Why accuracy over each k-fold cross-Validation differ alot? how can i improve that

Question

I have a data set with 91 Obervsation with 700 features. I have reduced the dimension of data using PCA. Then I split the data-set into training and test with ratio= 70/30.