Root Image Processing Lab
Plant and Soil Science Department
Michigan State University
Director Dr. Alvin Smucker

Computer Algorithms for Quantifying Video Recordings of Plant Root Images

Vera Bakic (CPS), Dr. Alvin Smucker (CSS), Dr. George Stockman (CPS)

Pattern Recognition and Image Processing Lab
Computer Science Department
Michigan State University

and

Root Image Processing Lab
Plant and Soil Science Department
Michigan State University

Minirhizotron image analysis

• Step 1 Images are recorded in the field by sliding a micro-video camera through minirhizotrons, which are clear plastic tubes inserted into the soil.

• Step 2 Images are manually or automatically digitized from the video tape and saved to the disk by the experiment name, date of recording, and the tube number.

• Step 3 Images are processed by MR-RIPL one by one.

• Step 4 Results for individual images are organized at the level of experiment performed, date of recording, and the tube number.

MR-RIPL 2.0---Ridge detection algorithm

• Based on the previous algorithm developed by John Ferguson (CSS) for VICOM image processor.

• Centerlines of the roots are found as zero crossing of the first derivative (c), (d).

• Centerlines are measured for length, average width, parallelism, intensity, and angle change (e).

• Classify each centerline as root or background based on the above features: (f) white lines are background, colored lines are roots, different colors are used for different width classes.

(a) Original image	(b) Enhanced image	(c) Peaks and gradients
(d) Centerlines	(e) Centerlines and edges	(f) Final segmentation

MR-RIPL 3.0---``Mole'' filter algorithm

• Figure above shows the image intensities in blue (high intensities denote root) and two moles.

• The red mole cannot be placed underneath the blue surface, that is the image has no bright object at the mole's location, and the filter gives no response.

• The green mole can be placed underneath the blue surface, that is, the image has bright object at the mole's location, and the filter gives a response.

• The mole filter is applied to the histogram equalized image (c), so that unique mole can be used for all images.

• The mole filter response (d) is trimmed with edge information to obtain accurate width information (e).

• The resulting image (e) is thinned (f). Neighbouring lines with similar slopes are connected into one line to obtain longer centerlines (g).

• Centerlines are measured for length, average width, parallelism, intensity, and angle change. Classify each centerline based on the above features: (h) white lines are background, colored lines are roots, different colors are used for different width classes.

(a) Original image	(b) Enhanced image	(c) Histogram equalized
(d) Mole filter response	(e) With edge information
(f) Thinned	(g) Centerlines	(h) Final segmentation

Classification methods and data sets used

• Data set Training and testing data consisted of 296 images, chosen randomly from 5 video tapes of different plant species. For training 118 images were used and 178 for testing.

• Ridge detection Four times more background length is obtained using this algorithm. Thus, the main objective is to minimize Type I (false alarms) error.

  A classification tree has been constructed, which separated data into 20 classes, using length and width information. Two classes were pure root or background. Discriminant analysis (DA) was used for the final classification of 18 classes.

• Mole filter This algorithm results in roughly equal length of root and background centerlines, and Type I (false alarms) error can be slightly higher than in ridge detection, and still be admissible.

  A classification tree has been constructed, which separated data into 8 classes, using length and intensity information. Five classes were pure root or background. DA was used for 3 classes.

Results of MR-RIPL

• The error is measured in root/background length misclassified.

• Manually classified vs. reported root length. The '+' sign means more length has been reported than manually classified.

Discussion of MR-RIPL results

• Type II (false dismissals) error varies a lot for different tapes. The reason is that different plant species have different root morphologies and they are grown in different soil types.

• The general trend is that error rates are lower for simple backgrounds and little root overlap. However, for complex backgrounds or significant root overlap, the error rates are high in both algorithms.

• The ridge detection algorithm was not acceptable, because it was very hard to design good classifier. In order to decrease Type II (false dismissals) error, Type I (false alarms) error would have to be very high.

• For mole filter algorithm it was possible to decrease Type II (false dismissals) error , without increasing Type I (false alarms) error significantly.

• Even though the mole filter does not give perfect results, it outperforms the ridge detection approach for the overall result.

Sample images

(a) Original image	(b) MR-RIPL 2.0	(c) MR-RIPL 3.0