Enhancing Land Cover Classification with Object-Based Contextual and Time Series Approaches

In this research, Object-Based Image Classification (OBIC) applied to satellite imagery is explored, particularly in the context of land cover classification. While Deep Learning and Object-Based Convolutional Neural Networks (OB-CNN) have enhanced accuracy, they require substantial computing resources and may have limitations in capturing object relationships. To address this, a novel OBIC framework that harnesses Graph Convolutional Networks (GCN) to construct a graph structure from objects, allowing for the consideration of object relationships is introduced. The study introduces the Contextual Object-Based GCN (OB-GCN). The research also emphasizes the significance of analyzing multitemporal imagery for accurate land object classification, introducing a Time Series Object-Based CNN (OB-CNN) that interprets time series data. By implementing the Contextual OB-GCN and Time Series OB-CNN models, overall accuracies of 83.1% and 78.8% are achieved, respectively. The study compares the performance of Contextual OB-GCN with Time Series OB-CNN to determine their effectiveness in classifying land objects.