ArcGIS Blog

Imagery & Remote Sensing

ArcGIS Pro

What's new for GeoAI in the Image Analyst extension of ArcGIS Pro 3.4

By Pavan Yadav and Sangeet Mathew

The latest ArcGIS Pro version brings exciting advancements in GeoAI for imagery. Discover the top features we’ve added to enhance your workflow.

Expanded support for external training data

The Train Deep Learning Model tool now offers flexibility in training data sources for object detection. Besides the data from the Export Training Data For Deep Learning tool, it can now use external data in Pascal Visual Object Classes or KITTI rectangles formats. Just organize the data into images and labels folders. This update makes the tool compatible with numerous open-source and commercial software tools that generate training data differently, thus enhancing its versatility and efficiency.

A Pascal VOC format training data example with training data organized in images and labels folders. The xml on the right shows an example label file.
A Pascal VOC format training data example with training data organized in images and labels folders. The xml on the right shows an example label file.

Enhanced AI-assisted labeling

The AI-assisted labeling experience has been enhanced by removing anomalies in the detections, as well as being able to automatically label image collections. By turning on the Remove Anomalies option in the AI- assisted labeling properties, the tool now automates the process of removing most, if not all, of the false positive detections.

Support for new foundation models

ClimaX is a Vision Transformer (ViT) based deep learning model that uses diverse datasets that cover various weather variables across different spatial and temporal resolutions. This foundational model can be fine-tuned for a broad range of climate and weather applications, including tasks involving atmospheric variables and spatio-temporal details not encountered during the pretraining phase.

An example multidimensional dataset used to predict weather change using ClimaX
An example multidimensional dataset used to predict weather change using ClimaX
ClimaX being used in ArcGIS Pro for sea surface temperature analysis
ClimaX being used in ArcGIS Pro for sea surface temperature analysis

Prithvi-100m is a cutting-edge temporal ViT, now accessible as a foundation model within ArcGIS Pro. It is trained on extensive Harmonized Landsat and Sentinel-2 (HLS) data. This model employs a self-supervised encoder built upon a ViT architecture and Masked AutoEncoder (MAE) learning paradigm. The model incorporates both spatial and temporal attention mechanisms to effectively process and analyze image data.

ViT architecture + 3D Patch embedding + 3D positional encoding source: https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M on 9/16/2024
ViT architecture + 3D Patch embedding + 3D positional encoding source: https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M on 9/16/2024
An example output from the Prithvi-100M-sen1floods11 pretrained model by fine-tuning Prithvi-100m with the Sen1Floods11 dataset, in which the blue color represents floodwater
An example output from the Prithvi-100M-sen1floods11 pretrained model by fine-tuning Prithvi-100m with the Sen1Floods11 dataset, in which the blue color represents floodwater

Object detection on oriented imagery

Detecting features with oriented imagery data has been a challenge using deep learning. Now the Detect Object Using Deep Learning tool has been enhanced to accept oriented imagery datasets as input. The tool detects features within an oriented imagery dataset in pixel space and then projects them to map space.

An oriented imagery dataset viewed in ArcGIS Pro with detected features
An oriented imagery dataset viewed in ArcGIS Pro with detected features

Feature extraction more accessible

Easily extract features from imagery with the Extract Features tool, which is now accessible from the Imagery tab. This powerful tool offers a variety of ready-to-use models, options to use your own custom models, and a range of inferencing and postprocessing options for enhanced output quality.

The Extract Features tool is now accessible from the Imagery tab
The Extract Features tool is now accessible from the Imagery tab

Expand your analysis: append to existing outputs

The Detect Objects Using Deep Learning tool and the Classify Objects Using Deep Learning tool now support appending new results to existing output feature classes. This is useful when you’ve already processed one area and want to expand your analysis to adjacent regions. Simply run the tool again, and the new results will be added to the existing feature class.

Customize object detection: focus on what matters

When using the Detect Objects Using Deep Learning tool with models capable of identifying multiple object types, you can specify the exact objects of interest. This allows you to tailor the tool’s output to your needs and improve efficiency.

Share this article

Subscribe
Notify of
0 Comments
Oldest
Newest
Inline Feedbacks
View all comments