This release includes the following new and improved features.

Highlights

New Licensing Engine

This version includes a new licensing engine. The same activation codes you used to activate your previous version of ENVI Deep Learning can also be used to activate your new license. If you still have your legacy license installed, the License Administrator will be able to detect it and migrate your license.

TensorBoard Updates

ENVI Deep Learning Preference Changes

With this release, TensorBoard automatically launches when training begins for Pixel Segmentation and Object Detection, and it reports detailed metrics.

With this change, the following settings were removed from File > Preferences > Deep Learning:

  • Compute Training Metrics

  • Display TensorBoard During Training

Updated Metrics

ENVI Deep Learning now uses many features in TensorBoard which provide additional insight throughout the training process. Metrics are also now consistent between Pixel Segmentation and Object Detection.

  • Scalars: Previously, scalar metrics only offered overall loss and accuracy at the end of an epoch. ENVI Deep Learning 2.1 now has metrics for overall accuracy, loss, precision, and recall. Additionally, per class metrics are now available for accuracy, precision, and recall. All metrics are now reported per epoch step providing addition insight regarding an epoch’s performance. This provides real-time updates on the current state of training.

  • Images: Confusion matrices for both training and validation. These offers insight into how well the model is learning, and what classes may be causing confusion.

  • Distributions: Show distributions to model weights over time.

  • Histograms: Show Tensor changes over time.

  • Time Series: Show one model to multiple model differences over time.

See View Training Metrics for additional details.

Pixel Segmentation Updates

Model Initialization Changes

The following InitializeENVINet5MultiModel task changes affect the task dialog in the ENVI Modeler:

  • The dialog display name has been changed to "Initialize Pixel Segmentation Model." The routine name has not changed.

  • A new drop-down list called Architecture has been added, which has the options described in the next section (Unet++).

  • The prefix "Model" has been removed from the parameter names Name and Description.

  • The single-class model InitializeENVINet5Model task has been removed. Single-class and multiclass models are now created using InitializePixelSegmentationModel, which offers a Number of Classes parameter. Trained models that use the deprecated single-class architecture will still be supported for classification using the TensorFlow Pixel Classification task.

Unet++

ENVI Deep Learning now provides two architectures for training a Pixel Segmentation model:

  • SegUNet++ (new architecture)

  • SegUNet (original architecture)

The new SegUNet++ architecture is a denser network, filling in the space between the encoder and decoder with additional convolution layers. The purpose of the additional convolution layers is to reduce the feature map gaps in the encoder and decoder subnetworks. This can result in cleaner, and more accurate detections during classification.

Classification Raster Band Update

Pixel classification raster bands have been renamed to the following:

  • ENVI Deep Learning Classification: SegUNet

  • ENVI Deep Learning Classification: SegUNet++

  • ENVI Deep Learning Classification: Legacy Single-Class

Updated User Interface Progress Feedback

Progress Dialogs

With the updated TensorBoard training metrics, ENVI Deep Learning UI elements such as progress dialogs have been updated to be more responsive. Training dialogs now report the current epoch of total epochs, step of total steps complete for the epoch, and the loss value for the current step of the current epoch. This provides real-time information on training progress and performance.

  • Pixel and Object progress dialogs report consistent information.

  • Pixel and Object progress dialogs contain Pixel or Object in the dialog title.

Test Installation and Configuration

The Guide Map Tool Test Installation and Configuration now determines GPU capabilities based on driver versions, drivers detected, and GPU total memory. Users will be informed whether the GPU is capable for training and classification, only classification, or not suitable for ENVI Deep Learning.

New Machine Learning Training Parameters

New properties were added to the TrainExtraTrees and TrainRandomForest tasks. This documentation is located in the Machine Learning section of the ENVI Help TOC.

Properties added to TrainExtraTrees and TrainRandomForest:

  • CUSTOM_MAX_FEATURES

  • MAX_FEATURES

This additional property was added TrainRandomForest:

  • OOB_SCORE

Updated Machine Learning Classification Raster Band Names

The output classification raster band name has been updated to show the algorithm the model was trained with. For example:

  • ENVI Machine Learning Classification: RandomForest

  • ENVI Machine Learning Classification: BIRCH

  • ENVI Machine Learning Classification: LocalOutlierFactor