Mapping floristic patterns of trees in Peruvian Amazonia using remote sensing and machine learning

Recognition of the spatial variation in tree species composition is a necessary precondition for wise management and conservation of forests. In the Peruvian Amazonia, this goal is not yet achieved mostly because adequate species inventory data has been lacking. The recently started Peruvian national forest inventory (INFFS) is expected to change the situation. Here, we analyzed genus-level variation, summarized through non-metric multidimensional scaling (NMDS), in a set of 157 INFFS inventory plots in lowland to low mountain rain forests (<2000 m above sea level) using Landsat satellite imagery and climatic, edaphic, and elevation data as predictor variables. Genus-level floristic patterns have earlier been found to be indicative of species-level patterns. In correlation tests, the floristic variation of tree genera was most strongly related to Landsat variables and secondly to climatic variables. We used random forest regression, under varying criteria of feature selection and cross-validation, to predict the floristic composition on the basis of Landsat and environmental data. The best model explained >60% of the variation along NMDS axes 1 and 2 and 40% of the variation along NMDS axis 3. We used this model to predict the three NMDS dimensions at a 450-m resolution over all of the Peruvian Amazonia and classified the pixels into 10 floristic classes using k-means classification. An indicator analysis identified statistically significant indicator genera for 8 out of the 10 classes. The results are congruent with earlier studies, suggesting that the approach is robust and can be applied to other tropical regions, which is useful for reducing research gaps and for identifying suitable areas for conservation.