Factorial experimentation with Convolutional Neural Networks in the context of Land Use Land Cover classification

Knudsen, Emil Falk

4. semester

Surveying, Planning and Land Management (cand.geom.), Master

2025

2025-05-24

72 pages

This master's thesis serves as a follow-up to my 9th semester report, Land Use & Land Cover Classification of Aalborg Municipality. The 9th semester report explored the possibilities of generating a Land Use/Land Cover map for Denmark through Machine Learning (ML). The accuracies achieved in said report were lower than desired with an Overall Accuray (Po of 74% and a Cohen’s kappa of 0.63. This was hypothesised to primarily be due to two reasons; The model was developed in Google Earth Engine (GEE), which put an upper limit on the amount of training data that could be employed and limited the access to ML models to non-Neural Network models. This thesis explores whether better results could be achieved using Neural Networks (NN) as opposed to GEEs Random Forest (RF) architecture. This is first explored through a literature study to determine the difference between RF and other decision tree ensemble methods compared to NN, which type of NN is most suited for Computer Vision tasks that require spatial awareness, and which factors are relevant to the accuracy and computational performance. Based on the literature, which was consulted, this master’s thesis is focused on the construction and implementation of Convolutional Neural Networks (CNN). As a result of the literature study, a partial factorial experiment was set up and executed with inspiration from Design for Six Sigma’s Design of Experiment methodology. The experiment included seven experimental factors, of which the first two were attempts to mitigate the impact of the relatively large background class, while the remaining factors regarded the CNN model’s architecture, the type of data augmentation applied, the class weights applied in the loss functions, the Learning Rate and the number of epochs that the model is trained on. The implementation of these factors, the justification for the chosen data, and the general structure of the GEE and Python scripts written for the NN factorial experiment are described as well. Upon having optimised the script for each factor, the accuracy measures of the CNN model are compared to the RF model from the previous report. In the comparison, it is found that the CNN model outperformed the RF model relatively by 15% for Po and 24% for kappa.

Machine Learning ; Neural networks ; CNN ; LULC ; Factorial experiment ; LandSat 8

Download
View record in AAU Student Projects