Mid Sweden University

This thesis focuses on building Deep Learning algorithms/networks to classify/predict the degree of damage to larch trees using a dataset from the Swedish Forest Agency. By predicting the damage with a Deep learning algorithm, faster findings to find damaged trees will likely lead to faster actions to remove damaged trees or cure them, thus making the forest environment healthier by removing and treating damaged trees. This project aims to investigate and implement different approaches using different learning-based models to predict healthy/damaged trees to see if the algorithm performs better or not in terms of accuracy. The algorithms chosen for this project are a simple CNN-based network and a Faster R-CNN-based network with a VGG16 image classification model by using Google Colabs free version using TensorFlow and Keras to create the algorithm. The simple CNN uses a binary/multiclassification only with cropped images and Faster R-CNN uses cropped images and the whole forest images. The performance evaluation is made by first finding the accuracy of training data through comparison by a different number of images, steps per epoch and using a fixed amount of five epochs. After finding the most feasible values to be used in network, then begin predicting the images. The results show that the Faster R-CNN with VGG16 performs better in terms of prediction accuracy when using cropped images, and thus, it is feasible to use to predict the damage of the trees, meanwhile simple CNN has worse prediction accuracy.