To discriminate specific crop of interest, in this work we have used temporal data. These temporal data sets were pre-processed with respect to geo-registration, band ratio and finally fuzzy based classification approach was applied. For accuracy assessment fuzzy based error matrix was applied and soft reference data was generated using IRS-P6, LISS-III temporal data sets. It has been tried to achieve defined objective of this research project and detailed methodology has been explained further.
Study Area
The present study is undertaken to explore the feasibility of utilizing AWIFS (IRS-P6) and LISS-III (IRS-P6) data for mapping and testing single class (Wheat) of interest. The test location was selected in regions providing a range of challenging environments for specific crop monitoring. india
The area is situated in the southern part of the Uttarakhand state between 29° 11' 21.99"N to 29 ° 42' 06.73"N and 78 ° 38'18.30"E to 79 ° 10' 191.05"E. The area is located in Terai region and is a part of Kumaon Division. The study area is famous for its agriculture and irrigation on synchronized patterns from the past as garner of popularity for its productivity in paddy crops in the whole Uttarakhand state. Khariff and Rabi are two major cropping seasons. The main Khariff crops are rice, soyabean, Urd, and Moong and the Rabi crops are wheat, barley, Gram, Masoor, Mustard, and Sunflower.
uttranchal pantnagar
Pre-processing of the Data
(Geometric Correction of (IRS-P6) LISS-III and AWIFS Data)
The reference LISS_III images available were already pre-processed as follows: SOI toposheets were scanned and converted into digital form before geo-registration. Images were co-registered with respect to the toposheet maps on 1:50,000 in UTM projection with WGS-84 North spheroid and datum, zone 45 (Kashipur Area) Uttarakhand. The images were re-sampled at 20m spatial resolution by using the nearest neighbour re-sampling method (first order polynomial) and total 17 GCPs were collected from the toposheet for the purpose of geo-registration.
The test images available for classification were similarly pre-processed. These AWIFS images of same area were co-registered with respect to the LISS-III dataset in UTM projection similar to LISS-III images. Similarly, the images were re-sampled at 60m spatial resolution by using the nearest neighbour re-sampling method and total 17 GCPs were collected from the LISS-III images for the purpose of geo-registration. The difference between test image (AWIFS) and reference image (LISS-III) is 1:3, which implies that, one pixel of AWIFS image is equal to nine pixels of LISS-III image.
Methodology Adopted
In the present study the medium spatial resolution AWIFS data sets are used. In order to accurately map the specific crop-wheat, we use temporal data sets. Different datasets with varying temporal resolution were taken in order to find out the most suitable time-series (multi-date) image set that gives the best accuracy of classification.
In 2008-2009 Rabi Season- Wheat Growth Period
Nov Dec Jan Feb Mar Apr
(image3)
(image4)
(image2)
(image5)
(image1)
Suitable band ratio method
Suitable fuzzy based supervised classification for extracting single class of interest
Accuracy assessment (Using soft reference data)
Figure: Methodology adopted
The enhanced images were available for our study. The images are, in general, enhanced by pre-processing like geometric and atmospheric corrections. The images are geo-referenced using ground control points and then registered. This work is done using commercially available ERDAS software tool.
Training sites for wheat were identified in the digital LISS III and AWIFS image with the help of ground control point details provided and the visually interpreted FCC images.
Band ratio techniques were employed to eliminate the effect of slope and aspect and the difference of illumination. Normalized Difference Vegetation Index (NDVI), MIN/MAX and Transformed Vegetation Index (TVI) were taken for all the images from AWIFS and LISS-III. This was done using the SMIC software tool.
The fuzzy set theory based sub-pixel classification technique was used for further classification using temporal data. The samples of wheat were taken from both AWIFS and LISS-III time-series images respectively. SMIC software tool at IIRS was used. Sample was taken in the manner that, one image was in the viewer while, other temporal images were arranged as the click on viewer pick the value of all the images at that particular pixel (class). The fuzzy based classifier was used for classification. The classification was done by Possibilistic c-Means classifier approach. The output was a single classified fractional image, from which we analyzed the single class (wheat crop).
Further, the accuracy assessment of all the classified AWIFS fractional images was taken with respect to LISS-III image. FERM (Fuzzy Error Matrix) was used for accuracy assessment.
Concepts Employed
Band Ratio Techniques
The process of dividing the pixel values in one spectral band by the corresponding pixel values in another spectral band is known as band ratioing. It is just a simple transformation procedure applied to remote sensing images. The shape of the spectral reflectance curves of different land cover / land use types can be identified by this technique. Secondly it can reduce the recorded unwanted topographic effect like slope and aspect and eliminate the effects of difference in illumination. The ratioing technique such as spectral vegetation index are widely useful and benefits the numerous disciplines like assessment of biomass, water use, plant health, crop production and plant stress. Vegetation Indices combine different spectral bands, quantitatively measure and evaluate the vegetation cover density, classify the crop and also assist in crop discrimination.
The ratioing techniques used in this study are as follows
Min/Max
It is one of the easy to compute band ratio technique. It is obtained by taking the ratio between the minimum band value and maximum band value corresponding to that pixel.
Ratio Image = Minimum Value/Maximum Value
NDVI
One of the first successful vegetation indices based on band ratioing was developed by Rouse et al. (1973). They computed the normalized differences of brightness values from MSS7 and MSS5 for monitoring vegetation. They called it the Normalized Differences Vegetation Index (NDVI) Jenson 1996. The NDVI is measured in scale of -1 to +1. Snow, water bodies, desert and exposed soils come in a range of -0.2 to 0.05. While the progressively increasing amounts of green vegetation come in the range of 0.05 to 0.7. NDVI data are strongly correlated with the fraction of photo synthetically active radiation (0.4 to 0.7 µ.m. wave length) absorbed by vegetation canopies. The brighter the image pixel after classification, the greater the amount of photosynthesizing vegetation present.
NDVI = (NIR-R)/ (NIR+R)
TVI
TVI was introduced by Deering et al. (1975). They added 0.5 to NDVI and took the square root, producing the transformed vegetation index (TVI). The TVI can be linearly correlated with leaf area index, and has a higher sensitivity than that of NDVI in high biomass area (Huete et al., 2002, Sakamoto et al., 2005). Since the biomass of crop fields is low compared with forests, TVI retains linearity in crop fields. Furthermore, TVI is more practical than NDVI when humidity is high. (Sakamoto et al., 2005)
TVI = √ (NDVI + 0.5)
Image Classification Techniques
For the preparation of thematic maps and quantitative analysis of the images, capability of computer to interpret the images, identify pixels and label them based on their numerical properties is exploited. The method is commonly referred to as Image Classification. While the images can processed in digital environment it is better known as digital image classification.
A traditional hard classification technique does not help in this type of situation. Fuzzy logic may be beneficial where multiple classes exist within a pixel. To incorporate the mixed pixel problem in past researchers have proposed the 'soft' classification technique that decomposes the pixel into class proportions; fuzzy classification is a soft classification technique, which deals with vagueness, ambiguity and uncertainty in class definition. Therefore fuzzy classification technique is probably the best technique to extract the single class from the image and differentiate with other classes.
The concept of 'fuzzy set' theory was introduced by Zadeh, to deal with the uncertainty in class definition. The fuzzy set theory introduces the vagueness by eliminating the crisp boundaries into degree of membership to non-membership function Binaghi et al. 1999. It represents the situations where an individual pixel is not a member for a single cluster, but member for all clusters with different degree of belongingness Dutta 2009.
In this study fuzzy logic based algorithm, which is independent of statistical distribution assumption of data, has been studied to extract single land cover class from remote sensing multi-spectral images. Fuzzy based classifier in this work has been implemented in such a manner that remote sensing image from any sensor can be used for single class extraction.
Temporal Data Approach
At the third level of classification where we map the specific vegetation, mapping based on classification using single date image has been only moderately successful.
The first problem in this study is to extract single crop of interest from the coarse resolution satellite image. There is problem when extracting single class with single date image. The image consists of pixels. A pixel value (brightness value) recorded is result of interaction of electromagnetic waves with the ground objects and/or atmosphere. In addition crops may have similar spectral response patterns having only slight differences. Hence, the spectral response recorded by sensor may differ for same or similar type of classes while it may be possible that the dissimilar entities may show similar spectral response, depending on ground or atmospheric conditions. This introduces errors. So it is impossible to accurately extract the single crop using the single date imagery.
To overcome the above issues, temporal data approach best for specific crop mapping. With the time series data the spectral response of the class proportions can be recognized and differentiate from other classes. Therefore single crop of interest has been processed using temporal data.
Accuracy Assessment
Accuracy assessment and validation for sub-pixel classifiers is still a subject of research. No standard methods are available for sub-pixel classifiers, unlike that for hard-classifiers such as error matrix and kappa coefficient. For the validation of the result, FERM (Fuzzy Error Matrix, Binaghi et al.) with the help of fuzzy set theory based sub-pixel classifier was used in this study. The accuracy assessment of AWIFS fraction images has been done with respect to LISS-III fraction image. The fuzzy error matrix (FERM) method has been employed to compute the accuracy.
Classified Image
Mixed Pixel
Reference Image
Figure: Sub-Pixel Accuracy Assessment Method
(Comparing coarse resolution image with fine resolution image)
Fuzzy Error Matrix (FERM):
For the assessment of soft classified data various suggestion have been made, among which fuzzy error matrix introduced by Binaghi et al., 1999 is one of the most suitable approach. FERM takes the fraction soft classified images (floating value or non-negative real number) as input instead of traditional hard classified images (integer value). The layout of a fuzzy error matrix is similar to that of the traditional error matrix that is used for accessing the accuracy of hard classifiers. The element of the fuzzy error matrix represent class proportions, corresponding to soft reference data (Rn) and soft classified data (Cm), in classes n and m respectively. Fuzzy minimum operator are used to construct the fuzzy error matrix and determine the matrix elements M(m, n) in which the degree of membership in the fuzzy interaction (Cm ∩‌ Rn) is computed as (Kumar et al., 2007)
M (m,n) = |Cm ∩‌Rn| ∑xεX min(µCm, µRn)
Where X is the testing sample dataset; x is a testing sample in X and µCn and µRn are the class membership of testing sample x in Rn and Cm, respectively.
In our study, fuzzy based error matrix was applied for accuracy assessment and soft reference data was generated using IRS-P6, LISS-III temporal data sets. Producer, user and overall accuracies of classification were computed by this fuzzy error matrix (FERM) accuracy assessment method.
Results and Discussion
Sample Output Images
Following are the corresponding subsections of output images from temporal data set-6.
Acquisition Date Input Images: Min/Max Band Ratio Image: Final Wheat Mapping
21 Nov 2008 C:\Users\Amol\Desktop\Images_ Report\a.jpg C:\Users\Amol\Desktop\Images_ Report\m6_a.jpg
08 Jan 2009 C:\Users\Amol\Desktop\Images_ Report\b.jpg C:\Users\Amol\Desktop\Images_ Report\m6_b.jpg
27 Jan 2009 C:\Users\Amol\Desktop\Images_ Report\c.jpg C:\Users\Amol\Desktop\Images_ Report\m6_c.jpg
15 Feb 2009 C:\Users\Amol\Desktop\Images_ Report\d.jpg C:\Users\Amol\Desktop\Images_ Report\m6_g.jpg C:\Users\Amol\Desktop\Images_ Report\m6_final.jpg
25 Feb 2009 C:\Users\Amol\Desktop\Images_ Report\e.jpg C:\Users\Amol\Desktop\Images_ Report\m6_e.jpg
11 March 2009 C:\Users\Amol\Desktop\Images_ Report\f.jpg C:\Users\Amol\Desktop\Images_ Report\m6_f.jpg
31 March 2009 C:\Users\Amol\Desktop\Images_ Report\g.jpg C:\Users\Amol\Desktop\Images_ Report\m6_g.jpg
Classification Accuracy
To understand and illustrate the efficiency of the fuzzy set theory based sub-pixel classifier along with various band ratio techniques and temporal data sets, used in estimation of specific crop of interest, accuracy assessment and review is required.
The accuracies of the fraction images (proportion of wheat crop) generated by sub-pixel classifier of Kashipur area are shown in table 4.
Table 4: Accuracy Assessment of wheat crop mapping in Kashipur area
Vegetation Index
Accuracy
Set 1
(%)
Set 2
(%)
Set 3
(%)
Set 4
(%)
Set 5
(%)
Set 6
(%)
Set 7
(%)
Set 8
(%)
Min/Max
User
86.50
88.06
86.01
87.15
89.26
89.58
87.83
90.64
Producer
96.30
92.92
95.36
95.00
95.34
96.07
95.62
95.63
Over All
96.30
92.92
95.36
95.00
95.34
96.07
95.62
95.63
NDVI
User
92.40
89.68
92.14
92.20
93.24
93.37
91.87
94.13
Producer
94.24
93.56
94.67
95.68
94.99
95.97
95.10
95.69
Over All
94.24
93.56
94.67
95.68
94.99
95.97
95.10
95.69
TVI
User
73.83
80.13
76.00
76.46
78.73
79.86
79.06
79.37
Producer
95.23
93.24
94.53
93.66
92.79
96.15
94.66
96.09
Over All
95.23
93.24
94.53
93.66
92.79
96.15
94.66
96.09
Graphical Representation
Results
The Min/Max band ratio technique gives the best accuracy of wheat crop classification in four sets of temporal data and is followed by TVI vegetation index which gives maximum accuracy of wheat crop mapping in three sets of temporal data.
The accuracy of classification of wheat crop, obtained by applying the Min/Max band ratio technique is highest in set-1 and almost close to this maximum value in set-6. The accuracy of wheat crop mapping using NDVI vegetation index is maximum in set-6 and same is true for TVI vegetation index.
Discussion
In this study fuzzy set theory based sub-pixel classifier has been undertaken for extracting the single crop type (wheat) using multi-spectral satellite images. Three indices Min/Max, NDVI and TVI were used to monitor, estimate and discriminate wheat other crop types. Fuzzy set theory based sub-pixel classifier which gives the fraction images was applied to band ratio images. Fuzzy Error Matrix (FERM) method was applied for accuracy assessment of wheat crop mapping and validation of the results.
Irrespective of the number of images in each multi-date temporal data set and irrespective the vegetation index used, this fuzzy set theory based sub-pixel classifier gives fairly good overall accuracy, which is always above 92.5%. It thereby establishes credentials of fuzzy set theory based classifier for mapping specific crop.
Further comparative study of the effects on accuracy, of vegetation index used to monitor and discriminate wheat crop shows that, Min/Max vegetation index is best suited to monitor and discriminate wheat crop while using temporal data sets. Min/Max vegetation index gives best accuracy in half of the temporal data sets tested. However it does not show specific pattern in giving accuracy as the temporal data sets vary. TVI vegetation index also gives maximum accuracy in three sets, but the variations, in the accuracies is maximum. Though NDVI ratio does not get the highest spot in the accuracy of classification, there is some pattern seen in the accuracy of classification under NDVI vegetation index. It in general increases from set-2 to set-6 which can be explained vaguely as better accuracy for better temporal resolution.
Review of the accuracy of classification, with respect to temporal resolution and the dates of the images used, suggest that in general overall accuracy increases from set2 to set6 as the number of images used increase. Classification accuracy is highest for set6. As can be seen from Table3, set6 contains images from sowing to harvesting of the crop and almost at temporal resolution of 20 days. There is a dip in the classification accuracy in set2. That is probably due to inclusion of image of date 6 Nov. 2008 (before sowing of the crop). It also explains the dip in classification accuracy for set7.
Conclusions and Recommendations
Motive was to map single crop of interest using fuzzy based classifier with the help of time-series multi-spectral satellite images. The temporal data helps in discrimination of crop (especially wheat as tested here) from other crops.
It has been observed from this work that time-series multi-spectral images used for specific crop mapping give good overall accuracy of classification. According to results obtained from this work, Min/Max vegetation index gives maximum accuracy of mapping for wheat crop. NDVI vegetation index is also fairly good and TVI can also survive the purpose.
Temporal data set containing multi-date images that evenly cover the entire crop life cycle (sowing to harvesting) gives the maximum accurate mapping of the crop. In case of wheat, the temporal resolution of 20 days gives better accuracy. However, inclusion of images of dates other then cropping season (those before sowing the crop and after harvesting the crop) will reduce the overall accuracy of classification.
