Tree species classification from AVIRIS-NG hyperspectral imagery using convolutional neural networks

This study focuses on the automatic classification of tree species using a three-dimensional convolutional neural network (CNN) based on field-sampled ground reference data, a LiDAR point cloud and AVIRIS-NG airborne hyperspectral remote sensing imagery with 2 m spatial resolution acquired on 14 June 2021. I created a tree species map for my 10.4 km2 study area which is located in the Jurapark Aargau, a Swiss regional park of national interest. I collected ground reference data for six major tree species present in the study area (Quercus robur, Fagus sylvatica, Fraxinus excelsior, Pinus sylvestris, Tilia platyphyllos, total n = 331). To match the sampled ground reference to the AVIRIS-NG 425 band hyperspectral imagery, I delineated individual tree crowns (ITCs) from a canopy height model (CHM) based on LiDAR point cloud data. After matching the ground reference data to the hyperspectral imagery, I split the extracted image patches to training, validation, and testing subsets. The amount of training, validation and testing data was increased by applying image augmentation through rotating, flipping, and changing the brightness of the original input data. The classifier is a CNN trained on the first 32 principal components (PC’s) extracted from AVIRIS-NG data. The CNN uses image patches of 5 × 5 pixels and consists of two convolutional layers and two fully connected layers. The latter of which is responsible for the final classification using the softmax activation function. The results show that the CNN classifier outperforms comparable conventional classification methods. The CNN model is able to predict the correct tree species with an overall accuracy of 70% and an average F1-score of 0.67. A random forest classifier reached an overall accuracy of 67% and an average F1-score of 0.61 while a support-vector machine classified the tree species with an overall accuracy of 66% and an average F1-score of 0.62. This work highlights that CNNs based on imaging spectroscopy data can produce highly accurate high resolution tree species distribution maps based on a relatively small set of training data thanks to the high dimensionality of hyperspectral images and the ability of CNNs to utilize spatial and spectral features of the data. These maps provide valuable input for modelling the distributions of other plant and animal species and ecosystem services. In addition, this work illustrates the importance of direct collaboration with environmental practitioners to ensure user needs are met. This aspect will be evaluated further in future work by assessing how these products are used by environmental practitioners and as input for modelling purposes