A Quantitative Assessment of Radiometric Calibration Errors on Crop Cover Classifications

Abstract:

Satellite Earth observation imagery is increasingly part of the decision making process for various sectors such as disaster response, urban planning, and agriculture. Each component in the data value chain, from the acquisition of this imagery by the hardware, orthorectification, radiometric and atmospheric corrections, classification and modeling, and the final analysis by the end-user, presents a potential opportunity for error to be introduced and amplified downstream. Small spacecraft, without onboard calibration systems, are increasingly providing high-frequency, high-resolution data that is being used for a variety of applications. It is therefore important to investigate and characterize how calibration errors can impact remote sensing data products in which sufficient and frequent calibrations may not have been performed. In this study, we investigate the impact of radiometric calibration errors on crop cover classification using Landsat 8 OLI data as a demonstration case. While Landsat data is calibrated with on-board systems, it serves as a useful case for investigating errors due to the availability of validated data sets and ground truth information. In the preliminary analysis, we investigated the digital number (DN) value to top of the atmosphere (TOA) reflectance conversion for Landsat 8 Level 1 precision and terrain corrected images, and the impact of errors on crop classification using the USDA Cropland Data Layer as ground truth. Radiometric errors for each band were introduced by perturbing the gain factors from the metadata. The error-introduced images were then classified using a classifier originally trained on the correctly-calibrated control image. Results using the Random Forest method indicate that perturbations of individual bands affect the validation accuracy of the classifier trained on the baseline, with bands 5 (NIR), 3 (Green), 1 (Coastal aerosol) and 4 (Red) being the most sensitive. For one specific image, perturbing the gain coefficient of the NIR band by 1% resulted in a loss in test accuracy of ~8%, while perturbing the green band by 1% resulted in a loss of ~5%. Additional data will be examined for advancing this analysis further using a variety of classification methods and cases.