A high-precision boundary identification method and its application to coal mine fire zone boundary interpretation

Abstract

Boundary identification of field sources is an indispensable task for interpreting field data. Initially, people used the distribution characteristics of data to obtain boundary information of field sources, making it difficult to identify weak anomalies amidst strong background anomalies. To address this issue, automatic control filters based on a certain window size were employed to identify the distribution of field sources. However, this method's results heavily relied on the window size and were not well applicable to complex anomalies. In recent years, features reflecting the boundary information of field sources have mainly been derived from the derivatives of scalar field data. Then, the correspondence between imaging results and boundaries is utilized to identify the horizontal boundaries of field sources. Specifically, extreme values of magnetic anomaly horizontal derivatives and zero values of vertical derivatives correspond to geological body boundaries. Existing boundary identification methods mainly utilize a balanced boundary identification filter composed of the ratio of first-order horizontal and vertical derivatives to delineate the positions of geological bodies, but the method has lower resolution and generality. Therefore, this paper proposes combining boundary detection filters based on ratios of derivatives of different orders with multiscale unsupervised deep learning. This approach utilizes different orders of derivative ratios to obtain higher-resolution edge imaging results. Additionally, a combination of Deep Image Prior (DIP) and Generative Adversarial Network-None Local (GAN-NL) networks for multiscale unsupervised deep learning is established to determine the horizontal position of sources based on extreme values of edge imaging results. The multiscale DIP network is used to identify the source position, and a self-attention mechanism neural network is added to the DIP network to enhance its learning ability, which can remove noise without requiring a large amount of data label.

Keywords

• potential field
• edge recognition
• multiple scale
• unsupervised deep learning
• burning area