Precision Agriculture: Remote Sensing

If maps of the spatial distribution of soil productivity potential (maps of expected yield) and maps of the spatial distribution of plant nutrients available from the soil are developed for a field, fertilizers and organic wastes can be applied in amounts per acre that are directly proportional to the soil's expected yield and adjusted for the soil's fertility at any location in the field. Such a procedure would optimize the economic potential of a field, yet minimize the leaching of nutrients.

The above protocol depends on having a good map of the spatial variation of expected yields for crop fields. Maps of past crop yields for a field could be used for this purpose. However, multiple years of spatial yield data would be needed to overcome variations caused by year to year differences in weather, especially rainfall, and there remains multiple factors which result in lack of year to year correlation, L7_YieldComposite.GIF and L6Yieldcomposite.GIF 2. An alternative to mapping of actual crop yields would be to use remote sensing to determine spatial distribution of plant status (health or efficiency) and the corollary expected yields. A major advantage of this approach is that remote sensing can provide a current assessment of the overall plant health of the crop rather than relying on past history of yields.

Several different approaches exist for using remote sensing data for this purpose. Most of the commonly recognized techniques depend of measuring the "greenness" of the field. Typically this is involves some relationship comparing the reflectance of a visible band (such as red light) to the reflectance of a near-IR band. Since green vegetation has a very sharp change in reflectivity across this range and other material do not, virtually any technique will in fact detect it. The approach suffers from several defects, for example it is a relative technique and can be significantly affected by soil conditions. We have pursued a different path in this research. We have examined the thermodynamic efficiency of the crop. The core of our approach depends on energy in the thermal-IR. Our experiment was to study the energy budget of the crops and demonstrate a relationship between multi-temporal thermal imagery and crop yield.

The basis of the thermal remote sensing is simple physics. The total radiation received at the ground is partitioned into reflected energy and absorbed energy. The bulk of the absorbed energy is converted into either sensible heat or latent heat, i.e. changing the phase of H₂O from liquid to gas. Plants are efficient at minimizing the production of sensible heat and maximizing the conversion of incoming energy to latent heat. Thus measurement of how a plant surface is warming during the day is a good indication of the rate of transpiration.

This has been formalized in the work of Luvall and Holbo (1989). They demonstrated here and in subsequent work that if measurements of surface temperature are taken repeatedly during the day the change in temperature with time is an excellent measure of the availability of water to plants. Practical implementing this knowledge requires several precise conditions must be satisfied. To obtain the measurements of temperature using remote sensing it is necessary to remove atmospheric effects. The instrument itself must be calibrated. And finally it must be practical to reference multiple data sets to a common geometric base. Very few airborne systems provide the necessary information. These have been accomplished for the Atlas sensor flown by NASA's Stennis Space Center. (Rickman, D., Ochoa, M., Holladay, K., Huh, O.; 1989; Rickman, 1986).

Use of the Atlas has the added advantage that it allows us to compare results obtained from thermal data with results obtained from more common techniques, which depend on visible and Near-IR bands, CompareBands.gif and Thermal2.gif.

Data were acquired by NASA's Atlas scanner, (need to make reference maps), on June 26, 1998 and August 4, 1998. Four targets were flown, two in north central Alabama and two in southwestern Georgia. Acquisition was designed to get 2.5 meter spatial resolution twice within approximately 30 minutes.

We have significant results in three topics, one of which is completely outside our expectations.

Figures L6_Yield&F1_61.GIF, L7Yield&F1_71.GIF and L5_yield&Atlas.GIF clearly show that an extremely strong correlation can exist between yield and appropriate remote sensing data. This is true for both corn and cotton. These fields are dry land fields.

However, such a strong correlation is not a universal situation. f3_72_2613b.gif is a later flight over the same field as L7Yield&F1_71.GIF. It was acquired after the crop had started to die back, thus the lack of correlation is not surprising. But L2_Yield&F2_21.GIF is much more problematic. It is from an irrigated soybean field; the total variation in yield is relatively small, between approximately 40 - 52 bushels/acre. The crop was green and near maximum canopy, therefore the lack of correlation is surprising. We believe there are two factors that account for this. First, these data clearly show an unexpected phenomena, wind induced variation in spectral reflectance. This will be discussed in a later section. Second, the mechanism that causes the strong correlation between yield and the remote sensing data in the first three cases is missing. There is essentially no variation in canopy density in this soybean field. Correlation Matrix

Number of points: 37840

Correlation yield to ratio of bands 6/4=0.7476

Our analysis shows the strong correlation in the other imagery is driven by the relative abundance of crop and soil detectable in each pixel. All bands for these fields show a strong correlation with yield, limited more by the sensor sensitivity and noise than by anything else. Given the strong spectral contrast between these soils and live vegetation it is not surprising the ratio of soil to vegetation will strongly affect the imagery. As a simple check, ratios were done between the bands covering visible yellow (Band 4) and a Near-IR band (Band 6). Any ratio between these bands is well known to be very sensitive to the relative abundance of vegetation in the pixel. Such ratios, for example see f3_51_Composite.GIF, as expected given the apparent correlation between the individual bands and the yield, also show very high correlation with the yield. For the soybean field, such a ratio shows that virtually no soil is visible, see Figure L2_ratio64.gif.

We draw three conclusions from this. First, given that seeding density is unchanged across the fields, the map of crop yield in the dry land sites is tightly correlated with plant health/green biomass. Second, the yield within a field is spatially highly variable. Third, for such cases as these virtually any remote sensing approach will yield essentially the same results. Thus, if the remote sensing is of value in monitoring crops at this growth stage, considerations other than band sensitivity to some particular phenomena become important. An analysis of these criteria are beyond this report's scope.

We do illustrate one such consideration, cloud cover. The data for line 5, Crisp County, Georgia, were affected by cloud cover. f3_51_Composite.GIF shows that even where significant shadows occur, appropriate bands and data manipulation can either eliminate or greatly reduce the problem. Given the weather during the growing season in the southeast U.S., this may be a consideration in selection of remote sensing platforms.

Our original hypothesis said there should be a difference between the rate of temperature change in healthy versus less healthy plants during solar loading. We were unable to find such an effect in the corn fields in Alabama. The Georgia soybean data were disturbed by a wind related phenomena, discussed in the next section.

We can not definitively explain our inability to detect a temperature effect related to a change in time. The time between flights was roughly 8 minutes and 30 seconds. This probably was not sufficient time, given the instrument sensitivity. There is also almost certainly a wind effect in these fields, as well as the soybeans, but the relatively rougher structure prevents its detection.

(Continue here.)

L7_f1_72.gif L7_F1_Ch13_DeltaFiltered.gif 14:04:39 14:13:08

Because there is overlap between some of the flight lines, several of the fields were actually imaged 4 times. Wind_Thermal.gif shows one such field and the GMT time stamps. The data are from the Thermal IR (Band 13 of the sensor). The first feature most people note is the streaking from upper left to lower right. The streaking changes between images and there are zones which on average are brighter or darker. The signal is essentially pure variation in temperature of the crop, soybeans.

Wind_NearIR.gif shows the same views in the Near-IR (Band 6 of the sensor). All of these images were processed to use the full dynamic range within the single field. Clearly the Near-IR reflectivity is varying rapidly within 3-4 minutes. What is especially significant is that this Near-IR reflectance variation is highly correlated with the emitted, thermal signal. This is illustrated by wind_skats.gif which shows scattergrams for Near-IR and visible (Band 3) versus Thermal-IR for each data set. Even the visible bands in at least on case are correlated with the thermal and the spatial distribution of the thermal signal is changing radically within minutes! Spatially, there are two distinct scales. One generates linear streaks with approximately a 10-30 meter wavelength perpendicular to the streaks. The other (L6_Channel13Composite.GIF) shows a spacing closer to a kilometer and individually goes not form such pronounced lineations, although there is a overall, preferential orientation.

We have considered several possible explanations for these observations. For example artifacts of the scanner, artifacts of the crop rows, artifacts of look angle or some combination of these. Multiple points prove these are not factors: 1. The same patterns are equally visible in many of the fields at the farm. 2. The track of the scanner's sweep is oblique to the average trend of the streaks. 3. Various fields have differing row orientations. 4. The crops involved include soybeans, cotton and peanuts. 5. The 4 images are taken from just two views, as the aircraft flew nearly identical tracks on the repeat flights. 6. The patterns cross field boundaries.

We think this is a response to wind. Wind is the only phenomena which can change this rapidly and on such a small spatial scale and affect a variety of crops several kilometers apart. The Albany, GA airport, which is less than 16 kilometers north of the fields provided the following data.

10:47 Z calm, less than 3 knots

11:47 Z calm, less than 3 knots

12:47 Z calm, less than 3 knots

13:50 Z 060 degrees @ 05 knots

14:50 Z 060 degrees @ 09 knots

The direction is in agreement with the observed streaking in the thermal images. Ground observation by a meteorologist supporting the data acquisition at the experimental farm at time of over flight reports clear skies, no clouds.

The effect is not limited to the Georgia data. It is also present in the August data from Alabama. L6_Channel13Composite.GIF clearly shows the same phenomena.

As yet we do not have a detailed explanation for the mechanisms involved. Probable factors include the incidence angle of irradiance on the leaves, the elastic behavior of plant stems, and the coupling of leaf with air temperatures. The phenomena is apparent in three crops, soybeans, cotton and peanuts, which have relatively similar structure. It is also present in forest canopy

The significance of the observations is that this is a previously unrecognized source of significant variation in visible and Near-IR imagery.

1. Comparison of yield maps for a single field over 3 crops. Note lack of detailed correlation.

2 a&b. Spectral Curves for Atlas scanner.

3. Panel of thermal images for GA site.

4. Panel of Near-IR images for GA site.

5. Panel of scattergrams for GA site.

6. Panel of yield versus June and August imagery for Line 7.

7. Panel of yield versus June and August imagery for Line 5.

At the Crisp County site, both the growth rate of cotton and the hand picked yield of cotton lint varied with location in the field (Figure 1 and Table 1). Five measurement areas were selected to cover the range in soil type conditions in the field. Dry weights of cotton were lowest on the sandier Norfolk in the NE part of the field, and the dry weights were highest on the SE Orangeburg, an area of soil with relatively high soil organic matter. Lint yields varied from to kg lint per hectare. The differences in growth and yield were also verified with the remote sensing image taken August and the cotton lint yield map obtained at harvest with the microtrac yield monitor on the cotton picker, also shown in the report.

These differences in growth ratesbetween soil types could not be explained on the basis of availability of nutrients, since analysis of soil samples from across the field indicated adequate plant available phosphorus, potassium, and other nutrients for these areas of the field (data not shown). Also, the concentration of nitrogen, phosphorus, potassium, and other nutrients of the youngest fully expanded leaves indicated that these nutrients were in the sufficient range with the exception of potassium for one sampling date at the Norfolk location (Figures 2, 3, and 4). Also, no toxicities from aluminum or manganese were apparent, judging from the normal manganese concentrations in leaves (from 100 to 300 ppm manganese) and soil pH above 5.5 at the measurement locations.

The differences in growth and yield appeared to be due to differences in the degree of water stress in the various parts of the field. Cotton emerged around May xxx, and differences in growth appeared early in the season, with the greatest growth rates in areas of lower elevation that received water as runoff from higher elevation areas. Lowest yields were for soils with sandy A and E horizons more than 100 cm thick and eroded soils on convex slopes. Highest yields were on soils in shallow closed depressions with dark colored surface horizons. The water stress would be indicated by the low rainfall during the growing season (Figure 5). Differences in the ability of soils to store plant available water (dependent on soil organic matter content and clay content) are likely the cause of the yield differences.

Responsible Official: Dr. Steven J. Goodman (steven.goodman@nasa.gov)
Technical Contact: Dr. Doug Rickman (doug.rickman@msfc.nasa.gov)
Page Curator: Diane Samuelson (diane.samuelson@msfc.nasa.gov)