Potential of machine learning and WorldView-2 images for recognizing endangered and invasive species in the Atlantic Rainforest

Key message We found high accuracy classification ( F measure = 95%, on cross-validation) of Araucaria angustifolia (Bertol.) Kuntze, an endangered native species, and Hovenia dulcis Thunb. an aggressive, invasive alien species in WorldView-2 multispectral images. In applying machine learning algorithms, the spectral attributes mainly related to the near-infrared band were the most important for the models . Context It is difficult to classify tree species in tropical rainforests due to the high spectral response’s diversity of existing species, as well as to adjust efficient machine learning techniques and orbital image resolution. Aims To explore the spectral and textural response of an endangered species ( A. angustifo lia) and an invasive species ( H. dulcis ) in WorldView-2 multispectral images, testing its recognition capability by machine learning techniques. Methods We used a WordView-2 (2016) image with 0.5-m spatial resolution. Then we manually clipped the canopy area of the two species in this image using two compositions: True color composition (R=660 nm, G=545 nm, B=480 nm) and near-infrared composition (NIR-2=950 nm, G=545 nm, B=480 nm). Thus, we applied spectral and textural descriptors (pyramid histogram of oriented gradients—PHOG and Edge Filter), which selects the most representative features of the dataset. Finally, we used artificial neural networks (ANN) and random forest (RF) for tree species classification. Results The species classification was performed with high accuracy (F measure = 95%, on cross-validation), essentially for spectral attributes using the near-infrared composition. RF surpassed the ANN classification rates and also proved to be more stable and faster for training and testing. Conclusion The WorldView-2 multispectral sensor showed the potential to provide sufficient information for classifying two species, proving its usefulness in this phytophysiognomy where hyperspectral sensors are generally used for this type of classification.