Introduction

Arid and semi-arid ecosystems are those in which the ratio of total annual precipitation to potential evapotranspiration is 0.05–0.50; these areas cover about 30 % (about 4 billion ha2) of global dry land (Beadle 1959; Bailey 1979; Paruelo et al. 2000; Diouf and Lambin 2001; Lal 2004; Austin and Vivanco 2006; Heisler-White et al. 2008; Rotenberg and Yakir 2010; Walker 2012; Riha et al. 2014; Letnic et al. 2015). Vegetation in dry-land ecosystems support global biodiversity, carbon sequestration, and the majority of the world’s livestock. The woody plant mosaic in dry-land ecosystems is a fundamental determinant of key ecosystem processes (e.g., evapotranspiration, fire disturbance) and associated abiotic patterns (Breshears 2006).

Monitoring long-term changes in an ecosystem over large spatial extents is critical for understanding the dynamics of woody plants in arid and semi-arid ecosystems and their responses to natural disturbance and rangeland management (Jian et al. 2012; Saranya et al. 2014). Monitoring long-term tree cover dynamics in semi-arid woodlands requires repeated retrieval of the tree canopy cover area with relatively high accuracy (Hostert et al. 2003; Cohen et al. 2003; Jian et al. 2012).

Field measurement of the tree canopy cover is expensive, labor-intensive, and often limited in temporal scope and spatial scale. Remote sensing (RS) has been demonstrated to be a suitable alternative for quantifying biophysical variables such as leaf area index and vegetation cover (Cohen et al. 2003; Tang et al. 2014; Béland et al. 2014). Tree canopy cover in dry land ecosystems can be confounded by shrub cover and can contribute considerably to leaf area and foliar biomass. Quantifying tree canopy cover in dry land ecosystems using RS data can be particularly challenging in areas with low tree cover and density, short tree stature, and coexistence of trees and shrubs at fine spatial scales (Ko et al. 2009). In arid and semi-arid regions, changes in woody plant cover have dramatic effects on ecosystems (Breshears 2006; Homet-Gutiérrez 2015). RS has been used to provide spatially explicit information about the heterogeneity of woody plant distribution over extensive dry land areas (Walker et al. 2012; Ferreira et al. 2015).

Landsat archival imagery (LAI) can be effective for monitoring woodland expansion and contraction in semi-arid landscapes over time (Zhu et al. 2012; Jian et al. 2012; Griffiths et al. 2014; Schmidt et al. 2015). Studies on different biomes, including cropland, plantations, and forests (conifer and deciduous stands) have successfully linked LAI measurements on the ground using direct or indirect techniques to RS data (Chen and Black 1991; Chen and Cihlar 1996; Colombo et al. 2003; Brantley et al. 2011). About 300 vegetation indices have been published; however, only a few of those based on biophysics or on specific methods have been adopted (Thenkabail et al. 2016). These studies correlate spectral vegetation indices (SVIs) from different types of satellite data (Adina et al. 2014). The most popular of these indices are the normalized difference vegetation index (NDVI), simple ratio (SR) index, and soil-adjusted vegetation index (SAVI).

The relationship between LAI and different combinations of SVIs has been analyzed in a variety of studies (Adina et al. 2014; Le Maire et al. 2011). Besides linear, quadratic polynomial, and cubic polynomial links, most studies have shown logarithmic relationships (Tucker 1979; Myneni et al. 1997; Chen and Cihlar 1996; Datt 1999; Mutanga and Skidmore 2004; Le Maire et al. 2011; Potithep et al. 2013). Time series analyses provide powerful alternatives with their ability to separate seasonal variation from long-term trends (Sulla-Menashe et al. 2014; Lanorte et al. 2014; Starr et al. 2015; Ahmed et al. 2015). Trend analysis has been frequently applied to characterize land surface phenology change from coarse-scale imagery (Slayback et al. 2003; Heumann et al. 2007; Bradley and Mustard 2008).

Our study tested instrumental and technological solutions that efficiently and accurately describe desertification dynamics through creation of vegetation degradation maps for rational management. Multispectral Landsat images and calculation of vegetation indices were chosen to accommodate the large area of the study site in the Zagros forest of Lorestan to decrease the cost of processing the technology. Field studies were carried out at single research sites using optimal vegetation index to extract results for desertification of the region. In past decades, deterioration of vegetation and land degradation in arid and semi-arid regions has forced the mangers to use new and rapid technology to extract vegetation maps on local and regional levels for monitoring and assessment. This ability increases understanding of the influences of humans and climate change on land degradation. It is crucial to calibrate and determine the best indices to obtain reliable results about the status of the vegetation.

Materials and methods

Study area

The study area is located in western Iran at 48°47′–50°3′ longitude and 32°44′–33°35′ latitude. This includes part of the central Zagros forests with an area of 32,231 ha (Fig. 1). Geomorphologic classification of area is mountainous and its climate is semi-humid cold. The species in order of frequency are Quercus brantii, Daphne mucronata, Amygdalus scoparia, Acer monspessulanum, Amygdalus lycioides, Cerasus brachypetala, Crataegus pontica, Amygdalus orientalis, Pistacia mutica, and Pistacia khinjuk (Henareh Khalyani et al. 2012; Ghanbari and Sefidi 2012).

Fig. 1
figure 1

Location of Zagros forests and study area in Iran

Full size image

Methodology

Our study improved the compatibility of the vegetation index (Table 1) with trends in the environment, climate, and cover change over a 10-year period from 2002 to 2013. Satellite imagery for the years 2002 (Landsat 5, TM), 2009 (Landsat 7, ETM+), and 2013 (Landsat 7 ETM+) with a spatial resolution of 30 m2 were used (Huang et al. 2009; Brandt et al. 2012). All images were recorded in the month of July to increase accuracy by comparison of similar time frames. Two topographic map sheets were used to identify and visit the area and as ground truth maps. Topographic map sheets were applied to identify ground control points to allow geometric correction of the satellite images and evaluate the accuracy of the geometric correction.

Table 1 Spectral vegetation indices
Full size table

Randomized systematic sampling was used in this study. The distance between the UTM grid lines was 1 km. The line transect method was used with a starting point for each transect at the cross point at the 1 km × 1 km grid lines for the 1:25,000 scale maps with forest coverage. Along each transect, every single tree or shrub species having crowns that intersect the line transect were listed by scientific name. The understory species at each transect were recorded as an index of human influence. The length of each transect depended on the density of the forest canopy. For canopy cover of <10, between 10 and 50, and >50 %, the cover transect length was 100, 200, and 400 m, respectively. The data was managed in the form of a database that was built, sorted, and filtered using Microsoft Access software. Descriptive statistical parameters such as the total number of each species, means, extremes, frequency histograms were helpful for analysis of the niches of the plants.

Because the images sent from the satellite were received at different moments, it is necessary to compare the vegetation indices by converting the digital number (DN) of the satellite images to spectral reflectance (Coppin et al. 2004; Macleod and Congalton 1998; Singh 1989). Atmosphere conditions (humidity, height of sun, azimuth of sun, accurate timing of images) from the Scanex image processing software were used to account for the effect of atmosphere on surface reflectance. It is crucial to remove the effects of atmosphere, but is not possible to assess it precisely without time data on atmospheric conditions in countries such as Iran. The following algorithm was used to convert the DN to spectral reflectance (Richards 1993; Lillesand and Kiefer 1994; Roni 2013):

$$ L = L_{\hbox{min} } + (L_{\hbox{max} } - L_{\hbox{min} } )/1023 \times DN $$
(1)

where L is spectral radiance, L min is 1.238 (spectral radiance of DN 1), L max is 15.600 (spectral radiance of DN 255) and:

$$ \rho_{p} = \frac{{\pi \times L_{\lambda } \times d^{2} }}{{ESUN_{\lambda } \times \cos \theta_{s} }} $$
(2)

where ρ p is the unitless planetary reflectance, L λ is the spectral radiance at the sensor aperture, d is the distance from the earth to the sun in astronomical units from a nautical handbook, ESUNλ is the ALI solar irradiance, and θ s is the solar zenith angle in degrees.The following formula was proposed for field research for the average percentage of vegetation cover in transects (Franklin 2001):

$$ F = \frac{{78.5T\sum {CD} }}{L} = \sum {CD} $$
(3)

where CD is the perpendicular diameter of the canopy cover of trees at the test points, TΣCD is total CD located on the transects, L is the length of the transect, and F is the density of the tree canopy on the transects. Relation 3 was used to calculate the biomass density of the trees at the test points on the transects.

Geographical coordinates were recorded by GPS in the statistical test and arc GIS 10.2 software was used to convert the geographical coordinates of the test points to vector points using Scanex. The components of each vector point in the vegetation indices were extracted separately. The average values of each vegetation indices of test points and RMSE for the transects were calculated to select the appropriate vegetation index regression formula, error coefficient, correlation coefficients of the total biomass density of the trees of all test points located in the experimental transects.

Results

The total biomass density of the trees (percentage) at all test points for the 7 transects are shown in Table 2 and the average numerical values of each vegetation index are shown in Table 3.

Table 2 Total biomass density of trees (%) at all test points on 7 transects
Full size table
Table 3 Average numerical value of each vegetation index at all test points on 7 transects
Full size table

Correlation diagrams for each indicator versus the biomass density of the trees are shown in Fig. 2. These regression diagrams were obtained using the data from more than 135 test points along the seven transects to calculate the numerical values of each vegetation index using the average values at select test points on each transect. The result shows the highest correlation coefficient, description coefficient, and the lowest RMSE for GEMI versus the density of the forest cover obtained from field calculation (Table 4).

Fig. 2
figure 2

Correlation diagrams for each indicator versus biomass density of trees

Full size image
Table 4 Results of linear regression analysis and RMSE
Full size table

A small boundary was chosen as the control area. Aerial photographs, Google Earth software, and the results of the field study were visually interpreted and the forest area was separated from bare land and imported into the software. The area of the control step was 577.3 ha2 (Table 5).

Table 5 Vegetation indices for different canopies
Full size table

A map of the control area was created for the indices using Scanex to determine the relation between the accuracy of the forest area and the indices (Fig. 3). A visual comparison of the maps (forest vegetation indices) results in percentages of <30 % for bare land, 30–75 % for forest, and 75–100 % for irrigated area. The result shows that GEMI was more accurate than the other indices used to calculate the forest area of each indicator by visual interpretation (Table 6).

Fig. 3
figure 3

Cover classes based on indices versus visual interpretation

Full size image
Table 6 Numerical value of indices in control area
Full size table

GEMI provided the most appropriate correlation for forest area to detect changes in forest area over a 10-period (Table 7). Satellite images for 2002 and 2013 from ETM+ (Landsat 7) were used to extract the vegetation indices for each image (Fig. 4).

Table 7 Changes in coverage density of study area from GEMI
Full size table
Fig. 4
figure 4

Land-use changes for study period from GEMI

Full size image

Discussion and conclusion

The results show that GEMI was the best choice for detecting changes in the study area and other areas (Torahi and Rai 2011; Jian et al. 2012; Henareh Khalyani et al. 2012; Adina et al. 2014). GEMI (R 2 = 0.94) was chosen because it produced the greatest accuracy and ability to separate the border of the canopy cover of trees. It was used to create a formula for calculating the biomass density of trees based on the geo-ecology of the study region. GEMI has been empirically shown to be insensitive to atmospheric influences, but other drawbacks to the index have not been uncovered. The value from GEMI correlated highly with the field survey and was able to determine the area of the biomass of trees. It can normally divide border regions with the low biomass density of trees from regions without canopy cover.

SAVI (R 2 = 0.81) for L = 1 showed the lowest correlation related to its maximum value of coefficient L as applied to arid land; this was a result of the minimizing effects of soil spectral reflectance (Colombo et al. 2003). NDVI was acceptable in most studies with relatively good coverage (Peters et al. 2002) and EVI was good for forest area (Chaban 2004; Galvao et al. 2011), but had a lower correlation in the study area (R 2 = 0.88) than GEMI. This could be the result of the higher impact factor in the near-infrared band in GEMI (the band in which vegetation has high reflectance).

Semi-arid woodland areas (such as the oak forest of western Iran) showed a diversity of ground cover along the transects together with GPS accuracy. It can produce high uncertainty when relying on the statistical analysis of a single pixel data. The averages of the records in each transect were used to cover spatial variation and the results revealed better correlation in some area, such as Zagros forest, in arid and semi-arid areas.

The accuracy of indices depends on field studies and expertise for classification for image analysis. The results showed that the RMSE error rate in this study was very small and field studies showed high accuracy. The biomass density of trees as assessed for regions with similar geo-ecological conditions for the study region using the proposed formulas. As seen in thematic maps covering 11 years, the total area of the forest (with a density of 30–75 % of biomass) has decreased by 2720 ha2 (16.91 %). It can concluded that damaging agricultural methods near forests (Pourhashemi et al. 2004), inappropriate use of trees (Henareh Khalyani et al. 2014; Sadeghravesh et al. 2015), unsustainable exploitation of water resources especially groundwater resources (Zehtabian et al. 2010; Mashayekhi et al. 2010; Shirmohammadi et al. 2013; Moosavi et al. 2013; Rahmati et al. 2016), misplaced structures (Zehtabian et al. 2014), and depletion of crops had increased desertification in the study region.