1 (2017), 1, 18-40

Remote Sensing of Land

2582-3019

Improved Change Detection of Forests Using Landsat TM and ETM plus data

Kishor R. Sonawane 1 , Vijay Bhagat 1

1.Post-graduate Research Centre in Geography, Agasti Arts, Commerce and Dadasaheb Rupwate Science College, Akole-422601, Ahmednagar, Maharashtra (India).

Dr.Kishor R. Sonawane*

*.Post-graduate Research Centre in Geography, Agasti Arts, Commerce and Dadasaheb Rupwate Science College, Akole-422601, Ahmednagar, Maharashtra (India).

Dr.Vijay Bhagat 1

1.Post-graduate Research Centre in Geography, Agasti Arts, Commerce and Dadasaheb Rupwate Science College, Akole-422601, Ahmednagar, Maharashtra (India).

18-08-2017
15-11-2016
10-02-2017
02-02-2017

Graphical Abstract

Highlights

  1. Reliable change detection of forest using remote sensing data is challenging task.
  2. Ground information about stable land objects is useful for correction of target image.
  3. Statistical approach is useful for robust correction and estimations.
  4. Improved change detection technique for forest achieves precise estimations.
  5. The technique can be useful for precise change detection of land.

Abstract

Landsat TM and ETM+ datasets are useful for forest change detection (FCD) at good accuracy level. Classified forest maps have been prepared using NDVI calculated from Landsat-5 TM (2009) and Landsat-7 ETM+ (2002) datasets for FCD using post-classification technique. About 58.59% of reviewed area shows positive changes, 33.69% no-changes and 7.72% negative changes with 77.84% accuracy. This accuracy insists limitations of present FCD analysis. Therefore, improved post-classification technique was formulated for precise FCD using field data and statistical techniques. Information about stable land surface (water bodies, rocky lands, deep forests, etc.) was used for normalisation of exaggerated reflectance in vegetation indices i.e. greenness. About 70.08% land estimated using second approach shows stable vegetation, 23.59% positive changes and 6.33% negative changes. Higher accuracy (95.21%) itself shows improvement in FCD technique and efficient applicability for sustainable land management.

Keywords

Stable land objects , Forest change detection , Tasseled Cap Coefficient , LST , NDVI , Landsat-7 ETM+ , Landsat-5 TM

1 . INTRODUCTION

A forest is a community of living organisms which interact mutually with the physical environment. Forest covers approximately 30% (Carlowicz, 2012) to 31% (FAO, 2015) of the earth surface with complexities and self-regenerating capacities (Bauer et al., 1994; Virk and King, 2006; Healey et al., 2008). It is a source of organic carbon, which helps to maintain planetary climate, freshwater, biodiversity and useful to manage hazards like soil erosion, landslides, floods, etc. They are habitat of wildlife and regulate different cycles including hydrological, nutrient, atmospheric, etc. and conserve soils, water, etc. (Southworth, 2004).

Land under natural forest is estimated about 33.36 million km² (World Resources Institute [WRI]) to 39.88 million km² (World Conservation Monitoring Centre [WCMC]) excluding marine forests (Mishra et al., 2003; Kumsap et al., 2005; Fettig et al., 2007; Forkuo and Frimpong, 2012). However, from last some decades forest destruction and land appropriation increase, globally (Turker and Derenyi, 2000). Land under forest in India declined to 21.31% of Total Geographical Area (TGA) (ISFR, 2009 and FSI 2015) due to imbalance climatic conditions, soil degradation, deforestation, desertification and water stresses in drought conditions. India endowed with an immense variety of forest resources (Southworth, 2004). However, adverse changes in ecosystem are taking place with continuous pressures of an exploding population and the subsequent domestic needs including food, fuel, fodder, timber as well as industrial demands (Keenan et al., 1999). There are significant losses of forests at an alarming rate (Pant et al., 2000; Hayes and Sader, 2001). The ‘hotspots’ are identified and declared for tropical biodiversity in Western Ghats. These regions are house of rich biodiversity and globally endemic species (Fang and Xu, 2000). However, it is widely believed that the natural vegetation in this tropical region is losing the biodiversity at unprecedented rates (Panigrahy et al., 2010). Around, 275 million rural people (27%) in India are depends on forests for their subsistence and livelihoods (Kim et al., 2011), earning from trade of fuel wood, fodder, bamboo and minor forest products. It is notable that 17% of rural population in India depends on forests to meet their domestic energy (Drescher and Perera, 2010).

Governmental and non-governmental agencies are involved in conservation of forests from last some decades (Cohen et al., 1998). Planning and management for conservation of forest demands reliable information, investigation and analyses (Cohen et al., 1998; Hansen et al., 2000, Hayes and Sader, 2001; Bhagat, 2009). Many analysts have reported sophisticated techniques like remote sensing (Kennedy et al., 2009), geographic information system (Healey et al., 2008), global positioning system along with mathematical and statistical methods (Fettig et al., 2007) for change detection. Estimations of forests cover using satellite data can give satisfactorily results (Cano et al., 2006). Furthermore, change detection analysis can provide analytical data about forest degradation and conservation (Pouliot et al., 2002). However, reliable change detection techniques for forests using remote sensing data remains a challenging task (Coppin et al., 2004; Im et al.,2008;  Schwilch et al., 2011; Sommer et al., 2011; Bhagat, 2012). The results are susceptible to data quality, technical efficiency and suitability of selected techniques (Bhagat, 2012). Therefore, modified change detection technique designed for this study based on field checks and statistical analyses to get more precise analysis about forest changes (Southworth, 2004). The Landsat-5 TM and Landsat-7 ETM+ datasets have been used for detection of changes in forest cover in the study area. Two approaches have been adopted for this analysis: Approach I) post-classification technique and Approach II) improved post-classification technique (Chandio and Matori, 2011). Changes in forest were detected using post-classification techniques using NDVI (Fettig et al., 2007) calculated for Landsat-5 TM and Landsat-7 ETM+ datasets, whereas normalized indices and coefficients were used for improved approach of change detection.

Change detection analysis provides a thematic views to understand the natural and artificial behaviour of changes in land (Sommer et al., 2011) including 1) increase and decrease in area, 2) seasonal changes in forests, snow cover, coastline, ocean water, 3) mapping of floods, landslides, volcanic eruptions, corral rifts, wild animals, birds, 4) changes in near surface atmospheric conditions like temperature, snowfall, rainfall, clouds, fog, storms, and 5) human activities like military actions, observation, planning and management for war areas, coal mining, etc. (Lunetta et al., 2006). Scientists have used different methods of Digital Change Detection (DCD) including classification of multiband satellite data based on image ratio, tasseled cap coefficient, spectral vegetation indices, principal component analysis, change vector methods, threshold based classification, post-classification comparison, univariate image differencing, simultaneous analysis of multi temporal data, fusion approach, spectral classification, algebraic methods, regressions, etc. (Petit and Lambin, 2001; Bhagat, 2012). However, all methods are not ideally suitable, reliable and applicable to all surface change conditions (Du et al., 2002; Bhagat, 2012). Therefore, corrective measure has been suggested to achieve more precise results of forest change detection in this study. Correlation techniques have been used to find the suitable parameter for corrections from different multiband ratios (Coppin and Bauer, 1996) and spectral indices (Huete et al., 2002). Here, correlation and regression (Mountrakis et al., 2010) techniques have been used to determine variables for detection, estimations of exaggerations and corrections using information collected for stable land surface (water bodies, rocky lands, deep forest, etc.). The suggested technique for forest change detection can be useful for land management and especially forests.

2 . STUDY AREA

The study area (13859 ha) is a mountainous range between Pravara and Mula river basin from Western (Ghatghar) to Eastern borders (Washere) of Akole tahsil in Ahmednagar District (India) (Figure 1). The altitude varies from 560 to 1646 m Mean Sea Level (MSL). Geologically, this area is formed by basaltic rock (Gareeau et al., 2009) which prevents water to percolate. The depth and water-holding capacity of soils in the region (Zolekar and Bhagat, 2015) varies according to variations in slopes. The soils are very shallow at hilltop and depth increasing to foothill zones (Kumsap et al., 2005). Very shallow loamy, shallow clayey soils are found on the moderate (1°- 3°) and stiff (3°- 6°) slopes. Soil moisture impacts on the distribution of vegetation cover (Mishra et al., 2003). The forest cover varies with height, slopes, soil qualities, rainfall, etc. Foothill zones in Western parts show dense forest than hilltops with thin soils. Annual rainfall varies from 4937 mm at West and 1904 mm at East (Zolekar and Bhagat, 2015). The mean annual maximum and minimum temperatures are 39.80°C and 8.70°C, respectively. Indigenous people are engaged in agricultural activities (Theiler and Perkins, 2011) on land reclaimed from forests and dependent on forests for domestic needs. 

Figure 1. Study area: location map

 

3 . DATASET AND SOFTWARE

The analyses are based on remotely sensed data (Shalaby et al., 2006), statistical models (Golmehr, 2008) and field check information. The satellite data of Landsat-7 ETM+ (06th Nov. 2002) and Landsat-5 TM (1st Nov. 2009) (accessed on 01 January 2012) has been used for detection of changes in forest cover. Field check data was collected for verification of inferences. The remotely sensed dataset and data obtained in field checks were compiled, merged and loaded in the GIS image processing softwares, ILWIS v3.4 Academic, ERDAS Imagine v9.2 (© Hexagon) and ArcGIS v9.3 (© ESRI Ltd.). Garmin’s Global Positioning System (GPS) was used for ground verification. Correlation analyses was performed using ‘Karl Pearson techniques’ (Chen et al., 2001; Yang et al., 2014) available in Statistical Packages for the Social Sciences (SPSS © IBM) (Zhang et al., 2002; Fastring and Griffith, 2009).

 

4 . METHODOLOGY AND APPROACHES

Two approaches were used for detection of changes in forest cover: Approach I - post classification technique and Approach II - improved post-classification technique. Accuracy assessment was performed and results of analyses compared to check the applicability of the approaches. Statistical techniques were useful to inculcate robustness in the analysis for change detection (Theiler and Perkins, 2011).

4.1 Co-registration

Matching of multiple images captured at different time is critical task for change detection studies. Mis-registration of images drops accuracy in results obtained in change detection analysis (Burnicki et al., 2010; Bhagat, 2012; Pajares et al., 2012). Authors such as Giri Babu et al., (2014) and Tsai and Lin (2007) have suggested automatic tie point registration, pixel to pixel matching, ground control point registration, semi-automotive registration, etc. to obtained reliable results (Clifton 2003; Julien et al., 2011). Therefore, Landsat-7 ETM+ (t1) (2002) and Landsat-5 TM (t2) (2009) scenes were co-registered with the help of: 1) ground truth information (Hara et al., 2012) collected using Garmin GPS device and 2) topo-maps (Survey of India) with pixel to pixel match with 0.00012 Root Mean Square Error (RMSE) in ERDAS Imagine software.

4.2 Image Enhancement

Scientific understanding of behaviour and functioning of vegetation landscape, using remotely sensed data with different spectral, radiometric, temporal and spatial resolutions, have been important task in last some decades (Lippitt et al., 2011; Schwilch et al., 2011). Enhancement of images selected for analysis is highly required before further applications like classifications, estimations, etc. (Chowdhury et al., 2005; Weng et al., 2009; Ehlers et al., 2010). Vegetation indices like NDVI, LAI, TVI, etc. (Silleos et al., 2006) have been calculated using different bands e.g. red, infrared, thermal infrared, middle infrared and widely used for detection of land objects. Furthermore, coefficients like Soil Wetness Index (SWI), Normalized Difference Salinity Index (NDSI), Brightness, etc. have been used to study the distribution of soil moisture, soil salinity, etc. Spectral indices viz. NDVI, LAI and Land Surface Temperature Index (LSTI) and Tasseled Cap Coefficient Transformation have been calculated for enhancement of satellite images in present study.

4.2.1 Spectral Indices

In present study, Difference Vegetation Index (DVI), Ratio Vegetation Index (RVI), NDVI, Soil Adjusted Vegetation Index (SAVI), Leaf Area Index (LAI), Ratio Difference Vegetation Index (RDVI), Modified SAVI (MSAVI), Infrared Percentage Vegetation Index (IPVI) and Modified Simple Ratio (MSR) (Mróz and Sobieraj, 2004) have been used for change detection analysis. However, NDVI has been widely used for detection of changes in forest cover (Bhagat, 2012). Therefore, NDVI (equation 1) has been calculated using Near Infrared (Band 4) and Red (Band 3) bands of Landsat-7 ETM+ and Landsat-5 TM sensors (Bhagat and Sonawane, 2010).

  \(NDVI = {NIR-RED \over NIR+RED}\)                   (1)

Calculated NDVI values vary between -1 to +1 depending on relative Digital Number (DN) of Near Infrared and Red bands. Maximum NDVI (Figure 2) indicates dense vegetation and minimum values represent less or absence of vegetation. 

Figure 2. Normalized Difference Vegetation Index (NDVI)

Leaf Area Index (LAI) [m2/m2] represents the amount of leaf material in an ecosystem and is geometrically defined as the total one-sided area of photosynthetic tissue per unit ground surface area (Breda, 2003; Jonckheere et al., 2004; Morisette et al., 2006). It appears as a key variable in many models describing vegetation-atmosphere interactions, particularly with respect to the carbon and water cycles (GCOS, 2010). LAI was calculated using following equation after Jensen (2002) (equation 2):

\(LAI = -2.42 +12.18 *(NDVI)\)        (2)

Where, -2.42 and 12.18 are the constants (Jensen, 2002). LAI is widely used to identify and delineate light interception, gross productivity, soil moisture and transpiration from vegetation (Chen et al., 2012). Therefore, LAI were calculated for statistical analysis to test applicability for Forest Change Detection (FCD) in second approach. Furthermore, many scholars have been used Land Surface Temperature Index (LSTI) for vegetation analysis (Yuan and Bauer, 2007). LSTI is measurements of earth surface temperature including  canopy, soils, barren lands, rocks, water bodies, snow, ice, roof of a building using satellite data. The DN recorded for satellite images (Julien et al., 2011) were converted (equation 3) using space reaching radiance i.e. Top of Atmosphere (ToA) Radiance (equation 4) (Chander and Markham, 2003).

\(LST(T) = {{K_2} \over in({k_1\over L_\lambda}+1)}\)                           (3)

Here, \(T\)  is an effective at-satellite temperature in k (kelvin), \(L_\lambda\) is a spectral radiance in W/(m2 in µm), \(K_1\) and \(K_2\) are pre-launch calibration constants for ETM+ and TM sansors.

   \(L_\lambda = ({{L_{max}-L_{min}} \over QCAL_{max}-QCAL_{min}})(DN-QCAL_{min})+L_{min}\)       (4)

Where, \(L_{max}\)  is a maximum spectral radiance (W/m2 in µm) at QCAL equal 0 DN, \(L_{mim}\) is a minimum spectral radiance (W/m2 in µm) at QCAL equal 255 DN and QCAL are the quantized calibrated pixel values in DN.

4.2.2 Tasseled Cap Coefficient Transformations (TCCTs)

TCCTs have potentials to derive forest attributes for different regional applications where, atmospheric noise correction not possible (Silleos et al., 2006; Ghosh et al., 2010). These transformations simply reduce the number of radiance noise density and provide high association in single response. Therefore, brightness, greenness, wetness, fourth, fifth and sixth indices were calculated using coefficients estimated by Crist et al., (1986) for Landsat-5 TM (Table 1) and Huang et al., (2002) (Table 2) for Landsat-7 ETM+. 

Table 1. Landsat-5 TM: tasselled cap coefficients at satellite reflectance

Index

Band1

Band2

Band3

Band4

Band5

Band7

Brightness

0.2909

0.2493

0.4806

0.5568

0.4438

0.1706

Greenness

-0.2728

-0.2174

-0.5508

0.7221

0.0733

-0.1648

Wetness

0.1446

0.1761

0.3322

0.3396

-0.6210

-0.4186

Fourth

0.8461

-0.0731

-0.4640

-0.0032

-0.0492

0.0119

Fifth

0.0549

-0.0232

0.0339

-0.1937

0.4162

-0.7823

Sixth

0.1186

-0.8069

0.4094

0.0571

-0.0228

0.0220

Source: Crist et al., 1986

 

Table 2. Landsat-7 ETM+: tasselled cap coefficients at satellite reflectance

Index

Band1

Band2

Band3

Band4

Band5

Band7

Brightness

0.3561

0.3972

0.3904

0.6966

0.2286

0.1596

Greenness

-0.3344

-0.3544

-0.4556

0.6966

-0.0242

-0.2630

Wetness

0.2626

0.2141

0.0926

0.0656

-0.7629

-0.5388

Fourth

0.0805

-0.0498

0.1950

-0.1327

0.5752

-0.7775

Fifth

-7252

-0.0202

0.6683

0.0631

-0.1494

-0.0274

Sixth

0.4000

-0.8172

0.3832

0.0602

-0.1095

0.0985

Source: Huang et al., 2002

 

4.3   Approaches of Forest Change Detection

4.3.1 Approach I: Post-classification technique

Satellite images selected for present study was processed for co-registration, enhancements, classification and finally, classified images were compared for Change Detection (CD) in forest cover. Calculated pixel values of NDVI have been broadly grouped into five classes (Table 3) i.e. no-vegetation, low to medium, medium, medium to dense and dense to very dense, using threshold observed in NDVI (2002) and NDVI (2009) images using ‘slicing’ operation in Ilwis.

Table 3. Broad classification of forest density based on NDVI

Classes

Index values

ETM+ (2002)

TM (2009)

No-vegetation

< -0.16

< 0.20

Low to medium

-0.16 to -0.02

0.20 to 0.23

Medium

-0.02 to 0.01

0.23 to 0.36

Medium to dense

0.01 to 0.16

0.36 to 0.45

Dense to very dense

0.16 <

0.45 <

 

The maximum value (0.41) in reference image (t1) (NDVI 2002) was observed for dense vegetation (Table 3) whereas minimum (-0.48) for barren land including water body, rocky and barren land with 0.05 mean and 0.16 standard deviation. Targeted image (t2) (NDVI 2009) shows the maximum value (0.71) for dense vegetation and minimum (-0.33) for no-vegetation with 0.31 mean and 0.16 standard deviation. Adegoke and Carleton (2002) have been used innovative hybrid image classification technique for change detection analysis of forest. Therefore, limit values of NDVI classes are different for images acquired for 2002 and 2009 (Table 3) and was decided based on repetitive field checks using GPS, comparison with high-resolution images of Google Earth Pro and FCC (Figure 4). Class ‘no-vegetation’ includes rocky and barren lands distributed at higher levels (> 1100 m) of mountain and water bodies at bottom (Figure 3). 

Figure 3. Distribution of forest depicted based on NDVI

 

Figure 4. False Color Composite (FCC) images

Measurements of differences between detected land classes in base image (t1) and target image (t2) based on radiance difference called change detection (Weng et al., 2009), for instance forest. The radiance difference mainly occurred due to real change in surface features, deviation in atmospheric conditions, deviation in sensor calibration, minimum to maximum error in Route Mean Square (RMS) at the time of georeferencing, deviation in illumination and difference in soil moisture conditions, etc. Here, classified images t1 and t2 combined using ‘cross’ operator (Figure 5) in Ilwis which produce 25 classes (5 classes t1 X 5 classes t2). Similar classes have been merged into three classes viz. no-change, positive change and negative change in forest cover (Table 4) using tool ‘merging the classes’ in Ilwis. Further, class wise changes (Table 11) in forest cover also detected successfully (Figure 14) using same technique (Table 4). However, overall accuracy of the classes detected to show changes in forest cover was estimated about 78%. Users’ accuracy for class ‘positive change’ was about 58% and producer’s accuracy 70%. Rozenstein and Karnieli (2011) have insisted vigorous and creative efforts to establish new algorithms for change detection. Therefore, statistical methods and corrective measures were adopted to achieve improvements in post-classification technique for CD in forest cover. 

Figure 5. Approach-I: schematic preparation

Table 4. Class merging scheme for overall FCD

Vegetation classes: ETM+ (2002)

Vegetation classes: TM (2009)

No-vegetation

Low to medium

Medium

Medium to dense

Dense to very dense

No-vegetation

No-change

Positive

change

Positive

change

Positive

change

Positive

change

Low to medium

Negative

change

No-change

Positive

change

Positive

change

Positive

change

Medium

Negative

change

Negative

change

No-change

Positive

change

Positive

change

Medium to dense

Negative

change

Negative

change

Negative

change

No-change

Positive

change

Dense to very dense

Negative

change

Negative

change

Negative

change

Negative

change

No-change

 

Table 5. Class merging scheme for class wise FCD

Vegetation classes:

ETM+ (2002)

Vegetation classes: TM (2009)

No- vegetation

Low to medium

Medium

Medium to dense

Dense to very dense

No-vegetation

NVNoC

NVPL

NVPM

NVPD

NVPVD

Low to medium

LNNV

LNoC

LPM

LPD

LPVD

Medium

MNNV

MNL

MNoC

MPD

MPVD

Medium to dense

DNNV

DNL

DNM

DNoC

DPVD

Dense to very dense

VDNNV

VDNL

VDNM

VDND

VDNoC

* Refer section- Abbreviations or Table  11 and  Table 13 for explanation.

 

4.3.2  Approach II: Improved post-classification technique

Many scholars have been successfully used combinations of different approaches for CD of land objects (Cano et al., 2006; Pu et al., 2008; Weng et al., 2009). The algorithm of FCD has been modified to improve (Figure 6) the results and processed through five steps: 1) sampling (Chapelle et al., 2002; Lu et al., 2011), 2) statistical analysis (Singh, 1989; Dhakal et al., 2002), 3) normalization of indices, 4) change detection and 5) accuracy assessment. Statistical analysis and information about stable land surface including deep forests, rocky lands, barren lands and water bodies have used for normalization of indices calculated for target image (t2) (Felkar et al., 1981; Singh, 1996; Turker and Derenyi, 2000). 

Figure 6. Approach-II: schematic preparation

a. Sampling

Many researchers have been successfully used single band, ratio indices and coefficients for detection and delineation (Chen et al., 2012) of objects like forests, water bodies, barren lands, rocky lands, snow covers, marsh lands, desert lands, etc. The variations in land phenomenon regulate surface radiance according to their reflectivity (Carlson et al., 1990; Chen et al., 2012). Surface reflectance values are sensitive to variations in vegetation cover, soil, background temperature, thermal inertia, and topography (Bhagat 2012). Goetz (1997) has used inverse relationship between SVI and Ts based on inversion soil-vegetation-atmosphere-transfer (SVAT) model for estimating soil surface moisture using NDVI-Ts triangle space. This approach is defined as the ‘Triangle Method’ includes vegetation indices, soil moisture indices and thermal indices (Weng et al., 2009; Lu et al., 2011). Objects detected using this approach are stood more precise than single bands. Surface temperature can be estimated using thermal band and vegetation can be detected with the help of near infrared band (Nemani and Running, 1989; Carlosan et al., 1990; Nemani et al., 1993; Tanriverdi, 2010). Thus, all surface elements have equal importance in estimations of objects (Owen et al., 1998).

Water body absorbs maximum amount of radiation and emits very little energy (Nemani et al., 1993). Forest cover reflects maximum amount of Green and Near Infrared (NIR) bands and a good absorber of Blue and Red band. Chemical component existing in tree leaves known as ‘chlorophyll’ effectively absorbs radiation from the Blue and Red band (Jiany et al., 2008). The inner structure of these leaves response as good emission of NIR wavelength (Roberts et al., 2011). On the other hand, visible and NIR radiation absorb more in the form of longer wavelength from water body and less from shorter wavelength (Lu et al., 2011). Thus, water appears blue-green or blue due to maximum reflectance of shorter wavelength and viewed dark because reflectance of red and near infrared depending on the suspended sediments amount in water (Fletcher and Everitt, 2007). Rocky lands are purely dry and rarely covered by thin grass and algae in wet seasons. However, rocky lands absorb very low amount of visible as well as NIR waves and reflect maximum as compare to other classes (Majed et al., 2011). Water bodies, deep forests and rocky lands are stable objects and have extreme reflectance (minimum and maximum) in the region.

The variations in reflectance recorded is satellite image have influence on accuracy of forest CD using post-classification (Pu et al., 2008). Therefore, corrective approach was adopted for detection of FCD (discussed in next section). Ground truth information (170 samples) was collected for these stable land objects which were appeared and identified on satellite images e.g. Landsat-7 ETM+ (2002) and Landsat-5 TM (2009). DN values of these land objects were used further statistical analysis and normalisation of indices calculated using satellite data (t2) (Bhagat, 2012). About 31.18% samples were collected from forest cover, 27.65% from rocky land, 29.41% from water body and 11.76% from barren land for further process (Figure 7). 

Figure 7. Ground Reference Triangle

b.  Statistical analysis

NDVI has been widely used for vegetation analysis, however, greenness (equation 5 and 6) (Crist et al., 1986; Huang et al., 2002) have more potential of vegetation analysis. Greenness values estimated for both images (t1 and t2) have positively correlated with NDVI. Therefore, greenness estimated for ETM+ (t1) and TM (t2) data was used as dependent variable for estimations corrected greenness values for t2 (Figure 8). 

Figure 8. Greenness Index

Greenness (ETM+) = ((Band1 * (-0.3344)) + (Band2 * (-0.3544)) + (Band3 * (-0.4556)) + (Band4 * (0.6966)) + (Band5 *(-0.0242)) + (Band7* (-0.2630))  (5)

Greenness (TM) = ((Band1*(-0.2728)) + (Band2*(-0.2174))+ (Band3 * (-0.5508)) + (Band4 * (0.7221)) + (Band5 *(0.0733)) + (Band7 * (-0.1648))    (6)

NDVI (2002) values show positive correlation with greenness estimated for t2 (2009) (r = 0.976) and NDVI (2009) with greenness for t2 (2009) (r = 0.975) (Table 6). Scatterplot (Figure 9) for NDVI (2002) and Greenness (2009) show clustered distribution of vegetation. Therefore, calculated greenness values used for estimations of dependent variable in regression analysis. Maximum greenness (-0.07) calculated for t1 (2002) and 65.06 for t2 (2009) whereas minimum (-110.16) for no-vegetation i.e. water, rocky land, shadow, etc. in t1 (2002) and -36.57 in t2 (2009). These are stable land surface in the study area. Theoretically, values estimated for the stable lands might be same which is not possible in certain conditions at the time of image capturing (Munyati, 2004; Yarbrough et al., 2005; Gill et al., 2012). Therefore, accuracy of FCD in the region shows less. Further, these estimated values have been normalized using values estimated for stable pixels in this approach. Differences between estimated greenness values for stable pixels of t1 and t2 have been calculated and correlation with estimated indices SWI (t1), brightness (t1), fifth (t1) and DN values of band 7 (t1), band 7 (t2), band 5 (t1) have been estimated to find independent variables for estimation of normalized dependent variable for t2 e.g. greenness. 

Figure 9. Scatterplot: NDVI (2002) and Greenness (2009)

Table 6. Correlations between NDVI and Greenness

 

Greenness (2002)

Greenness (2009)

NDVI (2002)

Pearson correlation

0.684**

0.976**

Sig. (2-tailed)

.000

.000

N

170

170

NDVI (2009)

Pearson correlation

0.571**

0.975**

Sig. (2-tailed)

.000

.000

N

170

170

** Significance at the 0.01 level (2-tailed).

 

Significant correlation of calculated difference between greenness estimated for stable pixels of t1 and t2 was estimated with band 7 (t1), SWI (t1), band 5 (t1) band 7 (t2), brightness (t1), difference in SWI, difference in fifth (Table 7). However, the strongest correlation was estimated for band 7 of t1 and therefore this band was used as independent variable for estimations of pixel values of correction in t2 using regression techniques. Samples (170) collected from stable land surface i.e. forest land, rocky land, water, barren land, etc. (Figure 7) was used for this analysis. 

Table 7. Correlation of difference greenness with selected criterion

 

Band7 of t1

SWI

of t1

Band7

of t2

Band5

of t1

Brightness of t1

Difference

SWI

Difference

Fifth

Difference greenness

Pearson Correlation

0.983**

-0.964**

0.960**

0.960**

0.949**

0.934**

-0.933**

Sig.

(2-tailed)

0.000

0.000

0.000

0.000

0.000

0.000

0.000

N

170

170

170

170

170

170

170

** Significance at the 0.01 level (2-tailed).

 

Regression (equation 7) gives mathematical tools (Lawrence and Wrlght, 2001; Cohen et al., 2003) to estimate the fit between two different multi-spectral images acquired at different time for same area. Pixel values close to the regression line indicate the suitability of the model (Table 8).

  \(Y=a+b(x)\)                 (7)

Where, \(Y\)  is an estimated variable (dependent), \(a\)  and \(b\) are constant ( \(a\)  is intercept and \(b\)  is rate of change in slope) and \(x\)  is independent variable.

Table 8. Regression analysis of difference greenness with selected criterion

 

\(Y\)

\(x\)

Correlation (r)

\(a\)

\(b\)

R2

Trend

Difference greenness

Band7 of t1

0.983

21.664

0.602

0.967

Positive

SWI of t1

-0.964

33.511

-0.432

0.930

Negative

Band7 of t2

0.96

24.208

0.891

0.922

Positive

Band5 of t1

0.96

20.554

0.399

0.921

Positive

Brightness of t1

0.949

-2.307

0.370

0.901

Positive

Difference SWI

0.934

30.447

0.769

0.871

Positive

Difference Fifth

-0.933

90.231

-2.110

0.871

Negative

 

 Scholars have been reported the highest change detection accuracy by using regression approach. Jiany et al., (2008) reported three steps of radiometric normalization using pixel values for stable land surface: (1) delineation of stable land based on near-infrared (NIR) bands; (2) DN for stable areas using regression models, and (3) the regression coefficients used for normalizing DN of the target image. Here, linear regression model depicts (Figure 10) the fit (R2= 0.967) of differences in Greenness of t1 and t2 with band 7 of t1 with high correlation (0.983) and clustered distribution. The values closer to zero on x-axis are for water bodies, next cluster for forests and at far from the zero for barren and rocky lands. Band 5 and band 7 of ETM+ and TM images show higher sensitivity to soils and vegetation.

Figure 10. Regression analysis

c.   Estimation of exaggeration and normalization of images

The maximum value (0.41) in reference image (t1) (NDVI 2002) was observed for dense vegetation (Table 3) whereas minimum (-0.48) for barren land and water body with 0.05 mean and 0.16 standard deviation. Maximum NDVI value (0.71) in target image, (t2) (2009) was observed for dense vegetation and minimum (-0.33) for no-vegetation with 0.31 mean and 0.16 standard deviation. Therefore, it is assumed that the values recorded in target image are exaggerated from ground reality (Figure 8). Regression model (equation 8) was formed to estimate the values of exaggeration in greenness (Ge) estimated for target image (t2) for normalizing (t2).

\(Ge=a+b(B7t_1)\)                                (8)

Where, \(a\)  is an intercept 21.664 and \(b\)  is rate of change in slope 0.602 and \(B7 t_1\)  is band 7 of reference image (t1). Finally, raster image for greenness estimated for target image (t2) was normalised \(Nt_2\)  (equation 9) for final FCD.

\(Nt_2=Greenness2009-Ge\)                    (9)

The histograms for estimated images viz. corrected greenness 2009 using band 7 (t1), SWI (t1), band 7 (t2) and band 5 (t1), etc. were prepared in SPSS and compared with histograms prepared for greenness for t1 and t2 (Figure 11). Histogram values distributed from -110.1568 to -0.0782 for greenness of ETM+ (2002) and from -36.5670 to 65.0585 of TM (2009). However, distributed values on histogram prepared for corrected greenness of t2 (2009) vary from -108.6296 to 17.5085. Similar results observed for normalised raster images of band 7 (2002), SWI (2002), band 7 (2009) and band 5 (2002). 

Figure 11. Histogram distributions

4.3.3  Change detection

Finally, raster image of greenness estimated for t1 (ETM+ 2002) and normalised raster image of greenness estimated (Figure 12) for t2 (TM 2009) were processed for FCD in this study. The maximum value was -00.08 estimated for greenness of t1 (2002) and 17.51 for normalized greenness (2009) whereas minimum value was -110.16 for t1 (2002) and -108.63 for normalized image t2 for ‘no-vegetation’ i.e. water, rocky land, shadow, etc. (Figure 13). Reference (t1) and normalised (Nt2) images were classified into five classes (Table 9) e.g. no-vegetation, low to medium, medium, medium to dense and dense to very dense, respectively.

Table 9. Broad classification of forest density based greenness index

Vegetation classes

Index value

No-vegetation

< -64

Low to medium

-64 to -48

Medium

-48 to -32

Medium to dense

-32 to -16

Dense to very dense

-16 <

 

Figure 12. Corrected Greenness index

Classified raster images (Figure 13) for 2002 and 2009 were combined using cross operation in Ilwis for forest CD. 25 classes were created in this combination and similar classes merged to get meaningful categories (Table 4). Overall changes were estimated as ‘negative changes’, ‘positive changes’ and ‘no-changes’. Further, class wise changes also estimated to know the transformations from one class to another (Table 5) as ‘positive change, negative change and no-change.

Figure 13. Distribution of forests depicted based on greenness index

5 . RESULTS

Changes in forest cover in the study area (13858.83 ha) was estimated using two approaches: Approach I) post-classification technique and Approach II) improved post-classification technique. NDVI based classified maps prepared for t1 and t2 were used for CD (Figure 14) in first approach and statistical analyses were performed for improvement in accuracy of post-classification CD in second approach (Figure 15). Primary field check data and pixel information of stable land units i.e. water bodies, deep forest and rocky lands were used in statistical analysis for improvement of the technique.

5.1  Approach I: Post-classification Technique

The change in forest area within period of images acquired has been detected and successfully demarcated (Figure 14). Approximately, 58.59% of reviewed area shows forests with increasing trends, 33.69% area shows no-changes and 7.72% area losing the forest. Forest cover increases for the class ‘medium’ vegetation about 22.41% and class ‘dense to very dense’ 16.84% (Table 10). Slightly negative change (-0.12%) observed for class ‘no-vegetation’ and the class ‘medium to dense’ (-2.81%) (Figure 14). Positive changes in forest cover was estimated for 8119.98 ha, no-change for 4669.29 ha and negative change for 1069.56 ha. However, these estimated values show generalized picture of changes within the classes therefore, dynamics of class wise changes also analyzed. 

Figure 14. Approach-I: distribution of changes in forest cover

Table 10. Approach I- class wise changes in area under forest

Vegetation classes

Area (2002)

Area (2009)

Change

%

ha

%

ha

%

No-vegetation

1713.51

12.36

1695.96

12.24

-0.12

Low to medium

5840.46

42.14

806.76

5.82

-36.32

Medium

968.4

6.99

4075.11

29.4

22.41

Medium to dense

3706.65

26.75

3317.76

23.94

-2.81

Dense to very dense

1629.81

11.76

3963.24

28.6

16.84

Total Area

13858.8

100

13858.8

100

 

 

Sparse vegetation has been changed to medium vegetation (22.68%) and dense vegetation to very dense vegetation (15.02%) (Table 11) with no-change on 10.89 % area. However, vegetation decreases with high rate from ‘low vegetation’ to ‘no-vegetation’ for 4.30% area. The accuracy of assessment has been estimated about 77.84%, discussed in the next section (Table 14).

Table 11. Approach I – change class wise distribution of forest

Class Code

Change classes of forest

Area in ha

Area in %

NVNoC

No-vegetation to no-change

1083.96

7.8214

NVPL

No-vegetation positively changed to low vegetation

259.92

1.8755

NVPM

No-vegetation positively changed to medium vegetation

347.31

2.5061

NVPD

No-vegetation positively changed to dense vegetation

21.24

0.1533

NVPVD

No-vegetation positive changed to very dense vegetation

1.08

0.0078

LNNV

Low vegetation negatively changed to no-vegetation

597.24

4.3095

LNoC

Low vegetation with no-change

531.09

3.8321

LPM

Low vegetation positively changed to medium vegetation

3143.34

22.6811

LPD

Low vegetation positively changed to dense vegetation

1393.38

10.0541

LPVD

Low vegetation positive changed to very dense vegetation

175.41

1.2657

MNNV

Medium vegetation negative changed to no-vegetation

9.54

0.0688

MNL

Medium vegetation negatively changed to low vegetation

8.19

0.0591

MNoC

Medium vegetation with no-change

253.98

1.8326

MPD

Medium vegetation positively changed to dense vegetation

500.94

3.6146

MPVD

Medium vegetation positively changed to very dense vegetation

195.75

1.4125

DNNV

Dense vegetation negatively changed to no-vegetation

5.22

0.0377

DNL

Dense vegetation negatively changed to low vegetation

7.47

0.0539

DNM

Dense vegetation negatively changed to medium vegetation

321.48

2.3197

DNoC

Dense vegetation with no-change

1290.87

9.3144

DPVD

Dense vegetation positively changed to very dense vegetation

2081.61

15.0201

VDNNV

Very dense vegetation negatively changed to no-vegetation

0.02

0.0002

VDNL

Very dense vegetation negatively changed to low vegetation

0.07

0.0004

VDNM

Very dense vegetation negatively changed to medium vegetation

9

0.0649

VDND

Very dense vegetation negatively changed to dense vegetation

111.33

0.8033

VDNoC

Very dense vegetation with no-change

1509.39

10.8912

Total area

13858.83

100

5.2 Approach II: Improved Post-classification Technique

Statistical techniques i.e. correlation analysis and regression model have been used to improve the performance of FCD (Figure 6). Area under forest has been detected and estimated using TCCT greenness index. However, estimated greenness has been corrected to reduce exaggeration in reflectance recorded in DN using regression model (Figure 15). About 13858.83 ha (23.59% TGA) shows positive changes, 70.8% no-change and only 6.33% area shows negative changes. Accuracy assessment shows preciseness 95.21% (Table 15) of the results compared to post-classification based FCD. ‘No-changes’ in vegetation observed for the class, ‘low to medium’ vegetation (Table 12) and positive changes for ‘medium to dense’ and ‘dense to very dense’ vegetation. This improved technique of FCD helps to reduce the exaggeration in estimations. Class wise changes also estimated to understand the dimensions of changes in forest cover (Table 13). 

Figure 15. Approach-II: distribution of changes in forest cover

Table 12. Approach II - changes in forest cover

Vegetation classes

Area (2002)

Area (2009)

Change

 

ha

%

ha

%

%

No-vegetation

3310.2

23.89

2975.22

21.47

-2.42

Low to medium

4228.47

30.51

3505.68

25.3

-5.21

Medium

4115.97

29.7

4251.24

30.68

0.98

Medium to dense

2083.95

15.04

2908.71

20.99

5.95

Dense to very dense

120.24

0.87

217.98

1.57

0.7

Total Area

13858.83

100

13858.83

100

 

 

Table 13. Approach II – change class wise distribution of forest

Class code

Change classes of forest

Area in ha

Area in %

NVNoC

No-vegetation to no-change

1931.31

13.94

NVPL

No-vegetation positively changed to low vegetation

1176.48

8.49

NVPM

No-vegetation positively changed to medium vegetation

196.02

1.41

NVPD

No-vegetation positively changed to dense vegetation

6.3

0.05

NVPVD

No-vegetation positive changed to very dense vegetation

0.09

0.01

LNNV

Low vegetation negative change to no-vegetation

338.67

2.44

LNoC

Low vegetation negatively changed to no-vegetation

1838.99

13.27

LPM

Low vegetation with no-change

1919.72

13.85

LPD

Low vegetation positively changed to medium vegetation

168.75

1.22

LPVD

Low vegetation positively changed to dense vegetation

2.34

0.02

MNNV

Medium vegetation negative changed to no-vegetation

8.73

0.06

MNL

Medium vegetation negatively changed to low vegetation

317.43

2.29

MNoC

Medium vegetation with no-change

2262.79

16.33

MPD

Medium vegetation positively changed to dense vegetation

1511.28

10.9

MPVD

Medium vegetation positively changed to very dense vegetation

27.73

0.2

DNNV

Dense vegetation negatively changed to no-vegetation

0.34

0.01

DNL

Dense vegetation negatively changed to low vegetation

16.54

0.12

DNM

Dense vegetation negatively changed to medium vegetation

401.13

2.89

DNoC

Dense vegetation with no-change

1466.17

10.58

DPVD

Dense vegetation positively changed to very dense vegetation

98.73

0.71

VDNNV

Very dense vegetation negatively changed to no vegetation

0.17

0.01

VDNL

Very dense vegetation negatively changed to low vegetation

6.94

0.05

VDNM

Very dense vegetation negatively changed to medium vegetation

73.92

0.53

VDND

Very dense vegetation negatively changed to dense vegetation

31

0.22

VDNoC

Very dense vegetation with no-change

57.26

0.41

Total area

13858.83

100

 

6 . ACCURACY ASSESSMENT

Accuracy of detected changes in forest cover was estimated as user’s accuracy, producer’s accuracy and overall accuracy. Rahman and Saha (2008) have suggested that the sample size should take minimum 30 for each class to estimate accuracy at 90%. Therefore, 170 samples well distributed within the classes were collected using GPS points and Google Earth Pro high resolution images.

Table 14. Approach-I: error matrix

Classified data

Reference data

Positive change

No-change

Negative change

Water

Row total

User’s

accuracy in (%)

Positive change

19

6

4

4

33

58

No-change

5

31

5

3

44

70

Negative change

2

3

32

3

40

80

Water

1

0

1

48

50

96

Column total

27

40

42

58

167

 

Producer's

accuracy (%)

70

78

76

83

   

Over accuracy

77.84

 

 

 

 

 

 

Overall accuracy of the FCD using approach I: post-classification technique was estimated about 77.84% (Table 14). User’s accuracy was estimated about 58% for positively changed forest cover, 70% for no-changed or stable and 80% for negatively changed forest and 96% for water bodies. Producer’s accuracy was estimated about 70% for positive changes, 78% for no-changes, 76% for negative changes and 83% for water bodies. Maximum accuracy was estimated for water bodies and negative changes in vegetation whereas positively changed vegetated area and stable land units show very less accuracy. Accuracy assessment suggests need of improvement in FCD techniques. Therefore, algorithm for improved FCD has been formulated in this study. Approach II has improved accuracy of FCD to 95.21% (Table 15) with user’s accuracy about 97% for positive changes, 95% for no-change, 93% for negative changes and 96% for water bodies. Producer’s accuracy was estimated about 91% for positive changes, 98% for no-change, 95% for negative changes and 96% water bodies. Thus, improved post-classification technique improves the accuracy of FCD.

Table 15. Approach-II: error matrix

Classified data

Reference data

Positive change

No-

 change

Negative change

Water

Row total

User’s

accuracy (%)

Positive change

32

0

0

1

33

97

No-change

1

42

1

0

44

95

Negative change

1

1

37

1

40

93

Water

1

0

1

48

50

96

Column total

35

43

39

50

167

 

Producer's

accuracy in (%)

91

98

95

96

 

 

Over accuracy

95.21%

 

 

 

 

 

 

7 . CONCLUSION

About 58.59% of reviewed area estimated using first approach of CD shows (Figure 15) positive changes, 33.69% no-change and 7.72% negative changes in forest cover in the area. However, improved post-classification analysis estimates (Table 16) about 23.59% area with positive changes, 70.08% no-change and 6.33% negative changes. Thus, normalised raster greenness image using statistical analysis removes illusional exaggeration of reflectance recorded in satellite data and useful to estimate more precisely FCD.

Table 16. Comparison of changes in forest cover estimated using two approaches

Change classes

Approach I

Approach II

Area (hectare)

Area (%)

Area (hectare)

Area (%)

Positive change

8119.98

58.59%

3269.25

23.59%

No-change

4669.29

33.69%

9712.89

70.08%

Negative change

1069.56

7.72%

876.69

6.33%

Total area

13858.83

100%

13858.83

100%

 

Authors are aware about limitations of finer temporal resolution (7 years) and medium spatial resolution (30 x 30m) of Landsat ETM+ and Landsat TM images. The methodology and techniques formulated in this study can be useful for researchers, scholars, environmental planners and managements for change detection of forest for sustainable land management and development.

Conflict of Interest

The authors declare no conflict of interest.

Acknowledgements

Authors gratefully acknowledge the anonymous reviewers for constructive comments and suggestions for improvement in the draft.

Abbreviations

CD: Change Detection; DCD: Digital Change Detection; DN: Digital Numbers, DNL: Dense vegetation negatively changed to low vegetation; DNM: Dense vegetation negatively changed to medium vegetation; DNNV: Dense vegetation negatively changed to no-vegetation; DNoC: Dense vegetation with no-change; DPVD: Dense vegetation positively changed to very dense vegetation; DVI: Difference Vegetation Index; ERDAS: Earth Resource Data Analysis System: ETM+: Enhanced Thematic Mapper Plus; FAO: Food and Agriculture Organization; FCC: False Colour Composite; FCD: Forest Change Detection; FSI: Forest Survey of India; GIS: Geographical Information System; GPS: Global Positioning System; ILWIS: Integrated Land Water Information System; IPVI: Infrared Percentage Vegetation Index; ISFR: India State of Forest Report; LAI: Leaf Area Indices; LNNV: Low vegetation negatively changed to no-vegetation; LNoC: Low vegetation with no-change; LPD: Low vegetation positively changed to dense vegetation; LPM: Low vegetation positively changed to medium vegetation; LPVD: Low vegetation positive changed to very dense vegetation; LSTI: Land Surface Temperature Index; MNL: Medium vegetation negatively changed to low vegetation; MNNV: Medium vegetation negative changed to no-vegetation; MNoC: Medium vegetation with no-change; MPD: Medium vegetation positively changed to dense vegetation; MPVD: Medium vegetation positively changed to very dense vegetation; MSAVI: Modified SAVI; MSL: Mean Sea Level; MSR: Modified Simple Ratio; NASA: National Aeronautics and Space Administration; NDSI: Normalised Difference Salinity Index; NDVI: Normalized Difference Vegetation Index; NIR: Near Infrared; NVNoC; No-vegetation to no-change; NVPD: No-vegetation positively changed to dense vegetation; NVPL: No-vegetation positively changed to low vegetation; NVPM: No-vegetation positively changed to medium vegetation; NVPVD: No-vegetation positive changed to very dense vegetation; RDVI: Ratio Difference Vegetation Index; RMS: Route Mean Square; RVI: Ratio Vegetation Index; SAVI: Soil Adjusted Vegetation Index; SOI: Survey of India; SPSS: Statistical Packages for the Social Sciences; SVAT: Soil Vegetation Atmosphere Transfer; SVI: Soil Vegetation Index; SWI: Soil Wetness Index; TCCT: Tasseled Cap Coefficient Transformation; TGA: Total Geographical Area; ToA: Top of Atmosphere; TM: Thematic Mapper; TVI: Transformed Vegetation Index; VDND: Very dense vegetation negatively changed to dense vegetation; VDNL: Very dense vegetation negatively changed to low vegetation; VDNM: Very dense vegetation negatively changed to medium vegetation; VDNNV: Very dense vegetation negatively changed to no vegetation; VDNoC: Very dense vegetation with no-change; WCMC: World Conservation Monitoring Centre; WRI: World Resources Institute.

References

10.

Dhakal A. S., Amada T., Aniya M. and Sharm R. R., 2002. Detection of areas associated with flood and erosion caused by a heavy rainfall using multitemporal Landsat TM data, Photogram. Eng. Remote Sens., 68, 233-239.

33.

Forkuo E. K. and Frimpong A., 2012. Analysis of Forest Cover Change Detection. International Journal of Remote Sensing Applications, 2 (4), 82-92.

70.

Pajares G., Alonso C., Cruz J. M. and Moreno V., 2002. Shade identification in urban areas through the bayesian classifier, Proc. European Symposium on Intelligent Technologies, 476-481.