Crop phenology date estimation based on NDVI derived from the reconstructed MODIS daily surface reflectance data

Crop phenological stage estimation based on remote sensing data is critical for evaluating crop progress, condition and crop yield. However, the coarse spatial and temporal resolutions of multi-day composited data products limit the phenology estimation accuracy. The finer resolutions mean more variations in the data. To solve this dilemma, this paper proposes to use NDVI and its derivatives derived from the 250m MODIS daily surface reflectance data MOD09GQ to estimate crop phenology stages. In this paper, the contaminated data of MOD09GQ are first filtered out using quality flag and cloud information from MOD09GA. The missing data are reconstructed with linear interpolation. To remove noise and to generate differentiable NDVI curve, a new temporally and spatially iterative smoothing procedure that uses Savitzky-Golay filter and area averaging is proposed and the double logistic function fitting method is also presented as a comparison. The phenology stages such as emerged, maturity, and harvest dates are detected from the NDVI curve and its derivatives while other phenological stages that are not characterized by NDVI and its derivatives are indirectly derived from all known information. The initial experimental results indicate that the overall mean error of phenological stage estimation is less than 2 weeks for both corn and soybean, which are better than the results produced using temporal composited products as reported by existing papers. The experimental results for corn and soybean phenological estimation also indicate that different denoising techniques may lead to different results on diverse land cover types.