Exploring Land Use Mapping with Multimodal Data Fusion and Convolutional Neural Network

Abstract. Accurate and efficient land-use mapping provides intuitive spatial information, which helps to rationalize the planning and deployment of land resources, and provides a basis for urban planning, agricultural development, environmental protection and other aspects. This study utilizes the Google Earth Engine platform and the Resnet-50 method to explore the spatial distribution of land use in Daxing District, Beijing in 2023, by combining point of interest (POI) data, nighttime light data, Sentinel-1 data, and Sentinel-2 data. The results of the study show that the accuracy of land use mapping using different data is different, and the accuracy of the Resnet-50 method is better than that of the Random Forest method. Making full use of the band features and index features of Sentinel-1 data and Sentinel-2 data, nighttime light data and POI data can improve the accuracy of land use mapping results. Among them, the land use mapping accuracy of the proposed method is the highest, with an OA of 88.11% and a Kappa coefficient of 0.83. Ranking the importance of different features found that VH band in January-March has the most important effect on the land use mapping results and the residential land in the POI data has the least important effect on the land use mapping results. This study provides a feasible reference program for efficiently and accurately obtaining land use mapping data for a large study area.