Bayesian Spatiotemporal Small Area Estimation for Forest Disturbance Monitoring

Climate change is increasing the frequency and severity of forest disturbances (e.g., droughts, wildfires, storms, and insect outbreaks), which are inherently localized and evolve rapidly across space and time. While national forest inventories (NFIs) provide unbiased estimates of forest resources at large scales, they are not designed to quantify localized events when local sample sizes are small. Remote sensing technologies (LiDAR, satellite imagery) can detect and map disturbed areas with high precision, but they do not directly measure essential forest attributes such as the volume of timber affected.

Model-assisted estimation offers a partial solution: by incorporating remote sensing data as auxiliary covariates, it is possible to reduce the variance of design-based estimators without sacrificing their design-based validity. However, this approach relies on asymptotic guarantees and on auxiliary data that are strongly correlated with field measurements, conditions that are difficult to satisfy at fine spatial and temporal resolutions or when disturbances are rare and localized.

This PhD will develop a methodological framework for high-resolution monitoring of forest disturbances using French NFI data. Funded by the Institut national de l’information géographique et forestière (IGN) as part of its commitment to advancing operational forest monitoring, the project will build on Bayesian small area estimation (SAE) methods that borrow statistical strength across space and time when local data are insufficient. Building on the Fay-Herriot model and its spatiotemporal extensions, the project will exploit approximately 20 years of French NFI data to produce coherent posterior distributions of forest disturbance indicators at fine spatial resolution. Inference will likely be carried out using Integrated Nested Laplace Approximation (INLA), which provides substantial computational advantages over MCMC and makes near-real-time updating operationally feasible.

A key methodological objective is to develop Bayesian spatiotemporal SAE methods that remain design-consistent, or approximately design-consistent, while benchmarking naturally to official inventory estimates. Particular emphasis will be placed on estimating forest attributes for disturbance-defined domains identified through remote sensing, even when these domains evolve over time (e.g., bark beetle outbreaks) and are not known at the survey design stage. More broadly, the project aims to develop computationally efficient statistical workflows for operational forest disturbance monitoring.

The methods will be developed at the Laboratory of Forest Inventory (LIF) using real NFI and satellite data to produce high-resolution maps of disturbance severity in terms of affected timber volume. Reproducible tools, including software packages and an interactive application, will be developed to facilitate operational implementation.

The Institut national de l'information géographique et forestière (IGN) is a French public establishment under the joint authority of the Ministries of Ecology and Forestry. As the national reference operator for geographic and forest information, IGN produces, qualifies, and disseminates the sovereign geospatial and forest data that underpin public policy, including responsibility for the French National Forest Inventory.

The Laboratoire d'inventaire forestier (LIF) is a research unit of IGN. Based in Nancy in the Lorraine region of France on a historic forestry and engineering campus now part of AgroParisTech, LIF combines forest inventory field observations, remote sensing, and statistical modelling to develop innovative methods supporting sustainable forest management, environmental monitoring, and responses to ecological and climate transitions.

A candidate with a strong Master’s in quantitative field (e.g. statistics, data science, applied mathematics, survey sampling, epidemiology, …). You should be comfortable with statistical modeling and programming in R or Python. Experience with survey sampling, spatial statistics, or Bayesian methods is clearly a plus — but motivation and the capacity to learn matter more than a checklist.

Required:

• Master’s degree (or equivalent 5-year diploma) by the start date
• Solid knowledge of statistical modeling
• Programming skills in R or Python
• Fluency in English (written and spoken)

Preferred:

• Experience with survey sampling or spatial statistics
• Experience with hierarchical Bayesian methods
• Interest in environmental or geospatial applications
• Interest in disease mapping (epidemiological analogy)
• French is a plus but not required