EMFNet: Enhanced Multisource Fusion Network for Land Cover Classification

Feature extraction and fusion are two critical issues for the task of multisource classification. In this article, we propose an enhanced multisource fusion network (EMFNet) to address them in an end-to-end framework. Specifically, two convolutional neural networks are employed to extract features from two different sources. Each network is mainly comprised of three convolutional layers. For each convolutional layer, feature tuning modules are designed to enhance the extracted feature of one source by taking advantage of the other source. After getting the features of two sources, a weighted summation method is used to fuse them. Considering that fusion weights should vary for different inputs, a feature fusion module is designed to achieve this goal. In order to test the performance of our proposed EMFNet, we compare it with state-of-the-art fusion models, including the traditional models and the deep-learning-based models, on two real datasets. Experimental results show that the EMFNet can achieve competitive classification results in comparison with them.