Abstract
This study presents a land use classification framework that integrates spatiotemporal image fusion with deep learning, focusing on farmland and industrial buildings in Taiwan. To address limitations from cloud cover and acquisition frequency, the ESTARFM model was used to fuse Landsat-8/-9 and Himawari-8 imagery. The classification workflow aligned satellite data with national land use surveys and applied PSANet for pixel-level semantic segmentation. Nine land use categories were analyzed, with farmland and industrial buildings emphasized due to their relevance to air quality studies. Accuracy assessments based on 300 randomly selected samples per class indicated that both precision and recall exceeded 0.90 for water bodies and roads. While bare land showed confusion with nearby land types, the combined ESTARFM-PSANet method improved classification performance compared to ISODATA and survey-only methods. The results demonstrate the effectiveness of combining fusion and deep learning for precise, timely land use mapping in support of spatial planning and environmental monitoring.