Feature extraction

Available with Image Analyst license.

Choose the Extract Features menu item from the Deep Learning Tool menu to open the Extract Features Using AI Model tool. The Extract Features Using AI Model tool extracts features of interest from imagery using pretrained AI models. The AI models can either be ArcGIS pretrained models from ArcGIS Living Atlas of the World or custom deep learning model packages (.dlpk). The tool accepts imagery and allows you to choose from several pretrained object detection or pixel classification models that extract features from the imagery. Several models can be chosen to extract multiple features of interest. In addition to detecting features and classifying pixels in the imagery, you can also use the tool to perform postprocessing on the output to generate refined results. You can also only postprocess previously generated results. The tool creates a group layer containing all the extracted features.

To run this tool, a machine hosting a GPU is required. If you have more than one GPU, specify the GPU ID instead.

Learn more about how Extract Features Using AI Models works

The Extract Features Using AI Models tool parameters are as follows:

Parameter	Description
Input Raster	The input raster on which processing will be performed. If the Mode parameter is specified as Only Postprocess, a raster with binary classification is required for this parameter.
Mode	Specifies the mode that will be used for the processing of the input raster. Infer and Postprocess—Features will be extracted from the imagery and postprocessed. This is the default. Only Postprocess—The input raster will be directly postprocessed. A single-band raster with binary classification is required for this option.
Output Location	The file geodatabase where the intermediate output from the models and the final postprocessed output will be stored.
Output Prefix	A prefix that will be added to the name of the outputs that will be saved to the output location. The prefix will also be used as the name of a group layer that will be used to display all outputs.
Area of Interest (Optional)	The geographical extent that will be used to extract features. Only features within the area of interest will be extracted.
Pretrained Models (Optional)	The ArcGIS pretrained models from ArcGIS Living Atlas of the World that can be used on the provided input raster. This parameter requires an internet connection to download the pretrained models.
Additional Models (Optional)	The deep learning models that can be used on the provided input raster and the postprocessing workflow that will be used for additional model files (.dlpk and .emd). Available postprocessing workflows are as follows: Line Regularization—The postprocessing workflow will extract line features from a single band raster with binary classification and generate a polyline feature class after refining it. This workflow also supports deep learning models that generate polyline feature classes. Parcel Regularization—The postprocessing workflow will extract parcels from a single band raster with binary classification and generate a polygon feature class after refining it. Polygon Regularization—The postprocessing workflow will generate a polygon feature class after refining it. This workflow is only compatible with object detection models. Polygon Segmentation—The postprocessing workflow will generate a polygon feature class containing the detected objects using their centroid or bounding box. The segmentation method is specified using the Prompt parameter. None—No postprocessing workflow will be applied. This is the default.
Confidence Threshold (Optional)	The minimum confidence of deep learning model that will be used when detecting objects. The value must be between 0 and 1.
Save Intermediate Output (Optional)	Specifies whether the intermediate outputs will be saved to the output location. The term intermediate outputs refers to the results generated after the model has been inferenced. Checked—The intermediate outputs will be saved to the output location. Unchecked—The intermediate outputs will not be saved. This is the default.
Test Time Augmentation (Optional)	Specifies whether predictions of flipped and rotated variants of the input image will be merged into the final output. Checked—Predictions of flipped and rotated variants of the input image will be merged into the final output. Unchecked—Predictions of flipped and rotated variants of the input image will not be merged into the final output. This is the default.
Buffer Distance (Optional)	The distance that will be used to buffer polyline features before they are used in postprocessing. The default is 15 meters.
Extend Length (Optional)	The maximum distance a line segment will be extended to an intersecting feature. The default is 25 meters.
Smoothing Tolerance (Optional)	The tolerance used by the Polynomial Approximation with Exponential Kernel (PAEK) algorithm. The default is 30 meters.
Dangle Length (Optional)	The length at which line segments that do not touch another line at both endpoints (dangles) will be trimmed. The default is 5 meters.
Input Road Features (Optional)	A road feature class that will be used for refining the parcels. The input can be a polygon or polyline feature class.
Road Buffer Width (Optional)	The buffer distance that will be used for the input road features. The default value is 5 meters for polyline features and 0 meters for polygon features.
Regularize Parcels (Optional)	Specifies whether extracted parcels will be normalized by eliminating undesirable artifacts in their geometry. Checked—Extracted parcels will be normalized. This is the default. Unchecked—Extracted parcels will not be normalized.
Post Processing Workflow (Optional)	Specifies the postprocessing workflow that will be used. Line Regularization—Line features will be extracted from a single-band raster with binary classification and a polyline feature class will be generated after refining it. Parcel Regularization—Parcels will be extracted from a single-band raster with binary classification and a polygon feature class will be generated after refining it. Polygon Regularization—A polygon feature class will be generated after refining it. This workflow is only compatible with object detection models.
Output Features (Optional)	The feature class containing the postprocessed output.
Tolerance Between Adjacent Parcels (Optional)	The minimum distance between coordinates before they are considered equal. This parameter is used to reduce slivers between extracted parcels. The default is 3 meters.
Regularization Method (Optional)	Specifies the regularization method that will be used in postprocessing. Right Angles—Shapes composed of 90° angles between adjoining edges will be constructed. This is the default. Right Angles and Diagonals—Shapes composed of 45° and 90° angles between adjoining edges will be constructed. Any Angles—Shapes that form any angles between adjoining edges will be constructed. Circle—The maximum distance from the boundary of the feature being processed will be used.
Tolerance (Optional)	The maximum distance that the regularized footprint can deviate from the boundary of its originating feature. The default is 1 meter.
Prompt (Optional)	Specifies the segmentation method that will be used when the Additional Models parameter is set to Polygon Segmentation. Centroid—The centroid of the detections will be used to indicate to the polygon segmentation model what to segment in the input raster. Bounding Box—The bounding box of the detections will be used to indicate to the polygon segmentation model what to segment in the input raster. None—No segmentation method will be used. This is the default.
Input Features (Optional)	The feature class on which postprocessing will be performed. This parameter is only supported when the Post Processing Workflow parameter is set to Line Regularization or Polygon Regularization.
Output Summary (Optional)	The table that will contain a list of outputs that were generated along with their respective paths.