Why i Get low accuracy when i give unseen data to Trained Model?

I have combine dataset of signals which have 14 classes. I have split them using

imds = imageDatastore('E:\SNR-Dataset\DATA-11-time\Data-for-training\', 'FileExtensions', '.mat', 'IncludeSubfolders',true, ...
                      'LabelSource','foldernames',...
                      'ReadFcn',@matReader);
[imdsTrain,imdsValidation, imdsTest] = splitEachLabel(imds,0.7,0.2, 'randomized');

.
.
.
[net2,tr] = trainNetwork(augimdsTrain,lgraph,options);

imdsTest_resize = augmentedImageDatastore([224,224],imdsTest);
[YPred,probs] = classify(net2,imdsTest_resize);

accuracy = mean(YPred == imdsTest.Labels)

Whenever i use imdsTest from splitEachLabel it give me 99% accuracy (Note that the train validation and test are in one folder)

I have unseen data which save in different folder and i use the following code to check the model accuracy on unseen data

imdsTest1 = imageDatastore('E:\SNR-Dataset\DATA-11-time\snr-test-data\Final-Test-data\snr30', 'FileExtensions', '.mat', 'IncludeSubfolders',true, ...
                      'LabelSource','foldernames',...
                      'ReadFcn',@matReader);
imdsTest_resize1 = augmentedImageDatastore([224,224],imdsTest1);

[YPred,probs] = classify(net2,imdsTest_resize1);

accuracy = mean(YPred == imdsTest1.Labels)

i got the 30% test accuracy

Please Assist why i get low accuracy when testing a model on unssen data which are in saparate folder?

Answer 1

[yanqi liu]

may be modify layers，add some dropoutLayer

if possible，may be upload data and code to debug

Answer 2

[john karli]

I am using the same link for data generation and training a model. the above code is modified version of the below link. you can generate the data and test it.

https://www.mathworks.com/help/deeplearning/ug/modulation-classification-with-deep-learning.html

Out of 10,000 sample. I have use first 5000 samples per modulation scheme for training. and used last 500 (9,501:10,000) for testing purpose (saved in different folder).