Why prediction of NIR spectral train and test datasets different when use PLSREGRESS

Hi,

I am using NIR data to build a model. I split the datasets into the train(80%) and test (20%). Then I used the training dataset and Partial Least Squares (PLS) regression model. I got quite good agreement between observation and predictions, but when I used test datasets to test the model, I notice that observation and predictions are not showing expected agreement.

Using training dataset:

Using test data:

Use Cross-validation to select the number component for plsregress function (I chose 15 for this datasets):

I would appreciate if someone helps me in this case. Thanks in advance.