Title: A Bayesian spatio-temporal model for map reconstruction of remote sensing observations and forest inventory prediction
Authors: Md Kamrul Hasan Khan - University of Arkansas (United States) [presenting]
Barry T Wilson - US Forest Service (United States)
Avishek Chakraborty - University of Arkansas (United States)
Giovanni Petris - University of Arkansas (United States)
Abstract: The USDA Forest Service aims to use satellite imagery for monitoring and predicting change in forest-conditions over the time across large geographic regions within the country. The auxiliary data collected from these imageries, such as brightness, greenness and wetness, are relatively dense in space and time and can be used to efficiently predict how the forest inventory map changes over time. However, it contains a huge proportion of missing values at every location due to practical limitations. Therefore, we develop a spatio-temporal model to reconstruct these missing values from posterior predictive distributions. The model consists of a temporal fixed effect based on periodic patterns and a spatio-temporal random effect based on conditional autoregressive (CAR) prior. To allow for presence of change points in the landscape (that should prevent spatial smoothing), we modify the neighborhood structure using binary boundary parameters with a Markov prior over time. Once we obtain full spatio-temporal map, we use it to model the presence/absence of forest and the amount of basal area across the region. These models are formulated using functional regression and elastic net regularization is performed to identify important auxiliary variables.