Lithological Scene Classification Based on Model Migration and Fine-Tuning Strategy

The distribution of lithologic categories has a significant uneven distribution, which means there may be lithologic categories in one domain that are not available in another domain. To tackle the problem of difficult identification of unabled lithology in cross regional prediction using conventional models, this article is based on the idea of transfer learning. To develop an improved dense connected network for the source domain and fine-tune the model with a small amount of data from target domain for achieving the lithology classification here. A classification experiment was conducted using dataset A1 and dataset A2. The results indicate that the proposed model is valid for identifying new lithology types that was not found in the source domain, and also improved the classification accuracy with only limited samples. OA and F1_score, and Kappa on the normal test set were 61.52 ± 0.95%, 55.58 ± 2.58%, and 52.18 ± 1.01%, respectively, and on a small sample test set were 47.40 ± 0.65%, 49.58 ± 0.41%, and 40.41 ± 0.45%, respectively, which were superior to the direct training.