I think creating another new features from each of the aforementioned features will solve the problem of high correlation rather than removing or dropping one of the features, since both features might likely have a lot of impacts in predicting the target variable. After creating these new features, both features can then be dropped.