2510 Rate this article:

New ENVI Deep Learning Object Detection Tool Reduces Labeling and Training Time

Jeremy Metcalf

The newest release of ENVI® Deep Learning can drastically reduce labeling and training time. If your features of interest are objects rather than landcover, the new object detection tool in ENVI Deep Learning 1.2 provides the ability to create object detection models alongside the existing pixel segmentation. Object labeling is performed by creating vector bounding boxes to encapsulate areas of pixels. You can also easily convert existing pixel segmentation projects to object detection training images by using the “Build Object Detection Raster from ROI” tool. It’s as simple as choosing a starting point and an ending point.

This blog outlines a disaster response project where I mapped rooftop tarps following Hurricane Delta in 2020. When I started the project, the latest imagery available was collected just after Hurricane Delta made its way through the state of Louisiana causing $3 billion in damage. Initially I made a pixel segmentation Deep Learning model to classify blue and green tarps using aerial imagery hosted by the NOAA’s Emergency Response Imagery website https://storms.ngs.noaa.gov/. This dataset contained 1,233 tiled images covering 2.8 by 2.8 km with a spatial resolution of 20 cm.

Analyze large areas of interest with ENVI Deep Learning

Figure 1: Spatial extent of imagery collected and hosted by NOAA Storms following Hurricane Delta in October 2020.

use envi deep learning to quickly identify objects

Figure 2: Pixel segmentation results uploaded to ArcGIS Portal in vector format showing accurate delineation of tarps.

Out of the 1,233 tiled images, I chose six tiles containing damaged residential areas to label both blue and green tarps to train a pixel segmentation deep learning model. Once the model was trained, I used ENVI Modeler to batch classify each of the 1,233 tiles, convert the classification images to shapefiles and then upload to ArcGIS Portal for public dissemination. The entire process took approximately 24 hours, including the time to label the tarps, train the deep learning model, and batch process the vector products.

When the ENVI Deep Learning module added object detection in version 1.2, I wanted to see if the same tarp detection project workflow would realize a benefit in terms of speed. The short answer: It did. Training a new object detection model was straight forward. I just needed to convert the labeling completed for pixel segmentation to bounding boxes, and this was easily accomplished using the “Build Object Detection Raster from ROI” tool. This tool asks for an image and associated Regions of Interest (ROIs) as input and the resulting product is an image with labels shown as bounding boxes. Once I had converted my six labeled images, I trained a new deep learning model to find blue and green tarps and output the results as vector bounding boxes. This bypassed the time it took to convert a raster classification into vectors.

Quickly create object detection workflows


Figure 3: Bounding box results for tarps after converting a pixel segmentation Deep Learning workflow to an object detection workflow.

After creating both pixel segmentation and object detection workflows, I found that the object detection workflow lends itself to detecting rooftop tarps since I just needed to know the location of the tarps and not an accurate delineation. Having the outputs already in vector format allows for a quicker turnaround for dissemination of tarp locations to disaster response stakeholders.

If you missed the “What’s New in ENVI Deep Learning 1.2” webinar, the video can be found here.





We offer the largest selection of geospatial products worldwide.



Process and analyze all types of imagery and data.



Get faster, more accurate results with ENVI Deep Learning.