Estimating the Spectral Reflectance of Natural Imagery Using Color Image Features

Relative spectral reflectance is an illumination invariant image feature that is related to many ecological phenom- ena that are difficult to measure, such as plant CO2 up- take. We describe a procedure to estimate the relative spec- tral reflectance of known subject using color image features. Through application, we show that this procedure produces accurate estimates in the presence of changing field condi- tions. Using this procedure, we can use imagers as sensors to measure natural phenomena that cannot be easily measured using any other available sensing modality. There are many important natural phenomena that tradi- tional sensors cannot measure directly. For example, accu- rately measuring a plant's rate of photosynthesis (release or absorption of CO2) requires encasing part or all of the plant in a chamber, regulating the air flow, and measuring the com- position of the air leaving the chamber. When direct mea- surement is difficult, imagers are the missing input required to accurately model natural phenomena. Imagers are typically avoided in traditional sensing appli- cations because they produce large quantities of uncalibrated data. The form of calibration required for an imager-based ecological sensor is dissimilar to that of typical sensors; there is no conveniently accessible reference that can be used to calibrate an imager used as a CO2 sensor, for example. We aim to take the first step in this calibration process: estimate the spectral reflectance of a known subject using an imager. We choose spectral reflectance because it had been shown that CO2 uptake is related to the plant's spectral reflectance (5). Other application have used spectral reflectance to suc- cessfully distinguish soil from vegetation (11) and clouds from land and ice sheets (14). Legleiter et. al. (9) even used spectral reflectance as the basis for estimating the depth of a river channel. In order to estimate the subject's spectral reflectance, we must account for the spectral power distribution (SPD) of the incident light. The general form of this calibration, known as color constancy (10), has traditionally been difficult. Unlike the general form, we assume that we have a single subject illuminated by a varying lighting conditions. We show that accurate estimates of a subject's spectral reflectance can be derived from images by modeling the possible illumination and relative reflectance spectra. Further, we show how to build these models from experimentally acquired data.