Controlling for misclassified land use data: A post-classification latent multinomial logit approach

Terrain and landscape complexities can limit the accurate discrimination of land use categories with similar spectral signatures, as well as the accurate detection of land use change in temporal analyses of landscape dynamics. Studies based on misclassified land use data can generate biased parameter estimates and standard errors, inaccurate predictions, and incorrect policy recommendations. To address these challenges and improve the accuracy of land use analyses, we implement a post-classification strategy to detect misclassified land use observations using a latent multinomial logit model. This strategy is tested using both Monte Carlo simulations and a time series dataset based on supervised classification of remotely sensed data corresponding to land use decisions observed in a Mexican coffee growing region during the period 1984-2006. The results indicate that the strategy is useful for identifying land use observations with a high probability of being wrongly classified, even between categories with low discriminative spectral signatures. Reclassification of the land use data, based on the model results, increases the magnitudes of the marginal effects of the analyzed land use drivers in the theoretically expected directions, and in some cases improves the statistical significance of the parameter estimates.