Identifying the categories of detritus collected from river sands is an important work in geological researches, including sediment source analysis, tectonic evolution and lithofacies palaeogeography. Among deep learning techniques developed in recent years, Convolutional Neural Network (CNN) can be applied to the detritus identification problem. However, due to both data insufficiency caused by the high cost of manual labelling, and data imbalance caused by the uneven distribution of different categories of detritus, existing CNN models are hindered to reach their best performance. In this paper, we propose a novel network architecture for the problem of detritus identification: Dual-Input Attention Network (DANet), which accepts both plane-polarized images and cross-polarized images of detritus as input, and uses Parametrized Cross-Entropy as the loss function in order to alleviate the poor performance of detritus identification caused by data insufficiency and data imbalance. Experiments based on the detritus collected from the Yarlung Zangbo River Basin prove both the effectiveness and potential of DANet for detritus identification.

识别从河沙中收集的碎屑类别是地质研究的重要工作，包括沉积物来源分析，构造演化和岩相古地理。在近年来开发的深度学习技术中，卷积神经网络（CNN）可以应用于碎屑识别问题。但是，由于人工标记成本高昂导致数据不足，以及不同类别碎屑分布不均导致数据不平衡，现有的CNN模型难以获得最佳性能。在本文中，我们针对碎屑识别问题提出了一种新颖的网络架构：双输入注意力网络（DANet），它接受碎屑的平面极化图像和交叉极化图像作为输入，并使用参数化交叉熵作为损失函数，以减轻由于数据不足和数据不平衡而导致的碎屑识别性能下降。基于从雅鲁藏布江流域收集的碎屑进行的实验证明了DANet在碎屑识别方面的有效性和潜力。