Prior Knowledge-Based Intelligent Model for Lithology Classification

Vegetation coverage can weaken the lithology information and increases inter-class similarity, making it difficult to effectively extract key feature information for lithology classification. To address the above issues, this study proposes a lithology scene classification model based on prior knowledge and improved dense connected networks. The steps of the improved dense connected network includes: Extracting edge information through edge detection operators to enhance the extraction of detailed features such as lithology, texture, and edges; Extracting multi-scale data features and fusing them using a dense connected network with enhanced feature fusion; Adding a random mixed attention mechanism to efficiently combine channel and spatial attention, while capturing dependency relationships on channels and pixel level relationships in space, improving the model’s ability to focus on key feature information; Using label smoothing to balance the classification accuracy of different categories, making the accuracy of each category more average. Building a dual branch network based on an improved dense connected network, one is to construct the main branch based on dataset A, and the other is to introduce the prior knowledge of 1:250,000 scale lithology classification based on dataset B to construct auxiliary branches; Building unsupervised loss based on label association prior of two datasets; Adaptively fusing the supervised loss and unsupervised loss constructed by two branches to construct a lithology scene classification model based on prior knowledge and improved dense connected networks. A classification experiment was conducted using dataset A and dataset B. The results show that the proposed classification model can effectively classify the lithology of the coverage area, and its performance is superior to other classic scene classification models. Especially, the addition of prior knowledge significantly improved the accuracy.