Land cover classification using machine-learning techniques applied to fused multi-modal satellite imagery and time series data

Sammanfattning: Land cover classification is one of the most studied topics in the field of remote sensing, involving the use of data from satellite sensors to analyze and categorize different land surface types. There are numerous satellite products available, each offering different spatial, spectral, and temporal resolutions. Consequently, several methodologies have been developed to efficiently determine land cover using remote sensing imagery according to the spectral characteristics of each land cover category. The objective of this thesis is to classify an area located in the Ionian region of Greece, identifying ‘Artificial’, ‘Bare Soil’, ‘Cropland’, ‘Dense Forest’, ‘Grassland’, ‘Low-density Urban’, ‘Low/Sparse Vegetation, and ‘Water’ classes. To do so, the study investigates the performance of different techniques for processing and integrating remote sensing data obtained from various sensors. Multi-spectral and thermal imagery are employed, as well as topographic data from the area of interest. Landsat 8 and Landsat 9 images were specifically chosen for this project, as they include both multi-spectral and thermal information in a single acquisition. Additionally, ASTER GDEM data was used for elevation information and the generation of two elevation derivatives, the aspect and the slope of the study area. These factors, along with their temporal variability, are considered crucial as the spectral properties of certain key classes (specifically those related to vegetation and agricultural activities) are influenced by the phenological cycle. The study addresses several research questions, including the impact of thermal information, elevation, and topography on the classification accuracy, as well as the utilization of time series data to enhance the results compared to using only the multispectral information as input. The findings indicate that combining multi-spectral data with either terrain information, thermal infrared bands, or both, significantly improves the classification results using both k-Nearest Neighbor and Random Forests classifiers. The highest performance in classification accuracy is achieved when incorporating the time series information of all the aforementioned factors as input to the Random Forests classifier. This integration yields improvements of up to 68% in specific classes, primarily those associated with vegetation.