Machine learning-based multimodal data fusion for the Swiss land use statistics

The official national Swiss land use statistics constitutes an important instrument of spatial planning and classify land use (LU) and land cover (LC) through manual expert annotation across a grid of over 4.1 million sample points, resulting in 46 LU and 27 LC classes. We investigate whether machine learning (ML) can streamline this process by integrating multimodal and multitemporal datasets. Aerial RGB and FCIR image-based Convolutional Neural Network (CNN) probability outputs are fused with multispectral Landsat-derived time series indices, digital elevation data, vegetation canopy models and cadastral information using a Random Forest (RF) post-classification step. This integrated CNN-RF pipeline achieved overall accuracies of 83% for LU and 86% for LC. Large classes, such as dense forest and agricultural areas, reached individual F1 scores above 90%, effectively reducing the manual workload. Our results show that fusing multiple data sources can substantially improve accuracy over CNNonly aerial image classification, demonstrating the potential of application to similar national or regional LU/LC mapping efforts worldwide.