Exploring geo-tagged photos for land cover validation with deep learning

摘要: Abstract Land cover validation plays an important role in the process of generating and distributing land thematic maps, which is usually implemented by high cost sample interpretation with remotely sensed images or field survey. With increasing availability geo-tagged landscape photos, automatic photo recognition methodologies, e.g., deep learning, can be effectively utilised for applications. However, they have hardly been processes, as challenges remain selection classification highly heterogeneous photos. This study proposed approach to employ photos using learning technology. The first identified automatically based on VGG-16 network. Then, samples were selected further classified considering distribution probabilities. implementations conducted GlobeLand30 product a area, western California. Experimental results represented promises validation, given that showed overall accuracy 83.80% samples, was close result 80.45% visual interpretation. Additionally, performances ResNet-50 AlexNet also quantified, revealing no substantial differences final results. ensures quality, supports strategy distribution, improvement from 72.07% 79.33% compared solely single nearest photo. Consequently, presented proves feasibility technology information identification has great potential support improve efficiency validation.