Data Pre-Processing

            Two sets of seasonal LANDSAT TM and ETM+ spring and summer images were obtained covering the period between 1985 and 2004.  Spring and summer images of the years 1985, 1987, 1991, 1997, 2001, 2003 and 2004 were obtained.  The criteria of image selection was based on the prevailing climatic conditions at time of image acquisition and on the climate profile of the year under investigation.  Effort was made to select images from years with similar climatic profiles (precipitation distribution and temperature variation).

            The satellite images were geometrically registered using a set of ground Control Points (GCPs) collected using a differential GPS.  GPS points used in image registration were taken at road intersections visible in the LANDSAT TM and ETM+ images. Geometric registration accuracy of 1 pixel (30 meters) was obtained in the registration process.  Topographic relief displacement was accounted for by orthorectification of the LANDSAT TM Images by using the GCPs and and a Digital Terrain Model (DTM).  The modification in the polynomial equation used with orthorectification as opposed to ordinary rectification is that it takes into account terrain elevation, local earth curvature, distance from nadir, and flying height above datum to get a polynomial transformation between the image and ground coordinates (ERDAS, 1997).

            The second step in data pre-processing was the atmospheric correction of the LANDSAT TM and ETM+ images.  The objective of an atmospheric correction is the elimination of atmospheric effects so that the energy levels present on a satellite image resemble those emitted by the features on the earth’s surface so that changes in satellite image energy levels can be attributed to changes in the surface and not to changes in the atmosphere.  The dark pixel subtraction method was utilized to atmospherically correct the images.  The effectiveness The effectiveness of the method in reducing the in between-scene illumination variation was quantitatively resolved. This was accomplished by regressing the surface-reflectance-corrected spectra of psuedo-invariant targets between the three images for all bands. It was observed that the regression coefficients are insignificantly different from 1 which indicates that the psuedo-invariant targets have the same reflectance values across the three images, therefore, it can be concluded that the in between-scene illumination variation was greatly reduced. However, the goodness of fit values for the regression models, particularly for bands 1 and 2, were different from 1. This shows that the resultant surface-reflectance-corrected spectra have residual artifacts. This can be further corrected for by the calibration of remote sensing data with in situ radiometric measurements made at the same time of data acquisition. Although the removal of such artifacts is necessary to extract biophysical information from water bodies, most land-cover-related remote sensing investigations do not require such high accuracy standards.



Top

Biomass Changes Detection

           Following the atmospheric and geometric correction of the satellite images, the images were analyzed.  Several spectral indices were derived that are indicators of green biomass relative abundance but not measures of Biomass.   These were the Normalized Difference Vegetation Index (NDVI), the Modified Soil Adjusted Vegetation Index II (MSAVI II) and the Vegetation Reflectance Model (VR).

           In order to measure biomass, the indicators of biomass relative abundance were regressed against a set of field biomass observations.  This was achieved as follows:

• First: fifty five plots were randomly selected in natural areas occupied with herbaceous plant species.  The plot dimensions were approximately 90 by 90 meters.

• Second: A number of quadrates ranging from 2 to 6 quadrates with a dimensions ranging from 2 by 2 meters to 12 by 12 meters were randomly selected within each plot.  Green biomass within each plot was harvested, oven dried and weighted.

• Third: A Kriging spatial interpolation method between the quadrates within each plot was used to derive a continuous biomass surface for each plot.

• Fourth: The plots were overlaid in a GIS over the derived vegetation indices; namely NDVI, MSAVI II and VR model.  The vegetation indices were derived from the spring 2004 image acquired over the same period of field observations.

• Vegetation indices' values which are spatially corresponding to plot locations were regressed against the measured Biomass values (g/m²).

• The regression equation was extended spatially to calculate biomass from vegetation indices for the entire study area and was after that extended temporally to the other images to monitor changes from the year 1985 to the year 2004.

           Figure 1 shows the regression results between the 55 plots and the Normalized Difference Vegetation index values obtained from the spring 2004 image.  Regression coefficients were significant at the 99% confidence level and the goodness of fit values for the regression model was 0.77.

Figure 1. regression equation between NDVI values and observed biomass (g/m2) for the training locations.

Top


              Comparing the results from natural biomass analysis obtained from the multi-temporal set of satellite images has revealed a significant drop in average natural biomass in the study area.  Figures 2a and 2b respectively show the drop in biomass as observed in spring and summer seasons over the years from 1985 to 2004. 

Figures 2.a and 2.b show the changes in average natural biomass in gram per square meter for the study area.

Top


            On the other hand, the measured biomass in agricultural areas has slightly increased from the years 1985 to the year 2004 which indicates a more intensive cropping pattern.  Figures 3a and 3b respectively show the increase in agricultural biomass as observed in spring and summer seasons over the years from 1985 to 2004. 

Figures 3a and 3b show the changes in average agricultural biomass in gram per square meter for the study area.

Top

Land Use and Land Cover

             Land use and land cover are two approaches for describing land.  Land use is a description of the way that humans are utilizing any particular piece of land for one or many purposes.  Comparatively, land cover is the bio-physical material covering the earth's surface at any particular location.  Together land use and land cover information provide a good indication of the landscape condition and processes that are occurring at a particular place.  Time series of land use/land cover maps tell us how much of the landscape is changing, as well as what changes have occurred and where the changes are taking place.  Accurate and timely mapping of land use/land cover provides vital information on the state of the environment, development trends and wildlife habitat among others. 

             The land cover types identified in the Dead Sea Basin were urban areas, road network, palm trees, banana trees, olive trees, vineyards, citrus plantations, other unclassified fruit trees, vegetables, wheat, natural trees (forest), shrubs land, Natural grass land, open space with little vegetation, open space with little or no vegetation, bare soil & rocks and fallow agricultural land and water bodies.  Several sampling points were collected for each of the aforementioned land cover types.  The spectral properties of the land cover types on the LANDSAT ETM+ image were inter-compared.  The results can be summarized as follows:

·         The three visible bands were not particularly useful for the discrimination between mixed fruit trees, olive grooves and banana plantations, these three classes had very low reflectance values which overlap in the visible bands (Figure 3). 

·         The near infrared (NIR) and the mid infrared (MIR) were the best spectral regions for the discrimination between vineyards and the other fruit tree classes found on the image. The NIR and MIR values of vineyards were significantly higher than the NIR and MIR reflectance values of other fruit trees; however, there is still some overlap between vineyards reflectance values in the NIR channel and the Palm trees reflectance values (Figure 3). These findings were also verified by the Euclidean-Distance separability index. The within land cover averaged Euclidean Distance was similar to the between land cover types Euclidean distance for the different fruit trees classes (table 2).

·         The LANDSAT ETM+ reflectance bands did not provide sufficient separability between all fruit trees classes except for the vineyards class which is spectrally separable from the other fruit trees.

·         Wheat and irrigated vegetables had similar reflectance values in all bands.

Figure 2: Spectral profiles of the different fruit tress types found in the Dead Sea Basin

Top
  

         An iterative process of refining the sampling areas in order to maximize the between land cover types separability was followed.  This included deleting the sampling areas with high within group variance values and adding new sampling areas with low within group variance values.The procedure resulted in a 10% average increase in between groups’ Euclidean distance.  Table 3 provides information as regards to the are of each land cover class in the Dead Sea Basin.

 

Top

            Accuracy assessment was applied to evaluate the classification results and the overall accuracy for the classification process.  An error matrix was determined and the Producer's (a measure of omission error), User's Accuracy (a measure of commission error) and the Kappa coefficient of agreement were calculated for each class and for the overall map.  The Kappa coefficient measures the agreement between the classified and reference data corrected for chance agreement (Congalton and Green, 1999).  A value greater than 0.80 represents strong agreement and a value between 0.40 and 0.80 represents moderate agreement. A minimum sample size of thirty points per class is generally recommended for a valid accuracy assessment for that particular class. 

Table 4 Accuracy Assessment results of the classification procedure

Top