A spatial–temporal approach to monitoring forest disease spread using multi-temporal high spatial resolution imagery

Abstract

Sudden Oak Death is a new and virulent disease affecting hardwood forests in coastal California. The spatial–temporal dynamics of oak mortality at the landscape scale are crucial indicators of disease progression. Modeling disease spread requires accurate mapping of the dynamic pattern of oak mortality in time through multi-temporal image analysis. Traditional mapping approaches using per-pixel, single-date image classifications have not generated consistently satisfactory results. Incorporation of spatial–temporal contextual information can improve these results. In this paper, we propose a spatial–temporally explicit algorithm to classify individual images using the spectral and spatial–temporal information derived from multiple co-registered images. This algorithm is initialized by a spectral classification using Support Vector Machines (SVM) for each individual image. Then, a Markov Random Fields (MRF) model accounting for ecological compatibility is used to model the spatial–temporal contextual prior probabilities of images. Finally, an iterative algorithm, Iterative Conditional Mode (ICM), is used to update the classification based on the combination of the initial SVM spectral classifications and MRF spatial–temporal contextual model. The algorithm was applied to two-year (2000, 2001) ADAR (Airborne Data Acquisition and Registration) images, from which three classes (bare, dead, forest) are detected. The results showed that the proposed algorithm achieved significantly better results (Year 2000: Kappa = 0.92; Year 2001: Kappa = 0.91), compared to traditional pixel-based single-date approaches (Year 2000: Kappa = 0.67; Year 2001: Kappa = 0.66). The improvement from the contributions of spatial–temporal contextual information indicated the importance of spatial–temporal modeling in multi-temporal remote sensing in general and forest disease modeling in particular.