ERDAS IMAGINE

History of machine learning

Machine learning has been in wide use in the geospatial industry since the 1980s. Processing satellite images using K-Means or ISODATA clustering algorithms was one of the first uses of remote sensing software like ERDAS IMAGINE. In more recent years, Bayesian networks, classification and regression trees (CART), support vector machines (SVM) and random forests have all been used to extract features from imagery in user-interactive workstations.

With the advent of powerful computers and proliferation of high-resolution images, deep learning has boomed. Deep learning technology is now used as one of the main technologies in addressing challenges in urban planning, agriculture, disaster management and defense.

Target/object detection and identification

Target/object detection and identification is one of the tasks geospatial analysts perform. With the increased amount of data being collected and shorter revisit cycles, manual identification of targets is becoming nearly impossible. Fortunately, this analysis can be performed faster using AI, by training a deep learning algorithm to detect and identify the targets/objects of interest. Once trained, the trained DL model can process new images faster and with higher accuracy. The work of the geospatial analyst is then be focused on quality control and final analysis.

Deep learning-based semantic segmentation

Pixel-level image classification is as old as images, and numerous techniques have been developed to perform the task. One of the drawbacks of pixel-based image classification is the salt and pepper effect in the results.

Deep learning-based semantic segmentation solves this problem by looking at the spectral and geometric relationships of neighboring pixels. Pixels that have similar spectral properties and form the shape of the training classes are grouped together to avoid the salt and pepper problem encountered in traditional classification techniques.

AI-based despeckle functionality

These two versions of a synthetic aperture radar (SAR) image clearly demonstrate the advantage of using AI-based despeckle functionality. The lower image, after despeckling, shows features at all scales more clearly defined and with better contrast. This cleanup is desirable for both human and machine interpretation.

The despeckle deep learning model is fully automatic; analysts do not need to make decisions to optimize the operation. This functionality is idea for rapid, automatic behind-the-scenes preprocessing of all radar images, regardless of their intended use.