Error 999999: something unexpected caused the tool to fail when performing Deep Learning analysis

Error Message

In ArcGIS Pro, Error 999999 can be returned for a number of reasons when performing analysis with Deep Learning.  Common causes and suggestions to resolve this error are as follows.

Cause

1. Ensure that the correct Deep Learning Framework is installed before attempting any Deep Learning workflows. This varies depending on the installed ArcGIS Pro version.
2. Test with sample data, the Palm Tree Detection training is recommended.
3. Ensure that the proper setting is made for the GPU or CPU to run the Detect Objects Using Deep Learning tool.
4. The training samples are being applied to imagery that is in a different format than the imagery where the training samples were collected.
5. There are not enough training samples in the collection to properly evaluate the image being analyzed. 
6. Ensure that the imagery used to collect the training samples is from the same region or area as the image to which the training samples are being applied. 

Solution or Workaround

1. Installing the Deep Learning Framework is necesssary for most functions in Deep Learning. It is important to match the installer version to your ArcGIS Pro version:  Download the proper version of the framework at the following link: Deep Learning Libraries Installers for ArcGIS
2. The Palm Tree Detection training sample assesses the health of palm trees, and serves two purposes.  Firstly, it helps to ensure that the software is installed properly to run the process, and secondly, it allows one to become familiar with this complex analysis.
3. When setting the environment parameters for the Detect Objects Using Deep Learning tool, set the Parallel Processing Factor environment only if Processor Type is set to CPU. There is no need to set the Parallel Processing Factor environment if Processor Type is set to GPU. By default, the tools in the Deep Learning toolset use GPU processing to perform analyses. Refer to ArcGIS Pro: An overview of the Deep Learning toolset for more information. See also: Unable to execute the Detect Objects Using Deep Learning tool in ArcGIS Pro
4. Ensure that the the format of the imagery matches that where the training samples were collected, .tif vs. .jpg for example. The bit depth and number of bands for the data where the training samples were collected must also match the parameters of the image being analyzed.
5. The Rule of 10 should be applied. Each degree of freedom in the analysis - the number of feature types that are being searched for - must be multiplied by ten to calculate a reasonable number of training samples required for the analysis.
   For example, to locate well pads, at least ten training samples of well pads from the original image are needed. When looking for the following four feature types in one classification process, at least 40 training samples would be necessary:
   • palm trees
   • swimming pools
   • manhole covers
   • fire hydrants
6. If the imagery used to collect the training samples is not from the same region or area as the image to which the training samples are being applied, the analysis may fail. As an example, it would be inadvisable to collect training samples around Palm Springs, California, which is in a desert, then try to apply those training samples to imagery in the state of Florida. Even though there are common features that exist in both places, such as palm trees, Florida is a lot wetter than Palm Springs and environmental moisture can have a profound effect on the results of the analysis.