Feature Selection Methods
Feature selection is required while using the multi-temporal data to reduce the number of unwanted scenes. It is a method to segregate the combination of useful bands or scenes for identifying a particular land cover using the temporal multispectral data (Bruzzone and Serpico, 2000). The increasing number of input features increases the computational requirement as well as cost. The feature selection method is used for generating the suitable combinations of datasets for achieving the maximum accuracy with low cost and less labor (Tso and Mather, 2009). The feature selection is different from the feature extraction method (e.g., principal component analysis, decision tree) in the sense that it uses the separability distances in the input feature space to derive the best sub feature dimensions. On the other hand, feature extraction compresses the information available in the original feature space with a drawback of losing the physical signif cance of features (Bruzzone and Serpico, 2000).
There are few methods, such as city block distance, Euclidean distance, angular separation, normalized city block distance, divergence, transformed divergence (TD), Bhattacharyya’s distance, and Jeffreys-Matusita (JM) to measure the separability distances (Ghosh, 2013). Out of all these methods, the last four methods are commonly used for feature selection by determining the separability measure of remote sensing data, which is further used as an important criterion for feature selection. The separability measure tries to select the best number of bands to be used out of the given dataset. Suppose that there are n numbers of bands in a given dataset, and an analyst is interested in f nding the best q number of bands, then the number of band combination C to be examined at a time can be expressed as in Equation (6.2) (Jensen, 1996; Ghosh, 2013):
Transformed divergence (TD) (Swain and Davis, 1978) and Jeffreys-Matusita (JM) distance (Swain and Davis, 1978) separability approach are commonly used separability measures. Transformed divergence (TD) can be expressed as in Equation (6.3):
/ and j are two signatures (classes) being compared and Djj is the divergence.
The divergence Д, can be calculated by the following Equation (6.4). where:
Ci is the covariance matrix of class /, u. is the mean vector of class /, tr is the trace function, and т is the transpose function of the matrix.
The TD values range from 0 to 2,000. According to Jensen (1996), for TD value greater than 1,900, there is no overlapping between classes thereby having good separation. If it lies between 1,700 and 1,900, the separation is fairly good, whereas for less than 1,700, separation is poor. On the other hand, the JM distance ranges from 0 to 1,414. The JM distance can be expressed as in Equations (6.5) and (6.6):
i and j are two signatures (classes) being compared,
Ci is the covariance matrix of class i, u( is the mean vector of class i, and |C, | is the determinant of matrix C, .
Some Case Studies for Temporal Data Analysis
The dynamic nature of a few land cover over a period is a triggering factor to utilize the temporal data for mapping a specif c land cover class. Let us take an example of a vegetation class in which the phenological changes over a period of time will be a unique factor to discriminate it from the other classes. In many studies, the temporal remote sensing data have been used for different applications such as estimation of forest biomass (Powell et al., 2010), f ood study (Sakamoto et al., 2007), forest f res (Goetz et al., 2006; Morton et al., 2011), forest mapping (Hilker et al., 2009) and landscape changes (Millward et al., 2006). The multi-temporal MODIS satellite data have been used for the crop studies (Xiao et al., 2006; Wardlow et al., 2007; Wardlow and Egbert, 2008, 2010; Pan et al., 2012, Upadhyay et al., 2016).
Due to its coarse spatial resolution, the MODIS dataset is suitable for crop mapping at local to regional scales. The MODIS time series data at 250 m spatial resolution has been used by researchers for identif cation of forest area estimation (Maselli, 2011), tropical forest phenology (Pennec et al., 2011), gross primary production (Schubert et al., 2012), and identif cation of cropping activity (Pringle et al., 2012). Some of the case studies for identif cation of land cover are discussed in the following.
- 1. Wang and Tenhunen (2004) used the multi-temporal NDVI data from NOAA-AVHRR for different vegetation mapping in northeastern China for the year 1997. Supervised minimum distance and unsupervised Л-means classif cation methods have been applied on the temporal NDVI data and its phenology based derived matrices such as maximum, mean, threshold, amplitude, total length of growing season, fraction of growing season during green up, rate of green up, rate of senescence, etc. The overall accuracy for NDVI temporal prof le for unsupervised Л-means and supervised minimum distance were 52% and 50%, respectively. The overall accuracy for NDVI derived matrix was below 50%. Thus, classif cations based on the NDVI temporal prof le were better than those with the derived matrices.
- 2. Blaes et al. (2005) used the three optical images along with a number of time series synthetic aperture radar (SAR) images for crop identif cation. The idea of using the SAR images along with the optical was to overcome the problem due to cloud cover conditions and to guarantee necessary temporal frequency throughout the growing season. The classif cation was performed by the different combination of optical or SAR imagery independently. The main focus was to study the effect due to inclusion of SAR images on the optical images. It was found that the classif cation accuracy increased by at least 5% when SAR images were combined with the optical images alone.
- 3. Xiao et al. (2006) mapped paddy rice f elds in south Asia and southeast Asia using the multi-temporal MODIS images. A MOD09A1 product with a spatial resolution of 500 m and composite period of 8 days was used for the study. Out of 46 tiles of MOD09A1 for the year 2002, only 23 were selected for the study. The paddy rice mapping algorithm that uses the time series of MODIS derived vegetation indices was used for the analysis. The resultant maps were compared with the agricultural statistical data at national and sub national levels. The outputs for the MODIS rice algorithm were similar to the database derived from the census statistics.
- 4. Wardlow et al. (2007) investigated that the MODIS 250 m 12-month time series (January-December 2001) VI data. It was found that the data had the suff cient spatial, spectral, and temporal resolution to discriminate the crop types for Kansas in U.S. central Great Plains. Climatic and management practice variation was also detected for the crop class of study in the time series data. The phenological prof les which were spectrally and temporally different for different crops had been observed. A similar cropping pattern was observed for the MODIS 250 m and Landsat ETM+ 30 m imagery. It was found that MODIS 250 m is an appropriate scale to measure the general crop mapping pattern for the U.S. central Great Plains with the f eld with size 32.4 hectare or larger. The possibility of sub pixel un-mixing was also carried out to estimate the proportion of specif c land cover class.
- 5. Wardlow and Egbert (2008) evaluated the applicability of time series MODIS 250 m normalized difference vegetation index (NDVI) data spanning from March 22 to November 1,2001, for large-area crop-related LULC mapping over the U.S. central Great Plains. A hierarchical crop mapping protocol was applied to a decision tree classifer using multi-temporal NDVI data collected over the crop growing season for the state of Kansas. Classif cation accuracies for the time series MODIS NDVI derived crop maps were greater than 80%. Overall accuracies ranged from 94% to 84% for the general crop map and summer crop map, respectively
- 6. Tingting and Chuang (2010) used the NDVI, NDWI, and normalized difference soil index (NDSoI) based time series spectral indices (12 periods out of the time series dataset from April 2007 to October 2007) for identifying the rice crop in the Chao Phraya Basin of Thailand. The f rst principal component corresponding to each of the three MODIS time series was combined to create a new dataset. A linear spectral un-mixing was then applied to this merged data to create another data set. Thereafter, using the composition of NDVI, NDWI, and NDSoI values in each pixel, agricultural crop land has been separated into upland and paddy f elds in Thailand by using support vector machine (SVM).
- 7. Wardlow and Egbert (2010) performed a comparative study between Terra MODIS-250 m NDVI and EVI data, acquired from March 22 to September 30, 2001, for different crop mapping. The study was carried out under the assumption that EVI is more sensitive for crop mapping studies. The study was carried for the U.S. central Great Plains for general crop types, summer crop types, and irrigated and non-irrigated crops. It was observed that the NDVI and EVI produced equivalent crop classif cation with a subtle difference in their multi-temporal behavior. The overall and class specif c classif cation accuracies were greater than 85% for both NDVI and EVI. The variation in the classif cation accuracy between the maps was of the order of 3%, and their pixel level agreement was greater than 90%. Since this study was performed for a small geographical area and for a single season, the applicability of this study can be verif ed after investigating its inter-annual climatic behavior and performing it for other major agricultural regions of the world.
- 8. Potgieter et al. (2010) provided the early season information on winter crop (wheat, barley, and chickpea) area estimates using the multi-temporal MODIS 250 m EVI data acquired for the period 2003-2004 for a study area in Queensland, Australia. This study was aimed to fulfil the requirement of the early estimates of net crop production before the harvest, which is useful for many applications such as the grain industry, disaster relief, and drought declaration. The unsupervised k-means algorithm was used for classif cation. The study shows that the multi-temporal remote sensing approach could be used for the early season crop area prediction, at least 1-2 months before the harvesting date.
- 9. Atzberger and Eilers (2011) used a time series data consisting of the 10-day maximum value composite images from the SPOT VGT for monitoring the vegetation activity and phenology in South America at a spatial resolution of 1 km, from April 1998 to December 2008. The Whittaker smoother (WS) flter was applied on the time series data to handle the missing data, flter noise, and construct high quality NDVI time series. The geostatistical variogram technique was applied to reveal signal to noise ratio (SNR) WS fltered images. It was found that the fltered time series had the potential to distinguish between various plant functional types, as well as a key for various phenological markers. Thus, it was concluded that the time series datasets have great potential for vegetation and environmental related studies.
- 10. Alcantara et al. (2012) used the multi-temporal Terra and Aqua MODIS 250 m NDVI data for mapping of abandoned agricultural felds in Eastern Europe. TIMESAT software-derived phenological parameters were used as input parameter for SVM based classifcation of these agriculture felds. For classifcation, an overall accuracy of 65% for growing season has been achieved. Although the multi-year MODIS NDVI data does not increase the classifcation accuracy, but by using phenology matrices the accuracy has been increased by 8%.
- 11. Gonqalves et al. (2012) applied univariate and multi-variate statistical forecasting models to compute the water requirement satisfaction index (WRSI) and NDVI from AVHRR time series satellite images to monitor the sugarcane felds in Brazil from April 2001 to March 2008. Although both the models successfully predicted the NDVI values, the accuracy of the univariate model was higher than the multi-variate model. The average relative prediction error in case of univariate and multi-variate models were 5.6% and 13.4%, respectively. The forecasting of WRSI has given higher prediction error of order 49.7% and 47% for univariate and multivariate respectively. This is due to the fact that the WRSI values vary frequently throughout the season. An autocorrelation between these two indices has shown a time lag of one month for NDVI, which means the NDVI values change after approximately one month of climate change occurs.
- 12. Pan et al. (2012) proposed a crop proportion phenology index (CPPI) to estimate the winter wheat crop area up to the sub pixel level by using MODIS EVI time series for two agricultural regions in China. The phenological variables from October 2006 to June 2007 were used as an input to calculate CPPI for both the study areas. CPPI has been estimated by f tting either the linear or non-linear regression models on the phenological variables. The inversion model has been used to calculate the regression coeff cients using the training samples. The utility of the index was tested on two experimental areas in China. It was found that the CPPI performed well in fractional crop area predictions, with RMSE ranging roughly from 15% in the individual pixel to 5% above 6.25 km2.