Image Analysis

Introduction to Automated Image Analysis and Biomass Estimation

The challenge. . .

The Automated MiniRhizotron (AMR) generates roughly 34K individual, high resolution (<10u detail) images per full tube scan. These images are displayed using our web-based RootView™ software using a mosaic created by thumbnails of each image. Our current production unit (the AMR-B) takes roughly 24 hours to capture a full tube of images. At this rate, it is easy to see how existing manual methods to identify and define structures and calculate estimated biomass can become tedious and overwhelming.

AMR instruments capture large numbers of root/soil images which must be inspected to determine the existence of roots and then measured to obtain quantitative measures of root properties. Because full tube scans can contain 34,000+ images each, and scans can be scheduled on a daily basis, and AMR machines are proliferating to the extent that hundreds of machines generate this data, we can predict that over a billion images can be stored annually. This far exceeds what humans can review and analyze manually.

What is required is an automatic method of scanning mosaics, tile groups, and tiles themselves that will detect the presence of roots and hyphae, and then automatically compute the biomass index for the image. Many other possibilities for new AMR features present themselves if automatic root detection (ARD) can be achieved.

METHODOLGY

Roots generally differ in visual appearance from the soil types they are in. The three principle differences are color, shape, and size.

Color: Roots have their own color characteristics, which vary from species to species. However, they are generally easily seen in an image because their characteristic color shades differ from that of the surrounding soil.

Shape: Roots are long and narrow, and therefore have a high aspect ratio. Soil, composed of granular mineral material, is generally “blob”-like, and have an aspect ratio closer to unity.

Size: Roots generally have a minimum volume, or visible area. To state this to the extreme, a 1-pixel object cannot be a root. Therefore, starting with an image, we want to select for the correct color space, the correct shape factor (related to aspect ratio), and a reasonable size.

The algorithm we have developed follows these selection criteria.

 

A small area (4×5 tiles) of a small mosaic (roughly 4,000 tiles) magnified to show detail.

 

 

One tile within the magnified area.

 

Characterizing a tube

As we can see, a tremendous amount of data is available even within a relatively small scan area.

To begin the automated analysis, a user defines the color space of the elements of interest. In this case, the goal is to “teach” the software what constitutes a fine root within the scan area of a given tube.

To begin, we select a representative image and identify the color space of the root(s). This is an iterative process with the original image file on the left and the color space on the right displayed in real time as each point is selected.

Here is result of choosing 14 points on the root.

With colors selected, the next step is to identify objects within the image that meet the size and shape requirements. To do this, we look for a minimum number of pixels as well as a reasonable aspect ratio. Items that are considered biomass are then rendered in green with quantitative results displayed on the far right.

Adding colors from multiple tiles results in more accurate object recognition. For this demonstration, color points were added from six of the 20 tiles in the mosaic resulting in a color space of 59 points. Colors may be viewed and deleted using the color management tool.

In this case, we have identified one object that meets the established criteria – Object #49 (of 191 candidate objects). Both mean diameter and length are displayed for each object. Length is calculated in one of two ways depending on the object shape. In most instances, objects are fitted with a polynomial where length is taken to be the greatest distance between vertices of the polynomial. In other cases one-half the object perimeter yields a more accurate result. The software automatically does this selection for the user.

Once satisfied with color, size and shape selections, the directory containing the image files is batch processed with the results from each tile within the mosaic analyzed and then combined for a volume (mm3) within the total image area (mm2).

Analysis of fungal hyphae is also likely possible for analysis as well, based on structure size (diameter <10u). We are still working on this functionality, but here is an example of where we are at this point.

As is obvious while watching the batch analysis run, biomass is under and over-estimated by as much 25% in each individual image. Fortunately, that percentage can be greatly reduced statistically due to the very large number of images captured by the AMR. Using the standard error of the mean formula (s/√n) it works out to be roughly 0.11% for an entire tube scan. [In this demonstration case it is 5.6%.]  With daily scans taken of the same areas, the statistical accuracy continues to improve with time.

Using automated numerical data rather than looking through and digitizing images by hand allows for easy entry into a database which then makes it possible to plot results against other relevant data including date and time, meteorological data and other soil data.

We are currenty testing the program with images captured over the past eight years in a wide variety of soil types. Our eventual goal is to incorporate it into our RootView™ product.

Initial testing indicates that our image analysis may also be compatible with images generated by other minirhizotron cameras. Should this be the case, we may also release a stand-alone version of the program.

Below are sample images from CID and Bartz processed through Root Detector. The high contrast in the CID image required only three color selection points.

The Bartz images are more problematic as the color is considerably red and green-shifted making differentiation from the background more challenging. In this case, we used color selection to exclude non-roots, essentially identifying the soil rather than roots.

An image from the ORNL SPRUCE site.