Dynamic Adversarial Adaptation Network with selective pseudo-labels for enhanced unsupervised domain adaptation in rock microscopic image analysis

The critical role of lithology classification in reservoir exploration is increasingly germinating interest in intelligent rock image classification applications. Nonetheless, the efficacy of these classification methods predominantly hinges on the premise that the independent and identically distributed principles underlie the training data. Real -world training data, amassed from disparate geographical regions, tend to manifest nonuniform distributions, which is a factor that muddles the accuracy of contemporary classification strategies. In light of the aforementioned scenario and considering the scarcity of labeled data in the target domain, we suggest a selective pseudo -label -based domain adversarial adaptation. This approach earmarks the generation of an adversarial network that aligns both the conditional and marginal distributions of the source and target domains concurrently during adversarial training. By infusing a dynamic factor, we straddle an adaptable balance between distributions and integrate label information from the target domain using selective pseudolabeling within the adversarial training phase. Notably, our class -wise pseudo -label technology application determines the pseudo -label threshold dynamically via the selection factor. The thorough evaluation of our proposed approach across four distinct geographic areas encompassing 12 tasks demonstrated that our model produced an average accuracy of 72.2%, signaling a 30% improvement over the source -only model. Importantly, relative to six traditional Unsupervised Domain Adaptation (UDA) algorithms, our approach performed remarkably better, recording an average accuracy surge of a minimum of 3%. Sensitivity assessments also portrayed the robustness of our model in responding to parameter alterations, causing a trivial 1% accuracy fluctuation.