Tips for Pixel Segmentation
You should not try to mask out unwanted features such as clouds or water pixels in your input images. Raster masks are not supported in ENVI Deep Learning as they can lead to invalid classification results.
Labeling features and training mask-based models are not always easy processes. Although you can usually get good results from labeling only a few features and accepting the default training parameters, you will get much more accurate results by investing some time in these steps. Once you build a good model to find specific features, you do not have to retrain it. You can build it once and use it multiple times to classify different images (as long as they have similar spectral and spatial properties).
If you are not sure how to set training parameters, you can use the ENVI Modeler to try many random combinations. See Randomize Training Parameters for details.
The single-class training process accessed through the Deep Learning Guide Map uses the ENVINet5 architecture that was introduced in ENVI Deep Learning version 1.0. The multiclass training process uses a new ENVINet5Multi architecture that was introduced in the ENVI Deep Learning 1.1 release.
If you get a class activation raster that is completely black, it is possible that the model could not accurately reproduce the training data. Or maybe the training did converge but to an incorrect solution. If this happens, rerun the training step to see if it produces a valid result. Also try increasing the Max values for Class Weight and/or Blur Distance.
tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape[6,128,392,392] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc