainty (resulting from difference in the edge sets inferred) as well as uncertainty in the functional relationship between flows and node & edge attributes. The key question is, how it affects the structural and dynamical properties of the network. For example, how do centrality measures of vertices vary with such perturbations or how does cluster decomposition change. For the multi-pathway model, machine learning and Gaussian process surrogates will be used to perform sensitivity analysis and parameter importance.

APPROACH: Scientific MethodsWe will be using spatial-interaction models such as gravity models and radiation models estimate the commodity flows. Machine learning techniques such as regression, decision trees and deep learning will be used to learn the functional relationships between the flows and their drivers. For the structural analysis of the networks will use various betweenness centrality algorithms, spectral methods, dense subgraph and network motif discovery algorithms, and community detection algorithms. These will be particularly aimed at directed weighted networks. For dynamical analysis, network epidemiological models will be designed incorporating the commodity flow networks. Simulations will be run on the networks to study hypothetical invasion and intervention scenarios, assess importance of nodes, vulnerable regions, hubs of spread, etc. Surrogate models will be used for parameter space exploration and sensitivity analysis. To assess total economic impact, we will use the partial equilibrium approach, which accounts for the dynamics of the market, efficacy of interventions, etc. Given that the models are complex and data for validation is generally inadequate (both for network construction as well as multi-pathway model), uncertainty quantification and sensitivity analysis are crucial to evaluate the robustness of the models and credibility of the results.EffortsResults of the project will be presented in conferences and workshops. Leveraging other projects that the PI and co-PIs are participating in, we will host workshops, conduct courses and webinars.Evaluation. This is organized by goals and objectives.1(a). Data exploration, curation and preliminary analysisWhen: Years 1 and 2Milestones:(i) Relevant datasets identified and collected for network construction,modeling and impact analysis.(ii) They will be analyzed individually and in combination to assess their utility and role in commodity flow and multipathway model.Measures of success:(i) Datasets made available in a central database or repository. Each dataset, depending on the type standardized and stored along with metadata.(ii) Documentation of datasets, processing steps and how we plan to incorporate them into the commodity flow networks and pathway models;(iii) Code for preprocessing and analyzing data will be verified (using test cases) and documented.1(b) Network constructionWhen: Years 1 and 2Milestones:(i) Seasonal trade networks of solanaceous crops at various spatial scales will be generated.(ii) Primary drivers of trade will be determined.Measures of success:(i) Validation of network structure and volume using sample flow data, production and consumption information;(ii) Networks will be made available in a central database or repository;(iii) Documentation of network construction pipeline.(iv) Code will be verified and documented.1(c) Structural and dynamical network analysisWhen: Years 1 and 2Milestones:(i) Identification of important nodes, edges, network motifs and communities in the networks.(ii) Uncertainty quantification and sensitivity analysis will be performed.Measures of success:(i) Network measures such as betweenness centrality, spectral methods, network reliability and hyperbolicity will be applied. The results will be compared with previous works on trade network analytics.(ii) Methods and results will be documented, published and presented.(iii) Code will be verified and documented.2(a) Model design and implementationWhen: Years 2, 3 and 4Milestones:(i) Verified and validated multipathway model.(ii) Emergent properties of the model will be analyzed through HPC simulations and machine learning surrogates. Collective roles of different pathways will be determined.Measures of success:(i) For validation of the model, incidence reports from some T. absoluta infested regions will be collected. The model will be calibrated and validated for these datasets. Transfer learning techniques will be applied to apply this validated model to the North American setting.(ii) Rigorous sensitivity analysis will be conducted.(iii) Methods and results will be documented.(iv) Code will be verified and documented.(v) Project towards thesis.2(b) Analyze spread under various hypothetical scenariosWhen: Years 3 and 4Milestones:(i) Identification of potential routes of introduction and possible response scenarios due to hypothetical invasion.(ii) Analysis of counterfactual scenarios.Measures of success:(i) Findings will be compared invasion records from Europe and other regions.(ii) Methods and results will be documented, published and presented.(iii) Code will be verified and documented.(iv) Projects towards student theses.3(c) Economic impact assessmentWhen: Years 3 and 4Milestones:(i) Analysis of production, pricing and loss data of commodities, and intervention costs(ii) Design and implementation of the economic surplus model.(iii) Analysis of counterfactual scenarios.Measures of success:(i) Findings will be compared invasion records from Europe and other regions.(ii) Methods and results will be documented, published and presented.(iii) Code will be verified and documented.(iv) Projects towards student theses.

PROGRESS: 2019/09 TO 2020/08
Target Audience: Nothing Reported Changes/Problems:Disruptions due to COVID-19: 1. Personnel effort: Our team has been involved in COVID modeling work since March 2020 to support policy making for the state and the university. In this regard, much of the resources had to be redirected. As a result, other projects such as this one have not received as much attention as planned at the beginning of this period. 2. Student hiring and mentorship: It has been difficult to hire and remotely mentor students. What opportunities for training and professional development has the project provided?Two graduate students and one postdoc were mentored during this period. One graduate student was tasked with downloading and analyzing various data sources (Goals 1 and 2). In this process the student was primarily mentored in database management systems, GIS and network analytics, and developed expertise in python programming language and PostgreSQL. The student was also tasked with presenting research papers related to network construction and related problems in linear and non-linear optimization. The second graduate student has been tasked with developing the invasive species spread simulation model, which is being implemented using python programming language (Goals 2 and 3). The main objectives are two-fold: (i) extend a previously implemented model to make it more general and efficient and (ii) enable integration with intervention algorithms framework when these are developed at a later stage in the project. The student is being mentored in vectorized operations in advanced python packages Numpy and Pandas. The student is also developing advanced optimization algorithms for monitoring and interventions. The student is being mentored on network dynamics and linear programming to this end. The postdoc is exploring the possibility of applying artificial neural networks for the problem of network construction (Goal 1). How have the results been disseminated to communities of interest? Nothing Reported What do you plan to do during the next reporting period to accomplish the goals?Goal 1: Data exploration, curation and preliminary analysis More datasets will be explored. This includes at least the following sources: (i) USDA APHIS Pest interception database to which we recently gained access (ii) data for economic impact analysis from ERS. Goal 2: Modeling and analysis of commodity flows. We will continue to work on modeling commodity flows from FAF networks. The main challenges are (i) estimating commodity specific flows from aggregate vegetable flows, and (ii) estimating time varying networks representing seasonal commodity flows from annual production and trade data. Literature survey indicates that there is no work done in this area to the best of our knowledge. However, there are several works on disaggregating available commodity flows to finer resolution spatial networks. Goal 3: Multi-pathway models for invasive species spread. The ongoing work on the multi-pathway model to make it efficient and more general will be completed (Goal 3a). The major work during this period will be the design of network-based monitoring and controlling algorithms, which has already been initiated (Goal 3b). We will also initiate the economic impact assessment (Goal 3c). Goal 4: Uncertainty quantification and sensitivity analysis All the methods developed under the three goals will be rigorously analyzed with regard to robustness. We will analyze the sensitivity of the modeled networks to functional and structural perturbations of the modeled networks. Under these perturbations, we will also analyze how forecasts and outcomes of intervention algorithms change. At least two publications planned for this period: 1. Analysis of food flows using network reliability. 2. Controlling the multi-pathway spread of invasive species using multi-scale network interventions.

IMPACT: 2019/09 TO 2020/08
What was accomplished under these goals? Goal 1: Data exploration, curation and preliminary analysis In this period, datasets -- vegetable and cereal commodity flows from Freight Analysis Framework (FAF), vegetable production from CROPSCAPE and trade matrix from FAO were downloaded and stored in a PostgreSQL database. Since the data is from multiple sources, we had to standardize the datasets in order to analyze them in combination. Using GIS tools in python programming language, production data was disaggregated to 25kmx25km cells in preparation for use in simulations. Goal 2: Modeling and analysis of commodity flows. Two sets of commodity flow networks were constructed from the above mentioned datasets -- US vegetable and cereal flows from FAF database for various years and country-to-country vegetable flow networks from FAO trade matrix. Structural and dynamical analysis of these datasets are underway. Important nodes in the network such as hubs, sources and sinks were identified based on structural analysis such as indegree, outdegree, betweenness centrality, etc. We are developing a novel approach to identify important dynamics-induced clusters of highly-connected nodes in a directed weighted network using Moore-Shannon network reliability. Goal 3: Multi-pathway models for invasive species spread. The multi-pathway model from our previous work [McNitt et al. 2019] is being extended. Firstly, based on feedback received more functional relationships are being implemented. Secondly, using vectorized operations in python programming language we have been making the simulator faster and therefore scalable to larger networks. This will enable us to run simulations at the region/country scale in the US. We are developing a multi-scale intervention framework with the objective of selecting few locations to setup traps or apply interventions such as pesticides in order to delay or stifle the spread of the pest in the event of its introduction. The approach will apply agent-based models along with linear optimization techniques. McNitt, J., Chungbaek, Y. Y., Mortveit, H., Marathe, M., Campos, M. R., Desneux, N., ... & Adiga, A. (2019). Assessing the multi-pathway threat from an invasive agricultural pest: Tuta absoluta in Asia. Proceedings of the Royal Society B, 286(1913), 20191159.

PUBLICATIONS (not previously reported): 2019/09 TO 2020/08
No publications reported this period.