A small-patched convolutional neural network for mangrove mapping at species level using high-resolution remote-sensing image

The understanding of mangrove forest structure and dynamics at species level is essential for mangrove conservation and management. To classify mangrove species, remote-sensing technologies provide a better way with high spatial resolution image. The spatial structure is usually viewed as effective complementary information for classification. However, it is still a challenge to design handcrafted features for mangrove species due to their non-structure texture. To leverage the advantage of convolutional neural networks (CNNs) in abstract feature exploration, a small patch-based CNN is proposed to overcome the requirement of fixed and large input which limits the applicability of CNNs to fringe mangrove forests. The function of down-sampling technology was substantially reduced to make deeper network for small input in our work. Meanwhile, the inception structure is used to exploit the multi-scale features of mangrove forests. Furthermore, the network is optimized with lesser convolution kernels and a single fully connected layer to reduce overfitting via reducing the training parameters. Compared to the features of grey level co-occurrence matrix with support vector machine, our proposed CNN shows better performance in classification accuracy. Moreover, the differences between mangrove species can be perceptive via CNN visualization, which offers better understanding of mangrove forests.