TY - THES ID - 136576841 TI - Potential and limitations of random forests and temporal convolutional neural networks for forest leaf type classification in Flanders using Sentinel imagery AU - Elen, An AU - Blaschko, Matthew AU - Somers, Ben AU - KU Leuven. Faculteit Ingenieurswetenschappen. Opleiding Master of Artificial Intelligence (Leuven) PY - 2022 PB - Leuven KU Leuven. Faculteit Ingenieurswetenschappen DB - UniCat UR - https://www.unicat.be/uniCat?func=search&query=sysid:136576841 AB - To define, monitor and evaluate forest management goals, it is important to know forest composition and its evolution through time. Remote sensing has brought new possibilities of automated forest classification based on majority stand type. In the here presented study, existing artificial intelligence methods are evaluated for the classification of forest in Flanders. The main challenge is the limited size of the data set and its class imbalance. A data preprocessing pipeline, based on the normalized difference vegetation index (or NDVI), is proposed to exclude erroneous plots and (cloud) corrupted data points. Erroneous plots are identified based on their yearly average NDVI. Corrupted data points are identified based on their low absolute NDVI value or their deviation from the expected value. Plots that are too sparse in the time dimension are removed. After preprocessing, random forests and temporal convolutional neural networks (TempCNN) are investigated. Several setups are evaluated. The impact of using training data of the same year as the testing data, extending this with training data from other years or - the most realistic approach - using only training data of other years are compared. The impact of limiting the data set to a subset of which the ground truth is more reliable is quantified, as well as the impact of class weighing to compensate for class imbalance. Using regional forest inventory data as ground truth for the classification of four tree types (pure broadleaf, mixed broadleaf, mixed conifer and pure conifer), overall accuracies of up to 88% and 89% are obtained with both the random forest and TempCNN methods respectively. However, it is found that, due to strong class imbalance and the nature of the mixed minority classes, neither of both methods succeeded in accurate prediction of these minority classes. For each class, the obtained accuracy correlated with the size of the training set of that class relative to the majority class. Class weighing can partially compensate for the differences in classification accuracy and when only making a split between two general broadleaf and conifer classes, accuracies up to 97% can be obtained. Excluding the year of the test data from the training data is found to result in comparable classification accuracy as when the year of the test set is included in the training set, which is of great practical importance. The presented experiments show the TempCNN model to perform slightly better than random forest, especially when class weighing is added to partially compensate for the class imbalance. However, the advantage was mainly situated in the majority classes and comes at the cost of extensive hyperparameter tuning and extra calculation time. ER -