基于深度学习的自然资源调查监测图斑辅助辨识方法

Auxiliary Identification Method for Natural Resource Survey and Monitoring Map Spots Based on Deep Learning

摘要:

自然资源调查监测所收集的图像数据部分存在标注不完整的情况，很多图像仅被赋予了有限的标签，导致难以全面学习到图像与自然资源类别之间准确映射关系，从而产生监测图斑辅助辨识效果差的问题。为此，本文提出了基于深度学习的自然资源调查监测图斑辅助辨识方法。通过监测图像空间到自然资源标签空间的映射，对自然资源调查监测图像数据标注及增强，防止过拟合问题。基于构建深度学习模型，结合自注意力机制对图斑分类与定位，分析图像与自然资源类别之间的映射关系。划分目标损失函数值，构建图斑辅助辨识判断式，实现图斑辅助辨识。由实验结果可知，该方法辨识的图斑点1、2、3、4对应的坐标位置分别为（0，30）、（60，12）、（60，55）、（40，30），与实际位置一致，且最短耗时为15 s，具有较强辨识性。

Abstract:

The image data collected by the Natural Resource Survey and Monitoring Institute suffers from incomplete labeling, where many images are assigned only limited tags, making it difficult to fully learn accurate mapping relationship between images and natural resource categories. This results in poor auxiliary identification performance for monitoring patches. In this paper, a deep learning-based method for auxiliary identification of natural resource survey and monitoring patches has been put forward. By mapping from spatial domain of monitoring images to the label space of natural resources, the method enhances the annotation of natural resource survey and monitoring image data and prevents overfitting. Based on constructing a deep learning model and incorporating a self-attention mechanism for patch classification and localization, the mapping relationship between images and natural resource categories has been analyzed. It divides the target loss function values, constructs a patch auxiliary identification formula, and achieves auxiliary identification of patches. As showed by experimental results, the identified patch points 1, 2, 3, and 4 correspond to coordinates (0, 30), (60, 12), (60, 55), and (40, 30) respectively. It is consistent with actual location. The minimum time consumption is 15 seconds with strong identification capability.