Tables 3and 4show error matrixes for the three residue cover categories using SLR for the test dataset and the complete
dataset
58bd8d2b482fb4daa58d4beaparison of time-series NDTI and NDVI values from the same pixel:non-conservation tillage (left);conservation tillage (right).
02
4681012
141618200
R a i n f a l l (c m )
N D T I
Day of Year
58bd8d2b482fb4daa58d4beaparison of time-series NDTI values with different levels of CRC.The magni-tude of change in NDTI differs depending on the type of tillage practice —i.e.,upon the amount of CRC left on the ground.In this instance,the tillage practice was applied around day 129(May 9),as is highlighted in the graph.Daily precipitation data (NOAA Cooperative Station 126340,Noblesville 3W)are also included.
Table 2
2010Indiana crop progress (Indiana Crop and Weather Report,2010).Week ending Corn (%)Soybean (%)Precip.ξ
(cm)
planted Emerged planted Emerged April 111NA ?NA NA 3.3April 1817NA NA NA 0April 2556512NA 1.3May 2712623NA 3.8May 98152359 1.9May 2388795034 3.4May 3094867052 1.4June 697928169 3.2June 20
100
100
91
85
8.6
*NA:not available.ξ
Precip.:weekly total precipitation for Central Indiana.
180 B.Zheng et al./Remote Sensing of Environment 117(2012)177–183
respectively.The overall accuracy is better than 90%and the Kappa coef ?cient (K )is 85%for both datasets.The K of 85%suggests that the classi ?cation accuracy is 85%better than chance alone.The user's accuracies are 72–100%,while the producer's accuracies are 83–100%for discriminating among three categories.5.2.Percentage change method
The high correlation between PC values and the CRC (R 2=0.80)(Fig.8)demonstrates the potential of this PC method for mapping CRC.Fig.8shows that the correct classi ?cation of classes with more than 70%CRC (green dots),more than 30%and less than 70%CRC (blue dots),and less than 30%CRC (yellow dots).Misclassi ?cation is shown in red dots.According to Fig.8,we determined that the classi-?cation rules for this study area are as follows:pixels that reveal a PC less than 40%are assigned to class CRC>70%;pixels with a PC larger than 40%but smaller than 70%are classi ?ed as 30%b CRC b 70%;and pixels with more than 70%change are assigned to non-conservation tillage (CRC b 30%).
The error matrix using the PC method is shown in Table 58bd8d2b482fb4daa58d4bea-pared to Table 4,the PC method resulted in the same overall accuracy and K .However,the user's accuracy of the class,30%b CRC b 70%,is slightly lower than that of minNDTI.We evaluated the difference be-tween the two classi ?cation accuracies using the minNDTI and PC methods using McNemar's test (Agresti,1996;Foody,2004)and
found no signi ?cant difference between two classi ?cation results (z=0.38b 1.96).6.Discussion
Both minNDTI and PC methods were able to classify CRC into three categories.minNDTI improves both continuous range mapping as well as categorical classi ?cation depending on user needs.Previous studies either classi ?ed CRC into two categories to achieve higher prediction accuracy (Gowda et al.,2001;Thoma et al.,2004),or found low correlations between tillage indices and CRC (Daughtry et al.,2006)using Landsat-based indices.The R 2of 0.11between CRC and NDTI reported by Daughtry et al.(2006)is probably because their Landsat image was acquired on June 12when most crops had emerged and confounded the NDTI signal.Other studies (Daughtry et al.,2005;Serbin et al.,2009b )suggested exclusion of pixels with green vegetation from the analysis,especially for Landsat-based till-age indices.Hyperspectral tillage indices are more effective for map-ping CRC than Landsat-based indices (Daughtry et al.,2005)because their narrow bands are more sensitive to crop residue and less sensitive to presence of green vegetation (Serbin et al.,2009b ).Nevertheless,pixels with green vegetation should be masked out using NDVI or other vegetation indices as suggested by Daughtry et al.(2005).Variation in soil moisture content may have negative effect on mapping CRC (Daughtry &Hunt,2008),but unfortunately,we don't have soil moisture data to examine the effect of soil moisture on our methods.No heavy rainfall happened immediately before our image dates (Fig.4).Therefore,there is no indication here that soil moisture in ?uences NDTI values.
Extracting minNDTI values from multi-temporal pro ?les can re-duce unmapped areas because this method can eliminate effects of green vegetation and avoid consideration of areas that farmers have not tilled yet.Watts et al.(2011)discovered that tillage classi ?cation accuracy was better using all ?ve available Landsat images instead
of
C R C %
minNDTI
Fig.5.Crop residue cover (CRC)as a function of minimum NDTI extracted from the time-series of Landsat images (calibration dataset:n =
31).
P r e d i c t e d C R C %
Measured CRC%
Fig.6.Measured vs.predicted crop residue cover (test dataset:n =
32).
C R C %
minNDTI
Fig.7.Crop residue cover (CRC)as a function of minimum NDTI extracted from the time-series of Landsat images (n =63).
Table 3
Error matrix for three residue cover classes using simple linear regression for test dataset.
Classi ?cation data
Reference data CRC b 30%
30%b CRC b 70%CRC >70%Total User accuracy CRC b 30%
100010100%30%b CRC b 70%260875%CRC >70%01131493%
Total
1271332
Producer's accuracy
83%
86%
100%
Overall accuracy:91%;Kappa coef ?cient:85%.Bold data are the number of pixels correctly assigned to each class.
181
B.Zheng et al./Remote Sensing of Environment 117(2012)177–183
using a single-date image,and demonstrated the importance of tem-poral frequency in tillage mapping.
Our methods are simple and can be easily adopted by others.The physical relationship is well explained by the SLR.Re ?ectance values at band 7(2080–2350nm)of TM/ETM+images decrease as CRC in-creases because crop residues have absorption features near 2100nm (Daughtry,2001).Thus,the NDTI has a positive linear relationship with CRC.Other non-linear methods,such as Arti ?cial Neural Networks (Sudheer et al.,2010),should also take into account temporal changes of agricultural surfaces when mapping tillage practices.Under similar soil moisture conditions,the PC method can mitigate soil color variation that could confound both single and multi-date approaches.The 40%break point of the PC method maximizes the classi ?cation accuracy for our study site.Logically,the break point should be 30%.The extra amount of change is probably due to residue weathering.Thus,one may adjust this value between 30and 40%regionally.However,further investigation is needed to con ?rm the causes.
Currently,image availability is one of the most important factors that constrains our ability to map CRC 58bd8d2b482fb4daa58d4beandsat provides global coverage at 30meter spatial resolution.Multi-temporal Landsat imag-ery is required to map CRC because its tillage indices can be biased by any green vegetation (Serbin et al.,2009b ).Multi-temporal methods for mapping CRC are subject to failure with insuf ?cient temporal cover-age of remotely sensed data.Watts et al.(2011)demonstrated the potential of STARFM-based synthetic dataset for mapping tillage practices,which is a potential solution for the lack of availability of cloud-free Landsat imagery.The planned version 2.0of Web-enabled Landsat Data (WELD)(Roy et al.,2010)could be another source of gap-free Landsat 7ETM+data.Our study area includes
ample Landsat 5TM and 7ETM+scenes to support a temporal anal-ysis for most of the years from 1999to 2010(Table 6).Table 6illus-trates the availability of Landsat data.Years with more than four Landsat scenes have a higher chance of success for mapping CRC ac-curately.Further studies are required to test the transferability of our method to new scenes.
7.Conclusions
