Report of Remoting sensing class project

Class Project

Mapping of Crop Residues Using Hyperspectral Data:

· Techniques and indexes for quantification,

· Data sources and

· Unsupervised classification for tillage systems

Table of contents / INDEX

	Topic	Page
1.	Problem / application	3
2.	Working hypothesis	3
3.	Project outcomes	4
4.	Literature review	4
5.	Data sources	8
6.	Methods	10
7.	Results	16
8	Issues and learning	20
9	Conclusions and future works	21
10.	Annex 1. Corrected bands and columns	22
11	Annex 2. Copy of in-running matlab code for de-striping	23
12	References	25

2

1. PROBLEM / APPLICATION

Agriculture is a widespread, basic activity around the world, which main purpose is to harvest food, fiber or/and energy. After every growing season residues are left in fields. It is important to quantify the amount and cover of agricultural residues for enhancing the understanding in global biogeochemical cycles, and for applications such as their role for preventing soil erosion and their contribution in carbon sequestration. However, it is not completely understood yet how to estimate crop residues cover, their discrimination under tillage or no tillage cropping systems, and its seasonal variability as well as their temporal changes. This class project proposes to explore the estimation and mapping of crop residues by remote sensing techniques using hyperspectral image data.

2. WORKING HYPOTHESIS

Crop residues cover and amount can be accurately estimated by remote sensing techniques. A wide range of crop species and their residues can be studied in the near future and they might be even differentiated by spectral classification. Future work might include description of temporal patterns upon analyzing hyperspectral data (EO-1 Hyperion) in complement with multispectral data (Landsat 7 ETM+ and EO-1 ALI).

3

3. PROJECT OUTCOMES

This class project will generate an estimation of crop residues cover in agricultural fields in Central Indiana in Tipton County. In addition, the amount of crop residues will be approximately calculated based upon yield/residues ratio assumptions. Also, unsupervised classifications for different tillage management (two classes: tilled areas and no-tilled areas) in agricultural fields in Tipton County. Finally, by this study we expect to integrate/use three different data sources (Landsat 7 ETM+, EO-1 ALI and EO-1 Hyperion) and to calculate Cellulose Absorption Index on hyperspectral data.

4. LITERATURE REVIEW

Crop residues are any portion of crop plants that is left in the field after harvest. Crop residues cover is a relevant topic to be studied because of three main reasons: they are widespread in the landscape of agriculture in the Midwest, they represent one of the most important organic inputs for soil carbon sequestration estimating input, and also they relate to soil conservation and reduction of soil erosion (Lal, 2002 & 2004). Remote sensing techniques are also a potential for these methods for monitoring compliance with conservation measures or perhaps in the future carbon credits.

Agricultural management practices such as tillage and crop rotation (West, 2002) are driving factors for crop residues accumulation in field. In any case, authors such as Nagler et al (2003) have pointed

4

out that remote sensing methods would be useful for providing wider area coverage in a regional landscape studies as well as estimation of spatial structure/variability of residues cover in agricultural settings. Spectral variability is another important outcome from remote sensing methods, which in general is overlooked in manual methods for crop residues determination.

One of the first studies assessing crop residues estimation by a remote sensing method was by Biard and Baret (1997). They appealed to multiband reflectance under an algorithm called Crop Residue Index Multiband (CRIM) which is based on any set of wave bands and consists on a linear mixing model of soil-residue complex and on soil and residue lines. However the same authors found dependency of the near infrared and middle infrared domains according to the aging state of the residues, which may be due to decomposition of key macromolecular compounds such as lignin and cellulose.

At about the same time as Biard and Baret (1997), Su et al (1997) also proposed another way to assess crop residues by SAIL model. It consisted in Scattering by Arbitrarily Inclined Leaves, which simply simulated the residue reflectance in wheat in near infrared band, and then it was possible to study the agreement between field measured reflectance and the simulated reflectance. They concluded that SAIL model was a promising technique for crop residues determination at that point in time.

Daughtry et al (1997) studies recorded potential problems that must be addressed to implement the fluorescence technique in the field,

5

which are: adequate excitation energy must be supplied to induce fluorescence, and the fluorescence signal is small relative to normal, ambient sunlight. The technique must be developed to either shield the system from sunlight or extract the fluorescence.

Nagler et al (2003) studied the use of Cellulose Absorption Index (CAI) as a way to quantify the plant litter cover after harvest even at low cover percentage (around 10 %). Their research conducted analysis from 0 % cover (bare soil) up to 100 % plant litter cover and using reflectance spectra (0.4 to 2.5 μm) for four crops: corn, soybean, rice and wheat; plus two tree species.

Following Nagler et al (2003) work, Daughtry et al (2004) established that the spectra of dry crop residues displayed a broad band absorption feature near 2100 nm due to absorption by cellulose/lignin compounds. Therefore, these authors proposed the combination of Cellulose Absorption Index (CAI) and Normalized Difference Vegetation Index (NDVI) base upon shortwave infrared reflectance in order to estimate crop residues cover. Also, it seems to be important to make a difference between dry residues and wet crop residues under this approach. However, once again these authors concurred in the usefulness of remote sensing methods in regional landscape surveys for crop residues.

Continuing with their work on CAI, Daughtry et al (2006) used multi/hyperspectral data (Landsat and Hyperion data) to distinguish two classes in corn and soybean residues. Landsat data did not yield good fit between field measurements and remote sensing analysis. Conversely,

6

Hyperion yield linear relationship for CAI. In addition, classification accuracy resulted in 80 % when working just in two classes of tillage (conventional: reduced + intensive, and conservational tillages). Consequently, it seems that an advanced multispectral or hyperspectral data is needed when assessing crop residues in a regional landscape scenario.

In this sense, Bannari et al (2006) also concurred with Daughtry et al (2006) when pointing out that a hyperspectral data (e.g. Probe-1) or an advanced multispectral data (e.g. high spatial resolution IKONOS data) is needed for precise evaluation of surface crop residues cover because of better spectral band characteristics, specially when looking at lignin/cellulose absorption features. In other words, Probe-1 hyperspectral data outperformed the IKONOS data because of the characteristics of the spectral bands in the hyperspectral sensor.

Lately, South et al (2004) compiled and analyzed a very interesting comparison of classification methods based upon spectral reflectance signatures for mapping senescent crop residues in Eastern Cornbelt soils. They concluded that out of five different methods (including parametric an non parametric ones) two spectral angle methods (spectral angle mapping and cosine of the angle) were the ones with higher performance, specifically the cosine of the angle concept algorithm had the highest accuracy (97.2 %) and kappa value (0.959). In contrast, minimum distance, Mahalanobis classifier and maximum likelihood had user’s accuracy below 84 % and producer’s accuracy around those values. Bannari et al (2006) also established a validation for their supervised classification with a divergence D of 0.86.

7

Utilizing aircraft and satellite image data (IKONOS hyperspectral data), Martin (2002) developed both unsupervised (using ISODATA algorithms) and supervised classifications (using ECHO: extraction and classification of homogenous objects and maximum likelihood algorithms) for different tillage and crop rotation treatments in a location in the Midwest. This author found that to compare tillage systems as classes is the best method to estimate residue cover. Four different treatments were investigated and those tillage systems were indicative of the amount of residue present within a given date but also that the total amount of residue differs during seasons of the year due to decomposition.

5. DATA SOURCES

Location of the study is between Windfall City and Tipton City in Tipton County IN, respectively. Coordinates are between 40º21’44” to 40º10’32” N and 85º57’25” to 86º1’11” W. This narrow area was defined by the EO-1 Hyperion stripe (about 7.7 km wide).

Images data are Landsat 7 ETM+ (a single scene corresponding to path 21 and row 32 radiometric and geometric corrected), EO-1 Advanced Land Imager ALI (a single stripe radiometric correcte), and EO-1 Hyperion (radiometric corrected). Landsat 7 and EO-1 correspond to the AM constellation as shown in Figure 1. They were acquired in the same date (April 12th, 2003) under a zero cloud cover. All three image were subset to extract about the same narrow area (sensor have different cover; ALI does not cover all the Hyperion view), corresponding

8

to a specific geographic location in the center of this county in order to study agricultural fields. Figure 2 shows the three subsets. The data sets are already downloaded in the Oxisol hard drive computer at LARS.

A B

Figure 1. AM Constellation including Landsat 7 and EO-1 spaceborne platforms (A) and swath with for three different sensors: Landsat 7 ETM, ALI and Hyperion (B).

A B C

Figure 2. Three original image data subsets showing field areas between two cities: Tipton City in the southwest and Windfall City in the northeast. Landsat 7 ETM (A), EO-1 ALI (B) and EO-1 Hyperion (C). Source: www.indianaview.org

9

The images in Figure 2 where prepared using RGB representation of Landsat 7 bands: 4 (NIR 0.76 – 0.90 μm), 3 (red 0.63 – 0.69 μm) and 2 (green 0.52 – 0.60 μm). Consequently, in case of ALI data bands 5 and

6 were averaged, and for Hyperion data average were prepared as B42-

45, B28 – B33 and B18 – B25, respectively.

In addition to image data, we accessed data from the Annual Crop Residue Management Survey 2002 (www.conservationinformation.org). It contains field data about crop residues and tillage systems. Therefore, it allowed us to do a comparison between image analysis and field data.

6. METHODS FOR DATA PROCESSING

Pre-processing include two main steps: de-striping and conversion/corrections.

Problems with calibrations of the Hyperion hyperspectral sensors results in stripes in the band images (which translates into “bad columns or bad detectors”). They may show up as dark stripes or clear stripes in vertical position.

Since in order to calculate the CAI index, one may need 9 SWIR bands: (183, 184, 185, 195, 196, 197, 204, 205, 206). After opening the bands and performing Gaussian and Linear 2% enhancement, the bands were examining and found that bands 183, 184, 185 needed stripe correction. Figure 3 A show one example (B196) of a band with any problematic columns, so de-striping was not run on those other six

10

bands. Problematic columns were visually identified in the first three bands. Annex 1 shows a list of bands and columns that were corrected.

De-striping was performed by iterated running a MATLAB® code (cor_str written at U. Texas) on the data. Annex 2 shows a copy of the in-running code as example. The code uses 32 adjacent pixels from the previous columns in order to fix the striped columns and bring it to a normal distribution. Figure 3 shows how the de-striping code yield positive results when removing the stripes from bands 184 and 185. However, the striped B185 was still noisy and during the index calculation, it was excluded.

A B C D E

Figure 3. Hyperion Hyperspectral Bands: 196 (A), 184 before (B) and after (C), 185 before (D) and after (E).

In order to analysis the effect of de-striping in noise reduction, we performed minimum noise fraction (MNF) transform. Generally MNF is used to reduce the dimensionality of hyperspectral data; however a forward transformation can remove the noise from data and in that way determine which bands contain the coherent images by looking at the

11

actual images or to the eigenvalues. After identifying the noisy bands an inverse MNF can be run on a spectral subset including just the good bands. Figures 4 and 5 show the eigenvalues/eigenvector/linear transformation before and after the running de-striping code. It is evident that at least one of the linear combinations has gotten better after the de-striping pre-processing of the three bands (B183-185) as shown by a decrease in the seventh eigenvalue. Here, the eigenvalues or eigenvectors are the result of linear combinations. In fact Figures 4 and 5 are also showing that one of the linear combinations (the first) is containing the most statistic information. In other words, one “new band” by linear combination through MNF transform has represented a large portion of the information containing originally in the nine bands. It is common that hyperspectral bands are correlated each other introducing redundancy and MNF is in fact a way to deal with this difficulty by reducing dimensionality.

	20
					Before De-striping
Eigenvalue	15				After De-striping
	10
	5
	0
	1	2	3	4	5	6	7	8	9

Eigenvalue Number

Figure 4. Eigenvalues for nine different linear combinations.

Figure 5. Image representing three of the linear combinations before and after the de-striping.

12

The second main step in pre-processing was the conversion and correction needed for the index calculation. Most of the crop residues

models and indexes are built on reflectance units. Therefore, digital number (DN) needed to be converted into radiance and then into

2

reflectance. The conversion to radiance [ W / (m sr µ) ] for Hyperion data was achieved by dividing the VNIR 50 bands (B8-57) by the scalar 40 and the SWIR bands (B77-224) by 80. Then, the conversion from

radiance into reflectance in order to normalize for the incoming radiation was done in FLAASH model.

A B C D

Figure 6. Hyperion RGB images for average bands (Red: B183-184, Green: B195-197 and Blue: B204-206) in digital number (A), radiance (B) (units= W·sr-1·m-2), reflectance from FLAASH without any aerosol/water corrections (C) and reflectance from FLAASH with both corrections aerosol and water retrieval by absorption band 1.1 μm (D).

Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) is an atmospheric correction modeling tool in ENVI 4.3 for retrieving spectral reflectance from hyperspectral radiance images. FLAASH incorporates the MODTRAN4 radiation transfer model to compensate for atmospheric effects. FLAASH is capable of doing aerosol correction as well as H2O vapor corrections.

We run both aerosol correction and water vapor correction. For H2O vapor corrections, the model has the choice of three different water absorption bands/features: 1.1, 0.9 and 0.8 μm. . We evaluated the first and the third features in Hyperion data resulting in better scenes and lower water retrieval in the feature 1.1 μm. The water retrieval in water absorption bands 770 - 870 nm resulted in 0.227 cm, while in the band 1050 - 1210 nm water retrieval was 0.120 cm. Both results can be compared to FLAASH output without application of water retrieval detected: 1.302 cm H2O vapor in atmosphere of scene.

In other words, FLAASH detected a total of 1.302 cm of water average in the atmosphere of the scene, and then the two absorption bands were able to drastically reduce the water vapor; however, the 1.1 μm features was more efficient to remove water vapor, so we keep using it for our correspondent analysis. The image was taken in a no cloudy day; however, FLAASH made it slightly better; besides it was an interesting exercise to do. The best way to correct from radiance to reflectance would be to set spot on the ground with different material (white paper, water, etc...) right in the moment when the airborne or spaceborne sensor is doing the readings, and then do the correction based on this data. It will also help to do atmospheric correction.

14

7. DATA PROCESSING: INDEX CALCULATION

It is important to understand how CAI works. The Cellulose Absorption Index is a result from absorption of the plant macromolecules that are cellulose (≈ 40 %) and lignin (6 – 14 %). Those plant parts exhibit absorption at around 2100 nm (as Figure 7 shows); therefore, most crop residues on soil surface also show this absorption feature. This index needs as input 30 nm-width spectral bands in reflectance (R). Those three bands are centered at 2015 nm (R2.0), 2106 nm (R2.1) and 2195 nm (R2.2). Because of the spectral resolution These 30 nm bands for CAI are averages over three Hyperion bands each one for R2.0: B183, B184, B185*; R2.1: B195, B196, B197; and R2.2: B204, B205, B206. Note*: in this project Hyperion band B185 was excluded because of severe striping and it was very noisy. The average can be done in Band Math by given more weight to the central bands as follow: Average Reflectance = 0.25*B1+0.50*B2+0.25*B3. Then CAI can be calculated with the definition: CAI  0.5R2.0  R2.2 − R2.1

λ (μm) 2.0 2.1 2.2

bare soil

residue cover Reflectance

Figure 7. Reflectance at bands in the Cellulose Absorption Index.

The CAI uses the bands at 2000 nm and 2200 nm in order to normalize the 2100 nm wavelength absorption feature. As Figure 7

15

shows when the 2100 nm feature gets deeper, then it implies more absorption (in other words less reflectance is happening), more crop residues are present, and as a result one obtains higher index.

Figure 8. Cellulose Absorption Index image for central Tipton County.

Image data results from the CAI (as a single band) for the selected agricultural areas (central Tipton County) shows greater abundance of crop residues as bright areas, and bare soil as dark areas as can be seen in Figure 8. There are also mixing pixel with intermediate colors. This can later on be classified into conservative and conventional tillage and crop residues management practices of the agricultural fields. The index image also shows the two cities (Tipton and Windfall) and some winter crops (C3 grasses and/or winter wheat) as bright areas

16

because of the presence of trees and growing vegetation in both of them; therefore cellulose absorption would be present.

After running the CAI in the selected Hyperion data subset for central Tipton County (without band 185), the next step was to register the image to the panchromatic Landsat subset (15 m spatial resolution) as shown in Figure 8. This was accomplished by 20 ground control points and a RMS error of 0.511m.

7. RESULTS

Unsupervised classification by ISODATA algorithm was performed on CAI image (previous Figure 8), and then crop residues classes were assigned as crop management classes. Three classification classes were assigned: one with less than 30% crop residues, conservation management, one with more than 30% crop residues, conventional management, and a class with other areas which corresponded to urban areas of the two cities, small houses farms scattered in the image, infrastructure, roadways and growing crops. The result of ISODATA is presented in Figure 9.

Figure 9. ISODATA classification on CAI image for central Tipton County with Tillage systems classes

17

Then the areas corresponding to the two tillage classes were extracted from ISODATA statistic output and finally normalized to 1. By comparing the classification results, table 1, with data from the Crop Residues Management Survey (www.conservationinformation.org) could a statistical comparing be performed by a Z- test, z ≈ (df: 1) = 7.56, P-value < 0.05. This showed that it were significant so both sources behave differently. However, in numerical terms, the data behave consistently similar for the tillage systems. The difference is just 3 %. It is very likely that the test statistic is not powerful enough to detect differences (z-test is a conservative statistic test), specially when having such limited amount of degree of freedom. In this case, we different just two classes conservative and conventional tillage; however, more classes can be differentiated, and then more detail and more powerful test might identify conclusive similarity.

Table 1. Classification results vs. Crop Management Survey field data.

Group /		CRM		CAI ISODATA
Category		Survey		Classification
No tillage		0.373		0.338
Tillage	0.627	0.662
Agricultural land		1.000		1.000

In addition to the previous analysis it was explored supervised classification by minimum distance on CAI image and maximum likelihood in ALI image. It is relevant to clarify that Envy 4.2 software requests two bands to perform a parametric supervised classification by

maximum likelihood as well as mahalanobis needs two or more bands, while the calculated index corresponds to a single band. Minimum distance can be performed in a single band; however, results were consistently poor as can be seen in Figure 10. Even after trying several times using different training sets. Minimum distance had an overall accuracy of = 90.8 % (1241/1366) and Kappa Coefficient = 0.82, which are good values; however, it totally confused the urban areas and the growing crops with no tillage areas yielding a bias result of 69 % of surface no tillage systems and 31 % under tillage systems. These results are totally different than field data collected during the Crop Residues Management Survey. Minimum distance classifier is way to simple to do a good job in this case, so we disqualified this classification result.

Tilled

No-tilled

Other areas

Figure 9. Minimum distance classification on CAI image for central Tipton County with Tillage systems classes

19

Because of the impediment of the Envy 4.2 software (requesting two or more bands to perform maximum likelihood or mahalanobis), we decided to use ALI image with spectral bands 5 & 6 (average), 4 and 3 as RGB (as shown in Figure 2) to perform classification by maximum likelihood with training sets for three classes and 1100 to 1200 pixels per class. Results were better than in minimum distance because cities were discriminated as well as growing crop. Maximum likelihood had an overall accuracy of = 98.8 % (3440/3480) and Kappa Coefficient = 0.98, which are good values; however, it underestimate the field under conventional tillage with 43.3 %, while the no tillage or conservational tillage system was 57 %. These values are closer than minimum distance on CAI image; however, not closed to the Crop Residues Management Survey. One has to remember that this classification was done not on the CAI image but in ALI RGB image. Even so, some agricultural fields were clearly identified by this classification.

Tilled

No-tilled

Other areas

Figure 10. Maximum likelihood classification on ALI RGB image for central Tipton County with Tillage systems classes

20

8. ISSUES AND LEARNING

During this project have we had several issues and difficulties but managed and learn a lot from each. The first difficulties were to destripe the Hyperion data to get rid of as many bad columns/stripes and noise as possible. Also the conversion from Digital Numbers, DN, to radiance and reflectance by the program FLAASH were not easy managed.

We have also found some limitations for the index, which are: living plants have also cellulose/lignin that will affect the results. Also the soil moisture content will change over time also that different kind of soils reacts in different ways by soil moisture content.

During this project have we used a number of software and that include matlab for the destriping code, Minimum Noise Fraction (MNF), FLAASH, band math, ISODATA algorithm and the Cellulose Absorption Index (CAI) and also ENVI as the program we have been working in.

9. CONCLUSIONS AND FUTURE WORKS

The results from this study can we draw the conclusion that the Cellulose Absorption Index is a very powerful tool for mapping crop residues with hyperspectral data. The results we performed could not statistically agree with field data at County level; however, in numerical they behaved very similar with a difference of just 3 %.

21

We would like to see future work in some comparisons/interpretations between hyperspectral data and multispectral data. Also to compare multispectral data indexes (Landsat and ALI) with hyperspectral indexes. Some indexes that could be interesting to investigate in can be Normalized Differential Index (NDVI), and Normalized Differential Senecent Index (NDSVI). One step can also be to perform a detailed supervised classification with more classes, and compare different algorithms for example Spectral Angle Mapper (SAM), Cosines angle, Maximum Liklehood, Minimum distance and Mahalanobis algorithms.

A project with multitemporal studies looking at fresh surface residues and how the decomposition rate changes with time would also be very interesting.

22

ANNEX 1. CORRECTED BANDS AND COLUMNS

Band 1 183 OK

[24 36 54 72 112 116 120 124 128 160 180 184 228 240]

[37 55 73 113 117 121 129 161 181 185 229 241] Band 2 184 OK

[2 13 24 40 51 60 69 72 88 90 93 102 104 112 123 126 134 144 152 159

165 176 183 190 196 204 206 216 228 230 240 248 253]

[3 25 61 73 113 124 127 145 153 166 184 185 191 197 198 207 217 229

231 241 249 254]

[114 154 186 199 232 242]

[115 116 117 155 156 157 187 121 122]

Band 3 185 OK

[2 7 12 22 24 32 36 40 52 55 60 62 76 80 84 90 100 112 116 130 134

152 156 170 172 182 187 196 206 216 222 230 234 240 246 252]

[3 13 25 33 37 41 53 56 77 81 85 91 95 101 113 117 131 135 153 157

173 183 188 197 207 217 223 231 235 241 253]

[102]

[103]

[22 94 112 122 180 194 200 246 226]

[23 113 123 195 201 247 227]

[181 196 202 248 228]

[182 203 249 229]

[63 124 172 186 210 246]

[125 211 247]

85 212

86 213

220

87

23

ANNEX 2. COPY OF IN-RUNNING MATLAB CODE FOR DE-STRIPPING

>> cor_str

File to De-stripe is:

OriginalB9s

Info file is:

B9.info.txt

Methods to correct streaks

[1] Additive

Please wait: creating destriped file.

if you left any band, you can run this code again.

It will not affect those bands that have been corrected

Bands that need to be corrected: put them into like [ 1 2 ], (0 for all) [3]

Enter your sliding window size, the default window size is 32 pixels

32

Band 3:

Steaked columns: put them into like [ 1 2 ], (0 for all)[12 57 162]

Band 3: 12 57 162

processing column 12

processing column 57

processing column 162

>>

24

8. REFERENCES

Martin, A. 2002. Detection of Crop Residues with four tillage systems using Remote Sensing Techniques. MS Thesis, Purdue University. 115 p

Bannari A, A. Pacheco, K. Staenz, H. McNairn, K. Omari. Estimating and mapping crop residues cover on agricultural lands using hyperspectral and IKONOS data. Remote Sensing of Environment 104 (2006) 447–459

Biard F, Baret F. Crop Residue Estimation Multiband Reflectance Using Multiband Reflectance. Remote Sens. Environ. 59:530-536 (1997)

Daughtry, C; J. E. McMurtrey III, M. S. Kim, E. W. Chappelle. Estimating Crop Residue Fluorescence Imaging Cover by Blue. Remote Sens. Environ. 60:14-21 (1997)

Daughtry, C., E.R. Hunt Jr., J.E. McMurtrey III. Assessing crop residue cover using shortwave infrared reflectance. Remote Sensing of Environment 90 (2004) 126–134

Daughtry C., P.C. Doraiswamy a, E.R. Hunt Jr.a, A.J. Stern a, J.E. McMurtrey IIIa, J.H. Prueger. Remote sensing of crop residue cover and soil tillage intensity. Soil & Tillage Research 91 (2006) 101–108

Lal, R. 2002. Soil carbon dynamics in cropland and rangeland Environmental Pollution 116:353–362

Lal, R. 2004. Soil carbon sequestration impacts on global climate change and food security. Science 304:1623-1627.

McNairna, H.; C. Duguayb, B. Briscoc, T.J. Pultz. The effect of soil and crop residue characteristics on polarimetric radar response. Remote Sensing of Environment 80 (2002) 308– 320

Nagler, P. L., Y. Inoue,1, E.P. Glenn, A.L. Russ,2, C.S.T. Daughtry, Cellulose absorption index (CAI) to quantify mixed soil–plant litter scenes Remote Sensing of Environment 87 (2003) 310–325

25

South, S, Jiaguo Qi, David P. Lusch. Optimal classification methods for mapping agricultural tillage practices. Remote Sensing of Environment 91 (2004) 90–97

Su, H.; M. D. Ransom, E. T. Kanemasu. Simulating wheat crop residue reflectance with the SAIL model. Int. J. Remote Sensing, 1997, vol. 18, no. 10, 2261 - 2267

Thomas J. Jackson Peggy E. O'Neill Remote Sensing of Environment. Microwave emission and crop residues. Volume 36, Issue 2 , May 1991, Pages 129-136

West, T; Post, W. 2002. Soil organic carbon sequestration rates by tillage and crop rotation: a global data analysis. Soc. Am. J. 66:1930– 1946.

26

27