Classification of rhizosphere components using visible–near infrared spectral images

Abstract

To establish new techniques for automatic classification of rhizosphere components, we investigated the utility of visible (VIS) and near–infrared (NIR) spectral images of the rhizosphere under two soil moisture conditions (mean volumetric water content: 0.39 and 0.16 cm³ cm⁻³). Spectral reflectance images of the belowground parts of hybrid poplar cuttings (Populus deltoides × P. euramericana, I45/51) grown in a rhizobox were recorded at 120 spectral bands ranging from 480 to 972 nm. We examined which wavelengths were suitable and the number of spectral bands needed to accurately classify live roots of four age classes, dead roots, leaf mold, and soil. VIS reflectance (<700 nm) of live roots first increased and then decreased with age, whereas NIR reflectance (≥700 nm) was stable in mature roots. The reflectance of dead roots was lower than that of mature roots in both the VIS and NIR spectral regions. VIS reflectance did not differ among dead roots, leaf mold, and soil, but the NIR reflectance was clearly lower in soil than in the other materials. The reflectance of leaf mold and soil increased mainly in the NIR spectral region with reducing soil moisture, but this increase did not affect the order of reflectance intensity among the rhizosphere components in general. Although the most suitable spectral bands statistically selected for classifying rhizosphere components differed somewhat between moist and dry conditions, the spectral bands 580–679 nm (VIS) and 848–894 nm (NIR) provided high reliability under both conditions. Classification accuracy was higher when using two to five VIS–NIR images (overall accuracy ≥87.8%) than three VIS images (red, green, and blue; accuracy <67.1%). The high accuracy with VIS–NIR was mainly due to successful separation of leaf mold and soil. Irrespective of soil moisture condition, the overall accuracy tended to be stable at 92–94% with use of four VIS–NIR images. The spectral bands effective in wet soil conditions could also be used for classification in dry conditions, with overall accuracies >86.9%. These results suggest that automatic image analysis using VIS–NIR images at four spectral bands, including red and NIR, allows for accurate classification of the growth stage or live/dead status of roots and distinguishes between leaf mold and soil.