Model Generalization Within a Compound Maximum a Posteriori Approach for the Classification of Historical Land Use and Land Cover Trajectories

The classification of remote sensing time series with 30+ years of data into land cover trajectories is influenced by the low availability of adequate observations in older dates and the collection of suitable reference data. In this study, we assess the Compound Maximuma posteriori(CMAP) algorithm within a model generalization design (i.e., using models trained with labeled pixels for only a subset of the time series to be classified) as an efficient method for classifying long remote sensing time series directly into valid only land cover trajectories. Here, we tested the simplest version of CMAP to classify 37 years of annual Landsat observations for forest regeneration analyses in a region within the Brazilian Amazon, in Pará state. These results were compared to those obtained with traditional Post-Classification Comparison approaches based on the same input data, with and without the use of post-classification filters. Our results show that independently classifying each image led to almost 90% of the pixels labeled with at least one invalid transition, as well as many areas with signs of deforestation labeled as old-growth forest. Applying the post-classification filter solved the problem of invalid transitions but led to a surplus of pixels labeled as secondary vegetation. CMAP was able to directly classify time series into valid only land cover trajectories without the need to design and implement additional post-processing filters. CMAP presented well-balanced results, with the highest global quality metrics values for the tested years and a distinct improvement over User's Accuracy for the secondary vegetation class. Our solution to deal with the lack of observations and the length of the time series can be extended to alternative forms of CMAP, which in turn could be a valuable tool for multi-temporal land cover classification projects.