Other Software and Data

Machine Learning and Image Processing

  • 1.  Machine Learning and Image Processing

    Posted 03-18-2019 15:00
    I was listening to an online learning webinar today on Machine Learning. For those of you who often use satellite imagery to do feature extraction, you might remember that when Landsat imagery first became available, that the challenge was how to discern urban areas from rural, sandstone from granite, trees from pastureland. The techniques used were generally supervised or unsupervised classification algorithms, that those of us who were using Landsat simply called "image processing" but now is called "machine learning."

    The technique of supervised classification called for identifying rock types, for example, by actually visiting the field area...we called this "ground truth." We took samples, photographed the area, then went back to the computer lab to identify the field area that we had visited by enscribing the area of pixels that we knew were "granite" vs. "sandstone" for example. Then we asked the computer to find areas that "looked like" these same sample sets that we knew were of a certain rock type. The result was a new map of a larger area that were delineated based on our ground truth and sample "training area" sets.

    Unsupervised classification simply allowed the computer algorithms to statistically find areas of similar spectral characteristics. This led to maps that were often "best guesses" at landforms, rock types or vegetative land cover. They were generally used as a "first pass" before supervised classification was understaken.

    This form of "machine learning" (aka image processing) was common and helped to define the early days of Landsat that formed the basis of land classification schemes. At the time, I was employed by the U.S. Geological Survey which undertook Landsat mapping for both domestic and international projects. Areas such as the Alaska National Wildlife Refuge (ANWR) and the mineralization of Nevada copper deposits set the stage for developing computer algorithms that were later refined to support further research.

    For fun, check out how machine learning is being used in pattern recognition for things as diverse as recognizing the difference between "puppies and bagels."

    Joe Francica
    Knowledge Community Shared Account
    Huntsville, AL