Landsat Correlation to Corn Yield
(Graphics and statistics shown below; including comparison between MSAVI, NNIR, NDVI, & GNDVI)
Different vegetation spectral indices correlate highly to spatial patterns of corn yield (as shown below) and leaf area index (LAI). Wu et al. (2007; pdf) showed (based on Quickbird imagery; has a much finer resolution than Landsat) that MSAVI (Qi et al., 1994) is better than other indices, including NDVI, for remote sensing of corn LAI. Hollinger (2011) showed MSAVI correlates higher to corn yield spatial patterns than most indices, but other indices correlate very similarly to MSAVI. Below, Landsat-based MSAVI from V12 to V19 is correlated (linear regression) with corn yield by averaging clean yield points within the pixel extent (based on 4-meter spacing; approximately 56 yield points per pixel); a comparison of correlations with other indices based on reflectance and digital numbers (DNs) is shown after the correlation graphics. Landsat imagery should not be used for corn prior to V12 because soil has too strong of an influence or at VT and later due to the effect of tassels.
Images transition every 4 seconds (or click arrows or numbers); statistics appear below graphic
(Landsat MSAVI from V12 to V19 is on left; corresponding clean corn yield data averaged to each pixel is on right)
Correlation (R²) with corn yield for above pixel groups for MSAVI and other indices based on reflectance and DNs; correlation level is largely a function of the standard deviation of the index value for a field (based on data from Hollinger ):
|Field||Stan. Dev.||MSAVI||NNIR||NDVI||GNDVI||DN NNIR||DN NDVI||DN GNDVI|
* MSAVI has a much higher correlation than others for this field because it weights NIR correlation within the index value relatively more than the other indices; this field has a very high correlation between NIR and corn yield.
NNIR = NIR / (NIR + red + green); NDVI = (NIR – red) / (NIR + red); GNDVI = (NIR – green) / (NIR + green)
Stan. Dev. is MSAVI standard deviation. Correlation level is largely a function of index standard deviation as shown here: R² = 0.9795 between MSAVI and standard deviation (polynomial 2nd order); R² = 0.7381 for linear correlation.
Pixels are only used for correlations if they do not average in areas from outside the field (this eliminates many near boundary) or major non-crop surfaces within the field, and if they correspond to valid clean yield points (for example, there can be a void of points adjacent to headland areas while points near obstacles such as electrical installations are typically erroneous). For both maps, darker green is higher value. Pixels are 30 x 30 meters.
Knowing that Landsat can provide high yield correlations as shown above, valid pixels from throughout a field (not just those that correspond to valid clean yield points, which is limited throughout a field but necessary for the above correlations) can be applied as the basis for a yield prediction map. The map only includes pixels that represent surface from within the boundary of a field (and excludes non-crop areas from within the boundary if necessary) and has the data extended to the boundary. The map can be calibrated to yield amounts based on an equation or any specified yield range and/or average value; the data is then essentially a generalized yield map (see below). An example of how Landsat can be used to produce a field yield prediction map of continuous data can be viewed though the link that follows:
Steps that show how Landsat can be used to produce a yield prediction map of continuous data or zones are included below (same process for any crop; soybean field shown below) and can be seen in different locations of the website. The map can be produced to a field extent of different shapes and sizes and can be calibrated with yield values based on an equation or a specified yield range and/or average value; the data is then essentially a generalized yield map.
The progression from raw Landsat imagery to maps to a field extent (or any other extent) is the same for any crop. From left to right the steps are: 1) produce Landsat with pixels that represent correct value to predict yield well enough; 2) use only pixels that represent the crop, not pixels that average in surface outside the field extent or major non-crop surface within the field perimeter; 3) data is interpolated or resampled to a finer resolution for a more coherent map (below, the pixels have been modified from the 30 x 30 meter native resolution to a one-meter resolution); 4) zones can be developed (the appropriate classification method, amount, and minimum size of zone need to be determined).
Calibration of Landsat Maps to Yield
Landsat yield prediction maps can be calibrated to yield based on solely a field average or a specified yield range and average can be applied. In either case, the calibrated map will keep the same proportions to the Landsat value map (in other words, the map symbology will look the same before and after calibrated to yield).
If just a field average is known, equations from Hollinger (2011) can be applied that predict yield for the entire field. Reflectance-based index values are necessary to apply the equations. The equations predicts normalized (to the mean) yield amounts based on normalized (to the mean) MSAVI or other index values. Once normalized yield values are calculated they are multiplied by a field average to acquire a yield prediction map. To produce the equation for MSAVI and other indices, pixel groups from many fields over two seasons from image dates from V12 to V19 were normalized to the mean and combined into one plot where they were correlated with corresponding clean yield monitor data (R² = 0.60 for MSAVI and yield; n = 1,086 pixels). The yield variability prediction model is based on fairly typical seasons; severe weather patterns in R-stages can cause yield variability to change a significant amount after the late V-stages. For example drought during earlier R-stages can cause a higher ratio of lower ground to higher ground yield.
Maps can also be calibrated based on a specified yield range and field average. An advantage of this is that DN values can be used (atmospheric correction is not necessary) which makes the process less time-consuming.
Reliable corn yield quantity maps at the field-scale cannot be produced with Landsat imagery because reliable crop assessment at the field-scale is only posssible during the later vegetative stages which is too far from harvest and excludes growth stages associated with greatest water needs (spatial patterns can still be predicted well from these growth stages though).
For the slideshow correlations, the NIR precipitable correction factor from Wu et al. (2005) was applied to the solar transmittance along the solar zenith angle and the path of the satellite view angle in the atmospheric correction process. The factor increased NIR reflectance significantly (increased average reflectance for the V-stage image date nearest tassel (late V-stage) from .3675 to .4964 [n=468 pixels]; applying the factor to only transmittance would retrieve a reflectance somewhere between the two. Although the factor increased reflectance, correlations between MSAVI and corn yield using either reflectance value were nearly identical; the largest absolute difference with and without the factor for any correlation was 0.0084 while the average absolute difference was 0.0034. Landsat 8 has a launch date in February, 2013; refined NIR bandwidth of Landsat 8 helps resolves the problems that Landsat 5 and 7 have with atmospheric water absorption of NIR wavelengths.
Hollinger, D. 2011. Crop Condition and yield prediction at the field scale with geospatial and artificial neural network applications. Dissertation. Kent State University.
Qi, J., Chehbouni, A., Huete, A. R., Kerr, Y.H., and S. Sorooshian. 1994. A Modified Soil Adjusted Vegetation Index. Remote Sensing of Environment 48: 119-126.
Wu, J., Wang, D, and M.E. Bauer. 2007. Assessing broadband vegetation indices and QuickBird data in estimating leaf area index of corn and potato canopies. Field Crops Research 102: 33–42.
Wu, J., Wang, D., and M.E. Bauer. 2005. Image-based atmospheric correction of Quickbird imagery of Minnesota cropland. Remote Sensing of Environment 99: pp.315-325.