A comparison of NDVI intercalibration methods

Sensor differences pose a challenge when using normalized difference vegetation index (NDVI) data calculated from different sensors. Determining an optimal intercalibration strategy is critical whenever a long-term comparison of NDVI record is required. In this context, the current study evaluated four intercalibration methods, namely linear regression (LR), quadratic regression (QR), neural network (NN), and radiative transfer (RT). Overall, the LR method performed less effectively over non-vegetated surfaces. The QR method yielded a comparable result to the NN method, indicating an excellent performance of these nonlinear methods. These statistical methods generally yielded unbiased NDVI values, whereas the RT method provided a high degree of correlation between the NDVI values (coefficient of determination, R-2 = 0.997). On the other hand, data-processing schemes had a large impact on NDVI intercalibration. The distributed scheme ('Band-to-NDVI') was more accurate than the lumped scheme ('NDVI-to-NDVI'). The differences were minimal for the RT method, followed by the NN, QR, and LR methods. The large differences associated with the statistical methods were likely due to the different behaviours of the spectral band differences in the red and near-infrared bands. Our findings can be useful in determining the optimal NDVI intercalibration methods and schemes for using long-term NDVI record.