About Crop Imaging
Related page: About Landsat & Crop Applications
Crop reflectance
The graph below shows how reflectance (total radiation emitted from a surface / total incoming radiation that strikes a surface) in different bands corresponds to agriculture. You can see that the more a pixel represents healthy green vegetation, the higher the NIR reflectance is and the lower the visible band (blue, green, and red) reflectance is.
dark green vegetation and soil = mixed pixels
The high ratio between the NIR and visible bands helps to differentiate crop condition within a field. Blue reflectance does not change as much between soil and vegetation as reflectance in the other bands, which is a reason why blue is not used as much (although is has been used). As the canopy fills, the visible bands become less effective than the NIR band at differentiating between crop condition and predicting yield patterns in a green field. Pixels that have blue, green, red, and NIR reflectance more similar to the dark green vegetation line in the above graphic represent better crop condition and will overall correlate to higher yield; pixels with reflectance more similar to the yellow line could represent a lower plant population or vegetation that has leaf damage due to drought, disease, or pest stress and will overall correlate to lower yield.
Types of imagery
Imagery from satellites and aircrafts are the main platforms for agricultural applications, though unmanned aerial vehicles (UAVs) continued to be applied (cannot be commercially used in USA but can in other countries such as Canada). Major differences between imagery include the pixel resolution (size of the pixels), bandwidths (spectral ranges of bands), radiometric resolution (range of values per pixel), and positional accuracy. In regards to satellites, in addition to the actual image characteristics, the revisit time (the amount of time it takes to image the same location) is very relevant. A necessity of imagery being applied to vegetation is that it include the NIR band (as will be described later on this page). Satellites including the NIR band have been launched in recent years that have resolution less than one meter, a high degree of positional accuracy, and have a revisit time of about a few days. A limitation for an individual farmer applying this data can be the cost. Recent high resolution imagery with less than one meter resolution needs to have a minimum area purchased (25 square km is on the low end) and will end up costing hundreds of dollars to purchase with the typical per acre price; however, the imagery can be very useful assessing crop patterns and the cost-effectiveness will increase if it can be applied to more than one field.
Landsat imagery is free. Currently, Landsat 8 and Landsat 7 ETM+ are operational and acquiring imagery; Landsat 7, however, has stripings of missing data due to a scan line corrector (SLC) instrument problem in May of 2003. Landsat 8 had successful launch in February of 2013. Landsat 8 data are based on a 12-bit dynamic range but are delivered as 16-bit images; Landsat 8 DNs have a maximum value of 65,535 (however, 55,000 corresponds to one hundred percent reflectance without factoring in a cosine of the solar zenith correction) as opposed to 255 for previous Landsat satellites. Landsat 8 has refined bandwidth which offers benefits for crop agriculture. Landsat imagery has 30 x 30 meter resolution in most bands (4.5 pixels per acre), and a 16-day revisit cycle. Landsat imagery dates back to July of 1982 (Landsat 4 TM) so historical assessments of a field can be made. The amount of imagery available is a function of whether or not a particular field is in a Landsat scene overlap zone and how much clear imagery there is of a particular field at the correct time of the growing season to sense crop condition and predict yield well enough.
Atmospheric correction limitations
Remote sensing values are normally referred to in terms of reflectance, as is shown in the above graph. Landsat 7, 5, and 4 surface reflectance data based on LEDAPS reflectance are now available, so image-based atmospheric correction is no longer necessary (not available for Landsat 8). Raw satellite values associated with pixels are digital number (DN) values that represent calibrated radiance (which is the amount of radiation emitted from a surface, not the percent reflected). Conversion from DNs to reflectance that takes into account certain atmospheric effects is known as atmospheric correction. Depending on the application, atmospheric correction may or may not be necessary.
A component of atmospheric correction accounts for the effect of "haze" or "path radiance". As solar radiation approaches Earth's surface it encounters atmospheric matter that reflects radiation back to the sensor before reaching the surface; this "scatter" can erroneously add to the values of bands recorded by the satellite sensor. If there is more of this matter in the atmosphere in different images or in different areas of the same image, then, with all else being equal, radiance values recorded by the sensor will be higher in those images or areas of an image. Methods to remove the scatter effects have been shown to improve data but are not totally reliable. Reducing the impact of scatter well enough in the visible bands is particularly challenging when the data represents crops. Green crops (and vegetation in general) have very low reflectance values in the visible bands (as shown in the graph above) and, as a result, any error in scatter amount can be too proportionally large. The correction for the scatter is typically done on an entire image basis (the same amount of scatter is applied to all areas of a scene) yet the actual scatter amount can vary throughout the extent of an image, particularly larger images such as Landsat scenes. The variable scatter amounts that can exist throughout the same Landsat image makes it important to compare areas that are close to each other to determine which fields have relatively higher values. Because NIR wavelengths are longer than visible wavelengths, less NIR is scattered under the same atmospheric conditions; this, coupled with the fact that green crops have a much higher NIR reflectance than visible band reflectance, makes scatter determination in the NIR band less problematic.
There is, however, a problem deriving accurate NIR reflectance for Landsat 4, 5, and 7 (and other satellite platforms with the same bandwidth) because NIR radiation is significantly absorbed by precipitable water in the atmosphere. Landsat 8 has refined NIR bandwidth that alleviates much of the absorption. The total atmospheric precipitable water amount and, hence, the amount of NIR radiation that may be absorbed depends largely on temperature, humidity, and the angle at which the radiation travels through (as is described by Wu et al. [2005; pdf]). The amount of atmospheric precipitable water can vary in locations within the same scene. A description of the effects of precipitable water on NIR wavelengths can be found in Wu et al. (2005; pdf), Guzzi and Rizzi (1984; pdf), and Eldridge (1967; pdf). This problem is partially resolved with Landsat 8 due to refined NIR bandwidth.
Vegetation spectral indices and correlation to crop yield
Correlation to corn yield Correlation to soybean yield
Vegetation spectral indices (VSIs) take advantage of the ratios and amounts of reflectance of the different bands shown in the graph. The growth stage that VSIs can effectively predict yield patterns can vary by crop and bands that can best be used to predict yield patterns can vary by growth stage. A common VSI is the Normalized Difference Vegetation Index (NDVI; [NIR – red] / [NIR + red]). NDVI tends to work better earlier in the season, before canopies have closed because visible bands start to saturate (values become the same or similar throughout the field even though crop condition is different) as the canopy closes. Wu et al. (2007) reported that MSAVI is better than NDVI, SAVI, TSAVI, and PVI for detecting corn and potato leaf area index (LAI). Hollinger (2008) reported that in northwest Ohio in mid-July (late vegetation stage) the average correlation (r) (n = eight fields) between the average of clean corn yield monitor data within the extent of Landsat pixels and the corresponding pixel vegetation spectral index values for NDVI, SAVI, and TSAVI were all between 0.75 and 0.80 (p < 0.01 in all cases). Other high correlations are reported in Hollinger (2011) and are shown in the Landsat Crop Yield Prediction folder in this website. High correlations have been shown between NDVI and GNDVI and corn and soybean yield (Adamchuk et al. 2003); GNDVI is NDVI with green reflectance substituted for red reflectance. Articles about remote sensing correlation to yield can be accessed in the "Remote sensing of crops and yield prediction" link of the Articles page which is in the Articles folder.
Although VSIs have been developed to account for the effects of the soil background, imagery should still only be used when soil visibility has been significantly reduced in order to improve the reliability of crop condition assessment and yield pattern predictions. The determination of whether and how imagery can predict yield patterns well enough needs to be determined on an image-by-image, field-by-field basis, and crop-by-crop basis. For example, Landsat imagery of a corn field from the time a field has tasseled to the end of the season should not be used because the tassels obscure NIR reflectance too much; on the other hand, soybean phenology (including blooms) does not obscure NIR reflectance much, so imagery can be used throughout much of the reproductive stages to effectively predict yield patterns (should be used from about R2 through R6; soybeans have a longer time window for effective imagery than corn).
References
Adamchuk, V.I., Perk, R.L., and J.S. Schepers. 2003. Applications of remote sensing in site-specific management. Precision Agriculture; University of Nebraska Cooperative Extension EC 03-702.
Eldridge, R.G. 1967. Water vapor absorption of visible and near infrared radiation. Applied Optics, vol. 6, no. 4, pp. 709-714.
Guzzi, R., and R. Rizzi. 1984. Water vapor absorption in the visible and near infrared: results of field measurements. Applied Optics, Vol. 23, no. 11, pp. 1853-1861.
Hollinger, D. 2009. A GIS-based method to predict county corn yield based on retrieved Landsat reflectance variability in western Ohio. Papers of the Applied Geography Conference 32: pp. 281-290.
Hollinger, D. 2008. Spatial correlation between Landsat 5 TM-derived vegetation spectral indices and corn yield in northwest Ohio, 2007. Papers of the Applied Geography Conferences 31: pp. 85-94.
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: pp. 33–42.
Wu, J., D. Wang, and M.E. Bauer. 2005. Image-based atmospheric correction of Quickbird imagery of Minnesota cropland. Remote Sensing of Environment 99:315-325.