Soil Bias and Effect on Correlation to Yield
Imagery applied to crops need to be from a time when soil influence has diminished enough. The graphics below (Figure 24 from Hollinger [2011]) are from a cornfield where higher yield is in darker soil. The 6/20/07 date is about V11 and the 7/13/07 date is about V19 (based on growing degree days). NIR radiation can transmit through leaves and reflect soil, so soil needs to be substantially covered. On the 6/11/07 and 6/20/07 image dates, there is more vegetation in the darker soil but there is not more NIR reflectance because there is too much background of darker soil averaged into pixel values. Soil is not sufficiently covered until the latest image on 7/13/07 (V19) when higher NIR reflectance is correctly located in darker soil where crop condition and yield are higher. To apply imagery to crops early on, NIR and visible bands (not shown) need to negatively correlate which only happens until the later image. Ultimately, knowing the growth stage at time of imaging is not as important as NIR and visible bands negatively correlating.
In the Landsat images below, lighter shades correspond to higher NIR reflectance
Imagery (6-inch; OGRIP, 2012) of soil and residue Landsat NIR (band 4) on 6/11/07
Landsat NIR (band 4) on 6/20/07 Landsat NIR (band 4) on 7/13/07
Certain indices tend to correlate to darker soil when calculated for soil imagery whereby higher index values will correspond to areas of darker soil. This biases correlations in fields where higher yields are in darker soil areas when the soil background is visible and averaged into pixel values, as it is on the 6/11/07 and 6/20/07 image dates. The 6/20/07 date is at about V11 and the 7/13/07 date is about V19 (based on growing degree days). The graphic below (Figure 19 from Hollinger [2011]) corresponds to a cornfield (different than the field shown above) and shows 19 Landsat-based indices and NIR (band 4); all are designed to correlated positively with yield. The graphs show that the indices that correlate better to corn yield earlier in vegetative stages do so, in large part, merely because they have higher values in darker soil when calculated for soil. (Indices such as NDVI have higher bias and higher correlation to yield when more soil is visible while soil adjusted vegetation indices have lower bias and lower correlation to yield when more soil is visible.)
Reference
Hollinger, D. 2011. Crop Condition and yield prediction at the field scale with geospatial and artificial neural network applications. Dissertation. Kent State University.
OGRIP. 2012. Ohio Geographically Referenced Information Program. Cited at: http://ogrip.oit.ohio.gov/. Last updated: 2014.