Factorial experimentation with Convolutional Neural Networks in the context of Land Use Land Cover classification
Author
Knudsen, Emil Falk
Term
4. semester
Publication year
2025
Submitted on
2025-05-28
Pages
72
Abstract
Dette speciale bygger videre på mit projekt fra 9. semester om klassifikation af arealanvendelse og arealdække i Aalborg Kommune med det mål at kunne lave et landsdækkende kort. I det tidligere arbejde blev der opnået en samlet nøjagtighed (Overall Accuracy, Po) på 74% og en Cohen’s kappa på 0,63, hvilket var lavere end ønsket. En sandsynlig forklaring var, at arbejdet foregik i Google Earth Engine (GEE), som både begrænsede mængden af træningsdata og adgangen til mere avancerede modeller som neurale netværk. Specialet undersøger derfor, om neurale netværk – specifikt konvolutionsneuronale netværk (CNN), som er velegnede til billedopgaver – kan give bedre resultater end en Random Forest (RF) model i GEE. Først gennemføres et litteraturstudie for at sammenligne RF og andre beslutningstræ-baserede metoder med neurale netværk, for at identificere hvilke netværkstyper der passer bedst til computer vision-opgaver med rumlig kontekst, og for at pege på faktorer, der påvirker både nøjagtighed og beregningshastighed. På baggrund af litteraturen designes og implementeres CNN-modeller. Der gennemføres et delvist faktorforsøg inspireret af Design for Six Sigmas Design of Experiments. Forsøget omfattede syv faktorer: to rettet mod at dæmpe effekten af en relativt stor baggrundsklasse, og de øvrige vedrørte selve CNN-arkitekturen, hvilken data augmentation der anvendes, klassevægte i tabsfunktionen, læringsraten samt antallet af træningsepoker. Specialet beskriver også valg af data og den overordnede struktur i de GEE- og Python-scripts, der blev udviklet til forsøget. Efter optimering af scriptet for hver faktor sammenlignes resultaterne fra CNN-modellen med RF-modellen fra det tidligere projekt. CNN’en klarede sig bedre med en relativ forbedring på 15% for Po og 24% for kappa. Her måler Po andelen af korrekte klassifikationer, mens Cohen’s kappa vurderer overensstemmelse korrigeret for tilfældigheder.
This thesis builds on my 9th-semester report on land use and land cover classification in Aalborg Municipality, with the broader aim of creating a nationwide map. The earlier work achieved an overall accuracy (Po) of 74% and a Cohen’s kappa of 0.63, which was lower than desired. The main suspected causes were working within Google Earth Engine (GEE), which limited both the amount of training data and access to more advanced models such as neural networks. The thesis investigates whether neural networks—specifically convolutional neural networks (CNNs), which are well suited to image tasks—can outperform a Random Forest (RF) model implemented in GEE. A literature review first compares RF and other decision tree ensemble methods with neural networks, identifies which neural network types best handle computer vision tasks with spatial context, and highlights factors that influence both accuracy and computational performance. Guided by the literature, CNN models are designed and implemented. A partial factorial experiment, inspired by the Design for Six Sigma Design of Experiments approach, tests seven factors: two aimed at reducing the influence of a relatively large background class, and the remainder covering CNN architecture, the type of data augmentation, class weights in the loss function, learning rate, and the number of training epochs. The thesis also describes data choices and the overall structure of the GEE and Python scripts developed for the experiment. After optimizing the script for each factor, the CNN’s accuracy metrics are compared to those of the previous RF model. The CNN outperformed the RF with a relative improvement of 15% for Po and 24% for kappa. Here, Po represents the share of correct classifications, while Cohen’s kappa measures agreement adjusted for chance.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Documents