The composition of time-series images and using the technique SMOTE ENN for balancing datasets in land use/cover mapping

Abstract

Monitoring Land-use/land-cover (LULC) changes are a significant challenge for sustainable spatial planning, particularly in response to transformation and degenerative landscape processes. These disturbances lead to the vulnerability of inhabitants and habitat and climate changes and socio-economic development in the region. Several studies have proposed different methods and techniques to monitor the spatial and temporal changes of LULC. Machine learning is a more popular method. However, the problem of data imbalance is a significant challenge, and the classification results tend to bias the majority classes for unbalanced data. Therefore, this study's objective is to develop a state-of-the-art technique to reduce the problem of data imbalance in LULC classification in Vietnam based on machine learning and SMOTE (Synthesizing Minority Oversampling Technology) with Edited Nearest Neighbor (ENN). Various statistical indices, including Kappa and Accuracy, have been used to assess the performance for the classification of Land-use/cover. The results indicate that integrating oversampling and under-sampling with SMOTE ENN gave better overall accuracy and generalization. We also find that the expected proportion of chance agreement after oversampling is higher than before (Kappa score before and after oversampling is 0.905244 and 0.974379, respectively). This study provides an effective method to monitor spatial and temporal land cover change in Vietnam; it plays a role as a framework for other relevant research related to land cover change, which can support planning and sustainable management of the territory.