Environ Monit Assess (2015)87:4 DOI 10.1007/s10661-015-4428-3
Assessing and monitoring the risk of desertification in Dobrogea, Romania, using Landsat data and decision tree classifier Iosif Vorovencii
Received: 3 June 2014 / Accepted: 12 March 2015 # Springer International Publishing Switzerland 2015
Abstract The risk of the desertification of a part of Romania is increasingly evident, constituting a serious problem for the environment and the society. This article attempts to assess and monitor the risk of desertification in Dobrogea using Landsat Thematic Mapper (TM) satellite images acquired in 1987, 1994, 2000, 2007 and 2011. In order to assess the risk of desertification, we used as indicators the Modified Soil Adjustment Vegetation Index 1 (MSAVI1), the Moving Standard Deviation Index (MSDI) and the albedo, indices relating to the vegetation conditions, the landscape pattern and micrometeorology. The decision tree classifier (DTC) was also used on the basis of pre-established rules, and maps displaying six grades of desertification risk were obtained: non, very low, low, medium, high and severe. Land surface temperature (LST) was also used for the analysis. The results indicate that, according to preestablished rules for the period of 1987–2011, there are two grades of desertification risk that have an ascending trend in Dobrogea, namely very low and medium desertification. An investigation into the causes of the desertification risk revealed that high temperature is the main factor, accompanied by the destruction of forest shelterbelts and of the irrigation system and, to a
Iosif Vorovencii holds a Ph.D. degree. I. Vorovencii (*) Forest Engineering, Forest Management Planning and Terrestrial Measurements Department, Faculty of Silviculture, Transilvania University of Brasov, Beethoven street nr. 1, 500123 Brasov, Romania e-mail: [email protected]
smaller extent, by the fragmentation of agricultural land and the deforestation in the study area. Keywords Desertification . Decision tree classifier . MSAVI1 . MSDI . Albedo
Introduction Desertification is a general degradation process that occurs on the surface of land because of the combined effects of human and natural causes that lead to decrease in, or loss of, the soil’s capacity to support the growth of plants (Helldén 1991). Desertification is associated with long-term changes in the functions of ecosystems (Dregne 1977), involving both the spatial and the temporal component and leading to a reduced productivity and biodiversity (Mouat et al. 1997). The surfaces affected by desertification are characterised by scarce precipitation and intense evapotranspiration which lead to sparse growth of vegetation and the occurrence of groups of simple structure plants. Over the last 2 decades, remote sensing was widely used in research concerning the assessment of desertification (Sun et al. 2005; Yan et al. 2009; Xu et al. 2009; Santini et al. 2010; Xu et al. 2010; Pan and Li 2013). The types of satellite images frequently involved in monitoring and assessing the risk of desertification are Landsat Multispectral Scanner (MSS), Landsat Thematic Mapper (TM), Landsat Enhanced Thematic Mapper Plus (ETM+) (Liu et al. 2008; Zhang et al. 2008; Xu et al. 2009; Pan and Li 2013), NOAA-
204
Environ Monit Assess402:781 )5 02(
Page 2 of 17
AVHRR (Arnab and Dipanwita 2011) and SPOT Vegetation (Huang and Siegert 2006). Most of the methods used in remote sensing assess and monitor the desertification risk using vegetation indices related to vegetation coverage (Chopping et al. 2008; Hanafi and Jauffret 2008; Xue et al. 2009). One such vegetation index frequently used in desertification studies is the Normalised Difference Vegetation Index (NDVI). In the semi-arid and arid regions, because of the sparse vegetation, the soil and its humidity have a significant impact on the NDVI (Wessels et al. 2004; Huang and Siegert 2006; Hill et al. 2008; Pan and Li 2013). In order to diminish the influence of the background of soil on the results, certain researchers have opted for the Modified Soil Adjustment Vegetation Index (MSAVI) or the Enhanced Vegetation Index (EVI) (Wu et al. 2007; Li et al. 2010). The assessment of the desertification risk using a single vegetation index (NDVI, MSAVI, EVI) is unilateral and fails to offer a comprehensive view of all the aspects of desertification (Pan and Li 2013). Certain studies (Xu et al. 2009) adopted quantitative analysis and monitored desertification by establishing quantitative relationships between the desertification process and biophysical features of the land surface such as surface albedo, land surface temperature (LST) and moving window. The surface albedo retrieved from remote sensing data is a physical parameter that renders reflective features of a surface and solar shortwave radiation. It is affected by vegetation coverage, land use, soil humidity, snow cover and other conditions of the land surface (Vorovencii 2014). Any change in the features of a land surface leads to a visible change in the surface albedo. The novelty of this article consists in assessing, for the first time, the risk of desertification in Dobrogea, Romania, using Landsat TM satellite images and the decision tree classifier (DTC). We have also proposed our own reference scheme for zoning the surfaces in the study area by grades of desertification risk and for analysing the causes. Knowing the expansion and severity of the desertification process is an important factor in the decisionmaking process for an effective control; this is the reason why the desertification phenomenon must be assessed and monitored. The general objective of this paper is to assess and monitor the desertification risk in Dobrogea, Romania, in the period of 1987–2011, using Landsat TM images. The specific objectives of the
research are the following: (1) to identify certain indicators and to develop a decision model using the decision tree classifier (DTC) in order to assess and monitor the desertification risk; (2) to use the indicators MSAV I1, Moving Standard Deviation Index (MSDI), the albedo, and their combinations in assessing the various grades of desertification risk; and (3) to analyse the causes that led to the desertification risk, including the LST analysis of the desertification phenomenon.
Materials and methods Study area The study area is the Dobrogea Region of Romania, including the counties of Constanta and Tulcea, and excluding the Danube Delta (Fig. 1). The total surface taken into consideration in the study is 1,176,582 ha. Out of 707,100 ha in Constanta County, 564,500 ha represents agricultural land and in Tulcea County; out of 849,900 ha, 362,218 ha represents agricultural land. The altitude is relatively low (200–300 m). In the southern part, the altitude is less than 200 m (with a maximum of 204 m in Deliorman Plateau), and in the northern part, the altitude reaches 350–400 m and even 467 m (Greci Peak in the Macinului Mountains). The most common climate in Dobrogea is an arid climate, with high average temperatures (10–11 °C), very high temperatures in summer (22–23 °C), scarce precipitation (on average under 400 mm/year), tropical days and frequent droughts. The region closer to the Black Sea is characterised by a more moderate climate, diurnal sea breezes and strong insolation. At altitudes exceeding 300 m, there is a low hill climate, with a lower average temperature (9–10 °C) and more abundant precipitation (500–600 mm/year). The average temperature of the coldest month (January) is, for most of the area, from −1 to −2 °C, and in the south-eastern part, it is positive. The annual thermal amplitude varies from 23 to 24 °C in the western part of Dobrogea to 21– 22 °C in the eastern part. Vegetation is varied: In the south, the vegetation is dry steppe covered by grass, which has been replaced mostly by cultivated plants; in the north, on the hills and in the low altitude mountains, the vegetation is silvosteppe combined with forest vegetation made up of red oak (Quercus rubra) and Hungarian oak (Quercus fraineto). The surface area covered by forests represents
Environ Monit Assess (2015)87:4
Page 3 of 17 204
Fig. 1 Study area. The studied surface is the Dobrogea region without the Danube Delta. In the eastern part of the study area (dashed line), there are the following lakes: Razim, Sinoe, Golovita, Zmeica, Histria, Nuntasi, Tuzla, Ceamurlia, Caraburum and Salcioara
only 3.5 % (24,750 ha) of Constanta County and 10.8 % (91,790 ha) of Tulcea County, as compared to the 26.7 %, average for Romania. The soils, which cover most of the surface area, belong to the class of chernozems, cambic chernozems and grey soils. A smaller surface area (around 5 %) is covered by brown luvic soils, rendzina, litosoils, erodosoils and regosol. Landsat and ancillary data The study used five frames clipped from Landsat 5 TM satellite images acquired on 14 July 1987, 01 July 1994, 01 July 2000, 21 July 2007 and 16 July 2011. The images are part of path/row 181/29 and were downloaded from the Internet (http://glovis.usgs.gov). They have a 1T level of correction and were georeferenced at Universal Transverse Mercator (UTM), datum WGS 84, zone 35 N. The spatial resolution of the multi-spectral and thermal bands, after resampling,
is 30 m. For an accurate assessment of the images obtained after processing and recoding, the following were used: the soils map (1:1,000,000), the vegetation map (1:100,000), the radiation distribution map (1:100,000), the land erosion map (1:1,000,000) and the drought risk map (1:1,000,000). Meteorological data, including monthly and annual averaged temperature and rainfall, of eight meteorological stations in and around Dobrogea, were collected from different reports (Bosneagu et al. 2010) and from the Romanian Meteorological Institute. Social and economic census data of Dobrogea was collected from local statistical yearbooks. Image pre-processing Relative radiometric normalisation (RRN) was carried out for all these images, except the 1994 image which represented the reference image. The RRN method used
204
Environ Monit Assess402:781 )5 02(
Page 4 of 17
was the Automatic Scattergram-Controlled Regression (ASCR) proposed by Eldvige et al. (1995). This method is based on the Scattergram-Controlled Regression (SCR) developed by Yuan and Elvidge (1993), but it contains some improvements concerning automation of the procedure. Application of the ASCR consisted in identifying the Bno-change^ pixel set, comprised of those pixels that occupy the core of the water and land data clusters visible in red near-infrared scattergrams. On the basis of the identified no-change pixels, we were able to calculate, by regression, the gains and offsets used in the normalisation of the subject images in order to match the reference image. The procedure is described in detail in Eldvige et al. (1995). Cloud removal was performed for the 1987 and 2011 images. Indicators of the desertification risk The changes resulting from desertification are mainly characterised by changes of vegetation biomass or cover, landscape pattern and micrometeorological conditions (Xu et al. 2009). In assessing vegetation biomass or land cover, various vegetation indices can be employed, such as NDVI, MSAVI and EVI. Taking into consideration the strong relationships among these indices, in the present study, the MSAVI1 index was chosen in order to assess land cover. MSAVI1 was shown to operate better than NDVI, SAVI and WDVI (Qi et al. 1994). In order to characterise the landscape pattern, MSDI was used, and for assessment of the micrometeorological conditions of the land surface, the land surface albedo was used. All these indicators were retrieved from Landsat 5 TM satellite images. MSAVI1 was calculated using the reflectance of red (RED) and near-infrared (NIR) bands and the L factor, employing the following formula: N I R−RED MSAVI1 ¼ ð 1 þ LÞ N I R þ RED þ L
ð1Þ
where L=1−2⋅s⋅NDVI⋅WDVI. In this formula, s is the slope of the soil line, and NDVI and the Weighted Difference Vegetation Index (WDVI) are calculated based on the following formulae, the terms having the same significance: N IR−RED N IR þ RED
ð2Þ
WDVI ¼ N IR−s⋅RED
ð3Þ
NDVI ¼
The slope of the soil line (s) and the intercept of the soil line (b) were calculated on the basis of the surface area of bare soil in each image. By identifying the range of pixel values representing bare soils in the NDVI image, we were able to obtain a mask image where bare soils have a value of 1, and all other pixels have a value of zero. Landscape pattern, or heterogeneity, in desertification risk was assessed by using MSDI, considered to be a good indicator in the assessment of landscape heterogeneity for land degradation (Xu et al. 2009). Several authors (Tanser 1997; Pickup 1990) showed that any process leading to an increase of landscape heterogeneity in time and space would result in its degradation. The hypothesis behind the heterogeneity index is that a healthy landscape shows less variability than a degraded landscape (Tanser 1997). Degraded landscapes in arid and semi-arid areas can be highly patterned because of increased erosion (Ienciu et al. 2012). The vegetation change from one location to another is often spatially variable due to redistribution of sediments and water (Friedel et al. 1993). When observing an increase of spatial heterogeneity, the landscape has moved from a state of equilibrium to non-equilibrium, and one can say that it has become dysfunctional (Ludwig and Tongway 1997). MSDI was calculated using a 3×3 filter that was passed across the red band, and we calculated the standard deviation for every nine-pixel window. Then, the standard deviation was placed onto a new map at the same location as the central pixel of each nine-pixel window (Tanser 1997; Tanser and Palmer 1999; Jafari et al. 2008; Xu et al. 2009). The formula used for MSDI was as follows:
MSDI ¼
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u N uX 2 u DN i −DN u t i¼1 N
ð4Þ
where N represents the pixel number of the filter (N= 9), DNi is the digital number of pixel, i, in each ninepixel window, and DN is the average digital number value of each nine-pixel window. The albedo was calculated on the basis of radiometrically normalised Landsat TM images, using the following formula:
Environ Monit Assess (2015)87:4
Page 5 of 17 204
Albedo ¼ 0:356α1 þ 0:130α3 þ 0:373α4 þ 0:085α5 þ 0:072α7 −0:0018
ð5Þ
where α1, α3, α4, α5 and α7 are narrow-band albedos for bands 1, 3, 4, 5 and 7 of Landsat TM. Land surface temperature retrieval LST is an indicator of micrometeorological conditions that was used in analysing certain desertification indicators (MSDI, albedo), without being included in the decision model for the assessment of the desertification risk. For each image, the LST was retrieved from Landsat TM thermal band (11.45–12.50 μm). The first step consisted in converting the digital numbers (DN) into spectral radiance, using the following equation (Landsat Project Science Office 2002): Lmax −Lmin Lλ ¼ ð6Þ DN þ Lmin 255 where Lλ is the spectral radiance and Lmax and Lmin (mW×cm−2 ×sr−1 ×μm) are the spectral radiance values for each band at digital numbers 0 and 255, respectively. The values Lmax and Lmin for Landsat TM are 15.303 and 1.2378 (Chander and Markham 2003). The next stage was to convert the spectral radiance into at-satellite brightness temperature (blackbody temperature) considering uniform emissivity, using the following formula: TB ¼
K 2 K1 ln þ1 Lλ
ð7Þ
where TB is the effective at-satellite temperature in Kelvin degrees, Lλ is the spectral radiance from Eq. (6), and K1 and K2 are the pre-launch calibration constants. For Landsat TM, the pre-launch calibration constants K1 and K2 in mW×cm− 2 ×sr− 1 ×μm, obtained from the Landsat data user’s manual, are 607.76 and 1260.56, respectively. The temperature values obtained by applying formula (7) are referenced to the blackbody. These temperatures were corrected taking into consideration the spectral emissivity (ε) for each type of land cover. Each land use/land cover category received an emissivity value according to the emissivity classification scheme (Snyder et al.
1 99 8 ) . T h e f ol l o w i n g f or m u l a (A r t i s a n d Carnahan 1982) was used: St ¼
T B TB 1þ λ lnε ρ
ð8Þ
where St is the corrected land surface temperature (in Kelvin degrees), λ is the wavelength of an emitted radiance (λ=11.5 μm), ρ=h×c/σ (1.438×10−2 m/K), σ is the Boltzmann constant (1.38 × 10−23 J/K), h is Planck’s constant (6.626×108 m/s) and c is the velocity of light (3×108 m/s). The temperatures obtained using Eq. (8) were expressed in Kelvin, and then they were converted into Celsius.
Method of assessment of the desertification risk In this study, DTC was used to assess the risk of desertification in Dobrogea. DTC is a type of multistage classifier, which can be applied to a single image or a stack of images. It is made up of a series of binary decisions (Teresneu 2012), specific to each stage, which are used to determine the correct category for each pixel in the image in the simplest way possible. Many studies (Hui et al. 2009; Xu et al. 2009; Elnaggar and Noller 2010) recommend the use of these techniques due to the substantial advantages for classification purposes: simplicity, flexibility, intuitiveness and efficient calculation. Applying DTC involved establishing in advance certain rules for construction of a decision model. The rules for each indicator were set on the basis of ancillary data available and of a thorough analysis of the range of values that each indicator can acquire in different conditions (Table 1). The fact that, for example, the albedo in normal conditions for agricultural land is placed in the Table 1 The rule sets for the assessment of the risk of desertification by using Landsat TM images Grade of desertification risk
Indicators MSAVI1
MSDI
Albedo
Non
0.4
≥0
0.25
204
Page 6 of 17
range, 0.18–0.25, was considered (Ahrens 2006). Figure 2 presents the DTC for risk assessment of desertification in Dobrogea. The DTC input consists of MSAVI1, MSDI and albedo images previously obtained through specific processing. The first decision rule has taken into consideration, MSAVI1, that segments images into four parts. If MSAVI0.4 (land covered by dense and healthy vegetation), (2) 0
Assessing and monitoring the risk of desertification in Dobrogea, Romania, using Landsat data and decision tree classifier Iosif Vorovencii
Received: 3 June 2014 / Accepted: 12 March 2015 # Springer International Publishing Switzerland 2015
Abstract The risk of the desertification of a part of Romania is increasingly evident, constituting a serious problem for the environment and the society. This article attempts to assess and monitor the risk of desertification in Dobrogea using Landsat Thematic Mapper (TM) satellite images acquired in 1987, 1994, 2000, 2007 and 2011. In order to assess the risk of desertification, we used as indicators the Modified Soil Adjustment Vegetation Index 1 (MSAVI1), the Moving Standard Deviation Index (MSDI) and the albedo, indices relating to the vegetation conditions, the landscape pattern and micrometeorology. The decision tree classifier (DTC) was also used on the basis of pre-established rules, and maps displaying six grades of desertification risk were obtained: non, very low, low, medium, high and severe. Land surface temperature (LST) was also used for the analysis. The results indicate that, according to preestablished rules for the period of 1987–2011, there are two grades of desertification risk that have an ascending trend in Dobrogea, namely very low and medium desertification. An investigation into the causes of the desertification risk revealed that high temperature is the main factor, accompanied by the destruction of forest shelterbelts and of the irrigation system and, to a
Iosif Vorovencii holds a Ph.D. degree. I. Vorovencii (*) Forest Engineering, Forest Management Planning and Terrestrial Measurements Department, Faculty of Silviculture, Transilvania University of Brasov, Beethoven street nr. 1, 500123 Brasov, Romania e-mail: [email protected]
smaller extent, by the fragmentation of agricultural land and the deforestation in the study area. Keywords Desertification . Decision tree classifier . MSAVI1 . MSDI . Albedo
Introduction Desertification is a general degradation process that occurs on the surface of land because of the combined effects of human and natural causes that lead to decrease in, or loss of, the soil’s capacity to support the growth of plants (Helldén 1991). Desertification is associated with long-term changes in the functions of ecosystems (Dregne 1977), involving both the spatial and the temporal component and leading to a reduced productivity and biodiversity (Mouat et al. 1997). The surfaces affected by desertification are characterised by scarce precipitation and intense evapotranspiration which lead to sparse growth of vegetation and the occurrence of groups of simple structure plants. Over the last 2 decades, remote sensing was widely used in research concerning the assessment of desertification (Sun et al. 2005; Yan et al. 2009; Xu et al. 2009; Santini et al. 2010; Xu et al. 2010; Pan and Li 2013). The types of satellite images frequently involved in monitoring and assessing the risk of desertification are Landsat Multispectral Scanner (MSS), Landsat Thematic Mapper (TM), Landsat Enhanced Thematic Mapper Plus (ETM+) (Liu et al. 2008; Zhang et al. 2008; Xu et al. 2009; Pan and Li 2013), NOAA-
204
Environ Monit Assess402:781 )5 02(
Page 2 of 17
AVHRR (Arnab and Dipanwita 2011) and SPOT Vegetation (Huang and Siegert 2006). Most of the methods used in remote sensing assess and monitor the desertification risk using vegetation indices related to vegetation coverage (Chopping et al. 2008; Hanafi and Jauffret 2008; Xue et al. 2009). One such vegetation index frequently used in desertification studies is the Normalised Difference Vegetation Index (NDVI). In the semi-arid and arid regions, because of the sparse vegetation, the soil and its humidity have a significant impact on the NDVI (Wessels et al. 2004; Huang and Siegert 2006; Hill et al. 2008; Pan and Li 2013). In order to diminish the influence of the background of soil on the results, certain researchers have opted for the Modified Soil Adjustment Vegetation Index (MSAVI) or the Enhanced Vegetation Index (EVI) (Wu et al. 2007; Li et al. 2010). The assessment of the desertification risk using a single vegetation index (NDVI, MSAVI, EVI) is unilateral and fails to offer a comprehensive view of all the aspects of desertification (Pan and Li 2013). Certain studies (Xu et al. 2009) adopted quantitative analysis and monitored desertification by establishing quantitative relationships between the desertification process and biophysical features of the land surface such as surface albedo, land surface temperature (LST) and moving window. The surface albedo retrieved from remote sensing data is a physical parameter that renders reflective features of a surface and solar shortwave radiation. It is affected by vegetation coverage, land use, soil humidity, snow cover and other conditions of the land surface (Vorovencii 2014). Any change in the features of a land surface leads to a visible change in the surface albedo. The novelty of this article consists in assessing, for the first time, the risk of desertification in Dobrogea, Romania, using Landsat TM satellite images and the decision tree classifier (DTC). We have also proposed our own reference scheme for zoning the surfaces in the study area by grades of desertification risk and for analysing the causes. Knowing the expansion and severity of the desertification process is an important factor in the decisionmaking process for an effective control; this is the reason why the desertification phenomenon must be assessed and monitored. The general objective of this paper is to assess and monitor the desertification risk in Dobrogea, Romania, in the period of 1987–2011, using Landsat TM images. The specific objectives of the
research are the following: (1) to identify certain indicators and to develop a decision model using the decision tree classifier (DTC) in order to assess and monitor the desertification risk; (2) to use the indicators MSAV I1, Moving Standard Deviation Index (MSDI), the albedo, and their combinations in assessing the various grades of desertification risk; and (3) to analyse the causes that led to the desertification risk, including the LST analysis of the desertification phenomenon.
Materials and methods Study area The study area is the Dobrogea Region of Romania, including the counties of Constanta and Tulcea, and excluding the Danube Delta (Fig. 1). The total surface taken into consideration in the study is 1,176,582 ha. Out of 707,100 ha in Constanta County, 564,500 ha represents agricultural land and in Tulcea County; out of 849,900 ha, 362,218 ha represents agricultural land. The altitude is relatively low (200–300 m). In the southern part, the altitude is less than 200 m (with a maximum of 204 m in Deliorman Plateau), and in the northern part, the altitude reaches 350–400 m and even 467 m (Greci Peak in the Macinului Mountains). The most common climate in Dobrogea is an arid climate, with high average temperatures (10–11 °C), very high temperatures in summer (22–23 °C), scarce precipitation (on average under 400 mm/year), tropical days and frequent droughts. The region closer to the Black Sea is characterised by a more moderate climate, diurnal sea breezes and strong insolation. At altitudes exceeding 300 m, there is a low hill climate, with a lower average temperature (9–10 °C) and more abundant precipitation (500–600 mm/year). The average temperature of the coldest month (January) is, for most of the area, from −1 to −2 °C, and in the south-eastern part, it is positive. The annual thermal amplitude varies from 23 to 24 °C in the western part of Dobrogea to 21– 22 °C in the eastern part. Vegetation is varied: In the south, the vegetation is dry steppe covered by grass, which has been replaced mostly by cultivated plants; in the north, on the hills and in the low altitude mountains, the vegetation is silvosteppe combined with forest vegetation made up of red oak (Quercus rubra) and Hungarian oak (Quercus fraineto). The surface area covered by forests represents
Environ Monit Assess (2015)87:4
Page 3 of 17 204
Fig. 1 Study area. The studied surface is the Dobrogea region without the Danube Delta. In the eastern part of the study area (dashed line), there are the following lakes: Razim, Sinoe, Golovita, Zmeica, Histria, Nuntasi, Tuzla, Ceamurlia, Caraburum and Salcioara
only 3.5 % (24,750 ha) of Constanta County and 10.8 % (91,790 ha) of Tulcea County, as compared to the 26.7 %, average for Romania. The soils, which cover most of the surface area, belong to the class of chernozems, cambic chernozems and grey soils. A smaller surface area (around 5 %) is covered by brown luvic soils, rendzina, litosoils, erodosoils and regosol. Landsat and ancillary data The study used five frames clipped from Landsat 5 TM satellite images acquired on 14 July 1987, 01 July 1994, 01 July 2000, 21 July 2007 and 16 July 2011. The images are part of path/row 181/29 and were downloaded from the Internet (http://glovis.usgs.gov). They have a 1T level of correction and were georeferenced at Universal Transverse Mercator (UTM), datum WGS 84, zone 35 N. The spatial resolution of the multi-spectral and thermal bands, after resampling,
is 30 m. For an accurate assessment of the images obtained after processing and recoding, the following were used: the soils map (1:1,000,000), the vegetation map (1:100,000), the radiation distribution map (1:100,000), the land erosion map (1:1,000,000) and the drought risk map (1:1,000,000). Meteorological data, including monthly and annual averaged temperature and rainfall, of eight meteorological stations in and around Dobrogea, were collected from different reports (Bosneagu et al. 2010) and from the Romanian Meteorological Institute. Social and economic census data of Dobrogea was collected from local statistical yearbooks. Image pre-processing Relative radiometric normalisation (RRN) was carried out for all these images, except the 1994 image which represented the reference image. The RRN method used
204
Environ Monit Assess402:781 )5 02(
Page 4 of 17
was the Automatic Scattergram-Controlled Regression (ASCR) proposed by Eldvige et al. (1995). This method is based on the Scattergram-Controlled Regression (SCR) developed by Yuan and Elvidge (1993), but it contains some improvements concerning automation of the procedure. Application of the ASCR consisted in identifying the Bno-change^ pixel set, comprised of those pixels that occupy the core of the water and land data clusters visible in red near-infrared scattergrams. On the basis of the identified no-change pixels, we were able to calculate, by regression, the gains and offsets used in the normalisation of the subject images in order to match the reference image. The procedure is described in detail in Eldvige et al. (1995). Cloud removal was performed for the 1987 and 2011 images. Indicators of the desertification risk The changes resulting from desertification are mainly characterised by changes of vegetation biomass or cover, landscape pattern and micrometeorological conditions (Xu et al. 2009). In assessing vegetation biomass or land cover, various vegetation indices can be employed, such as NDVI, MSAVI and EVI. Taking into consideration the strong relationships among these indices, in the present study, the MSAVI1 index was chosen in order to assess land cover. MSAVI1 was shown to operate better than NDVI, SAVI and WDVI (Qi et al. 1994). In order to characterise the landscape pattern, MSDI was used, and for assessment of the micrometeorological conditions of the land surface, the land surface albedo was used. All these indicators were retrieved from Landsat 5 TM satellite images. MSAVI1 was calculated using the reflectance of red (RED) and near-infrared (NIR) bands and the L factor, employing the following formula: N I R−RED MSAVI1 ¼ ð 1 þ LÞ N I R þ RED þ L
ð1Þ
where L=1−2⋅s⋅NDVI⋅WDVI. In this formula, s is the slope of the soil line, and NDVI and the Weighted Difference Vegetation Index (WDVI) are calculated based on the following formulae, the terms having the same significance: N IR−RED N IR þ RED
ð2Þ
WDVI ¼ N IR−s⋅RED
ð3Þ
NDVI ¼
The slope of the soil line (s) and the intercept of the soil line (b) were calculated on the basis of the surface area of bare soil in each image. By identifying the range of pixel values representing bare soils in the NDVI image, we were able to obtain a mask image where bare soils have a value of 1, and all other pixels have a value of zero. Landscape pattern, or heterogeneity, in desertification risk was assessed by using MSDI, considered to be a good indicator in the assessment of landscape heterogeneity for land degradation (Xu et al. 2009). Several authors (Tanser 1997; Pickup 1990) showed that any process leading to an increase of landscape heterogeneity in time and space would result in its degradation. The hypothesis behind the heterogeneity index is that a healthy landscape shows less variability than a degraded landscape (Tanser 1997). Degraded landscapes in arid and semi-arid areas can be highly patterned because of increased erosion (Ienciu et al. 2012). The vegetation change from one location to another is often spatially variable due to redistribution of sediments and water (Friedel et al. 1993). When observing an increase of spatial heterogeneity, the landscape has moved from a state of equilibrium to non-equilibrium, and one can say that it has become dysfunctional (Ludwig and Tongway 1997). MSDI was calculated using a 3×3 filter that was passed across the red band, and we calculated the standard deviation for every nine-pixel window. Then, the standard deviation was placed onto a new map at the same location as the central pixel of each nine-pixel window (Tanser 1997; Tanser and Palmer 1999; Jafari et al. 2008; Xu et al. 2009). The formula used for MSDI was as follows:
MSDI ¼
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u N uX 2 u DN i −DN u t i¼1 N
ð4Þ
where N represents the pixel number of the filter (N= 9), DNi is the digital number of pixel, i, in each ninepixel window, and DN is the average digital number value of each nine-pixel window. The albedo was calculated on the basis of radiometrically normalised Landsat TM images, using the following formula:
Environ Monit Assess (2015)87:4
Page 5 of 17 204
Albedo ¼ 0:356α1 þ 0:130α3 þ 0:373α4 þ 0:085α5 þ 0:072α7 −0:0018
ð5Þ
where α1, α3, α4, α5 and α7 are narrow-band albedos for bands 1, 3, 4, 5 and 7 of Landsat TM. Land surface temperature retrieval LST is an indicator of micrometeorological conditions that was used in analysing certain desertification indicators (MSDI, albedo), without being included in the decision model for the assessment of the desertification risk. For each image, the LST was retrieved from Landsat TM thermal band (11.45–12.50 μm). The first step consisted in converting the digital numbers (DN) into spectral radiance, using the following equation (Landsat Project Science Office 2002): Lmax −Lmin Lλ ¼ ð6Þ DN þ Lmin 255 where Lλ is the spectral radiance and Lmax and Lmin (mW×cm−2 ×sr−1 ×μm) are the spectral radiance values for each band at digital numbers 0 and 255, respectively. The values Lmax and Lmin for Landsat TM are 15.303 and 1.2378 (Chander and Markham 2003). The next stage was to convert the spectral radiance into at-satellite brightness temperature (blackbody temperature) considering uniform emissivity, using the following formula: TB ¼
K 2 K1 ln þ1 Lλ
ð7Þ
where TB is the effective at-satellite temperature in Kelvin degrees, Lλ is the spectral radiance from Eq. (6), and K1 and K2 are the pre-launch calibration constants. For Landsat TM, the pre-launch calibration constants K1 and K2 in mW×cm− 2 ×sr− 1 ×μm, obtained from the Landsat data user’s manual, are 607.76 and 1260.56, respectively. The temperature values obtained by applying formula (7) are referenced to the blackbody. These temperatures were corrected taking into consideration the spectral emissivity (ε) for each type of land cover. Each land use/land cover category received an emissivity value according to the emissivity classification scheme (Snyder et al.
1 99 8 ) . T h e f ol l o w i n g f or m u l a (A r t i s a n d Carnahan 1982) was used: St ¼
T B TB 1þ λ lnε ρ
ð8Þ
where St is the corrected land surface temperature (in Kelvin degrees), λ is the wavelength of an emitted radiance (λ=11.5 μm), ρ=h×c/σ (1.438×10−2 m/K), σ is the Boltzmann constant (1.38 × 10−23 J/K), h is Planck’s constant (6.626×108 m/s) and c is the velocity of light (3×108 m/s). The temperatures obtained using Eq. (8) were expressed in Kelvin, and then they were converted into Celsius.
Method of assessment of the desertification risk In this study, DTC was used to assess the risk of desertification in Dobrogea. DTC is a type of multistage classifier, which can be applied to a single image or a stack of images. It is made up of a series of binary decisions (Teresneu 2012), specific to each stage, which are used to determine the correct category for each pixel in the image in the simplest way possible. Many studies (Hui et al. 2009; Xu et al. 2009; Elnaggar and Noller 2010) recommend the use of these techniques due to the substantial advantages for classification purposes: simplicity, flexibility, intuitiveness and efficient calculation. Applying DTC involved establishing in advance certain rules for construction of a decision model. The rules for each indicator were set on the basis of ancillary data available and of a thorough analysis of the range of values that each indicator can acquire in different conditions (Table 1). The fact that, for example, the albedo in normal conditions for agricultural land is placed in the Table 1 The rule sets for the assessment of the risk of desertification by using Landsat TM images Grade of desertification risk
Indicators MSAVI1
MSDI
Albedo
Non
0.4
≥0
0.25
204
Page 6 of 17
range, 0.18–0.25, was considered (Ahrens 2006). Figure 2 presents the DTC for risk assessment of desertification in Dobrogea. The DTC input consists of MSAVI1, MSDI and albedo images previously obtained through specific processing. The first decision rule has taken into consideration, MSAVI1, that segments images into four parts. If MSAVI0.4 (land covered by dense and healthy vegetation), (2) 0
