Interference from nonspecific binding imposes a fundamental limit in the sensitivity of biosensors that is dependent on the affinity and specificity of the available sensing probes. The dynamic single-molecule sensing (DSMS) strategy allows ultrasensitive detection of biomarkers at the femtomolar level by identifying specific binding according to molecular binding traces. However, the accuracy in classifying binding traces is not sufficient from separate features, such as the bound lifetime. Here, we establish a DSMS workflow to improve the sensitivity and linearity by classifying molecular binding traces in surface plasmon resonance microscopy with multiple kinetic features. The improvement is achieved by correlation analysis to select key features of binding traces, followed by unsupervised k-clustering. The results show that this unsupervised classification approach improves the sensitivity and linearity in microRNA (hsa-miR155-5p, hsa-miR21-5p, and hsa-miR362-5p) detection to achieve a limit of detection at the subfemtomolar level.