Tree Species Identification in Urban Environments Using TensorFlow Lite and a Transfer Learning Approach

Abstract

Building and updating tree inventories is a challenging task for city administrators, requiring significant costs and the expertise of tree identification specialists. In Ecuador, only the Trees Inventory of Cuenca (TIC) contains this information, geolocated and integrated with the taxonomy, origin, leaf, and crown structure, phenological problems, and tree images taken with smartphones of each tree. From this dataset, we selected the fourteen classes with the most information and used the images to train a model, using a Transfer Learning approach, that could be deployed on mobile devices. Our results showed that the model based on ResNet V2 101 performed best, achieving an accuracy of 0.83 and kappa of 0.81 using the TensorFlow Lite interpreter, performing better results using the original model, with an accuracy and kappa of 0.912 and 0.905, respectively. The classes with the best performance were Ramo de novia, Sauce, and Cepillo blanco, which had the highest values of Precision, Recall, and F1-Score. The classes Eucalipto, Capuli, and Urapan were the most difficult to classify. Our study provides a model that can be deployed on Android smartphones, being the beginning of future implementations.