Image-by-image calibration of thermal infrared data

Surface temperature retrieval through thermal infrared imaging (TIR) is currently being applied in a multitude of natural science disciplines (e.g. agriculture and ecology). Enormous progress in sensor design, electronics, and computer science render TIR systems as easily applicable and accessible for research, industry and even private use. However, despite the existence of factory-set theoretical calibration models that are supposed to facilitate accurate conversion of digital pixel values into temperature values, complex environmental noise and handling parameters hamper the reliable and stable collection of temperature data. Here, we present an image-by-image calibration method that can potentially account for such environmental noise.

We used custom-built thermal calibration panels, a close-range thermal camera and a thermocouple to collect thermal images and ground truth temperatures of peanut plants (Arachis hypogaea L.) in an open and sheltered agricultural setting. Linear models were trained and tested to investigate whether an image-by-image calibration approach improves the accuracy of digital value to temperature conversion over calibrating once before an experiment, as well as before and after an experiment. For both, the sheltered and open setting, we collected data on multiple days.

Our data indicate that there are marked differences in calibration model stability and accuracy between the open and sheltered setting. For the open setting the image-by-image calibration resulted in lower mean absolute temperature errors (MAE = 0.9°C) compared to the sheltered setting (MAE = 4.37°C). We also found that the intercept and slope of the image-by-image calibration models varied substantially under open conditions. Between two images, both captured less than two minutes apart, the digital number to temperature conversion (intercept) could vary by up to 15°C. By contrast, the intercepts derived from the sheltered scenario rarely varied by more than 5 °C.

Our results show that an image-by-image calibration can be preferable to obtain reliable and accurate temperature data. Such data can be crucial to monitor and detect abiotic and biotic stress in animal and plant food production systems where differences in temperature can be very subtle. A reduction of stressors in such systems is often coupled with an increase in yield. At the EGU 2020, we would like to share our research, and some extensions of it, to receive constructive feedback to drive future research on how to reliably and accurately collect sensitive surface temperatures in industry and research.