Unsupervised Learning-Based Interpreting Architecture for Remote Sensing Imagery

Le résumé fourni par la source

The scarcity of labeled training data can significantly limit the performance of Remote Sensing (RS) interpreting. To address this limitation, studies have explored unsupervised methods to fully leverage latent representations from additional unlabeled data. However, in real interpreting scenarios, unsupervised methods are hardly employed due to the complex architecture. To this end, we propose an explicit unsupervised learning-driven architecture for RS interpreting, structured around four key stages: encoder design, pre-training, decoder design, and fine-tuning. First, the encoder is carefully designed to capture essential features from input data. Next, an appropriate unsupervised learning method is employed to pre-train the model, utilizing a customized pre-training structure. Subsequently, a task-specific decoder is developed to cater to various RS interpreting needs. Finally, the pre-trained model is fine-tuned and applied to specific interpreting tasks. The architecture was applied to road extraction in the Jiangbei New Area of Nanjing City. The experiments have shown the importance of the pre-training in RS interpreting and the flexibility of the proposed architecture, providing an effective way for practical RS interpretations.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.