Machine Vision for Measurement of Rice Canopy Dimensions

Abstract

A preliminary study was conducted to evaluate the potential of a machine vision based system for nitrogen (N) requirements in rice (Oryza sativa). The system consisted of an image acquisition unit (IAU), camcorder, frame grabber, and a microcomputer. The IAU was designed and constructed for staging images of the plants using natural light in flooded field conditions. An 8 mm camcorder was used to acquire an analog image of the end-view of an enclosed section of row 28 cm (11 in.) long. Image processing techniques were developed to classify plant and background pixels in images of rice plants acquired in a field experiment involving two rice cultivars, Millie and Lemont. The naturally illuminated images were characterized by uniformly varying pixel intensity values. Background and plant pixels from several locations in the image were sampled to create spatially varying mean intensity vectors for the two pixel categories. These vectors were used with the minimum distance classifier to assign pixels as either plant or background pixels. Mathematical morphology techniques were used to remove shadows and isolated noise pixels. The segmented images were then used for measuring plant features. Individual image extracted plant features such as plant height, width, and area correlated well with the corresponding manual measurements, which provided evidence that the segmentation algorithm could correctly identify plant pixels. For Millie, the correlation coefficients (r) between the individual manual measurements and the corresponding image extracted feature using the segmentation algorithm were 0.896, 0.874 and 0.885 for height, width, and area, respectively. Similarly, the corresponding r values for Lemont were 0.831, 0.706, and 0.790.