P.Rama Chandra Prasad
National Collateral Management Services Limited,
Hyderabad, India.
Ch. Sudhakar Reddy, G.Rajasekhar
Forestry and Ecology Division,
National Remote Sensing Agency, Dept of Space,
Balanagar, Hyderabad, India – 500037.
C.B.S.Dutt
Indian Space Research Organization,
Dept of space, Anthariksh Bhavan,
Bangalore, India.
ABSTRACT
A detailed vegetation map is required for the effective management of natural resources. An attempt was made for the first time to prepare vegetation type of north Andaman Islands using the high resolution LISS III satellite data. In the present study visual interpretation along with digital supervised classification aided in preparation of more accurate and precise vegetation type map of north Andaman Islands. In digital classification the accuracy of map was found to be low due to limited spectral bandwidth available with IRS 1C/1D satellite data, which is about 80 – 100 nano meters. Patch characterization studies showed presence of large patches in semi evergreen forest which forms dominant as well as diversified community of the study area.
Abbreviations
IRS – Indian Remote Sensing, DEM- Digital Elevation Model, LISS – Linear Imaging Self Scanner, IGBP – International Geosphere Biosphere Programme, GPS – Global Positioning System.
Introduction
Vegetation acts as an integrator of many of the physical and biological attributes of an area, (Specht 1975, Austin 1991) and a vegetation map can be used as an approach for vegetation evaluation. International Research Programes like IGBP, mainly aims at the global scale vegetation mapping (Mayaux et al., 2002) and one of the important objectives of vegetation mapping is to define, distinguish vegetation units (Nilsen and Elvebakk, 1999) for the effective management of natural resources (Trisurat et al., 2000). The advent of remote sensing and GIS techniques facilitated the researchers to carry out detail studies pertaining to vegetation type mapping, their distribution, modeling, and endemic species habitat zonation etc.
Many studies revealed the process of mapping using imagery from different dates, resolutions, sensors and classifiers. Porwal and Pant (1989) used visual interpretation technique for the forest cover type mapping by combining 2, 3, 4 bands of Landsat TM data for Charkrata in western Himalayan region, India. Roy et al (1991) adopted both visual and digital technique for mapping forest cover in parts of Assam region using Landsat MSS data. Initially visual interpretation was adopted to stratify the forest types than supervised classification using maximum likelihood technique for further detailed mapping. Visual interpretation of Landsat TM for characterizing ecological parameters in tropical forest of Bakultala range, middle Andaman was done by Roy et al., (1993). Kimes et al (1999) used SPOT HRV images for mapping secondary tropical forest using unsupervised classification in conjunction with neural network with an overall average accuracy of 95.2%. Neig et al (1999) visually interpreted IRS-1B LISS II supplementing with Landsat TM and SPOT data to increase the accuracy of the map. Gao (1999) carried out an interesting study by comparing the spatial and spectral resolutions of satellite data for mapping mangrove forests in western waitemata harbour, New Zealand. Mapping was done from SPOT HRV and Landsat TM at 10, 20 and 30 m resolution by maximum likelihood method and found that accuracy of mapping is more for Landsat TM. Sudhakar et al (1999) mapped forest types of Jaldapara wildlife sanctuary using IRS-1B LISS II by three classifiers viz. maximum likelihood, contextual and neural network. They proposed the neural network classifier to be the best in assessing high accuracy followed by maximum likelihood. Jaykumar et al (2000) showed that visual interpretation of Landsat TM for delineating forest types was most effective classification. Khoruk et al (2003) adopted a technique of classifying AVHRR images using IDRSI software, where data was subjected to segmentation process followed by supervised classification.
The present study mainly focuses on preparation of detailed vegetation type map and analyzes the forest patch characters and their distribution with in North Andaman Islands. Assessment of forest patch characters with reference to size and numbers provides information on the species richness, composition, abundance and diversity pattern with in particular forest type.
Study area
North Andaman, the present study site, is one of the important district of Andaman & Nicobar archipelago, a group of green islands, found floating in deep blue Indian Ocean.. They are located between 12°.95’ N and 92°.86’ E, constituting about 70 large and small islands. The terrain is rough with hills enclosing narrow longitudinal valleys formed of territory sand stone, lime stone and shale. Soils are derived from sandstones, serpentines, conglomerates and are acidic non calcareous with low organic matter and high nitrogen content. Lush forest vegetation is found in these islands due to continuous rainfall brought by monsoons with a short dry period. As per the champion & Seth (1968), the study area has been classified as Andaman evergreen (1A/C2), Andaman semi evergreen (2A/C1), Andaman moist deciduous (3A/C1), Littoral (4A/L1) and Mangrove forest (4B/TS2). Materials and Methods
IRS 1C/1D LISS III data of 1st March 1999, 115 / 064 path and row with resolution of 23.5 m and 4 (2,3,4,5) bands viz, Blue (0.45– 0.52), Green (0.52-0.59), Red (0.62-0.68) and Infra Red (0.77-0.86), was used for vegetation mapping. (Fig: 1).The geometrically rectified image was subjected to the process of classification using sequential steps by unsupervised, supervised and visual interpretation techniques for the generation of final vegetation map of the study area (Fig .1). Initially image was classified by unsupervised isoclustering method to separate the pixels of the image into different spectral clusters representing various land use / land cover types. Hard copy of the spectral cluster map was generated for reconnaissance field survey to get acquainted with the general pattern of vegetation of the area and to identify the spectral clusters representing different features on ground. Traverses, along all roads and major drainage, hill tops, creeks and sandy beeches were made for collecting ground truth. The existing literature survey and interaction with forest department and local institutions was also made for collecting knowledge base. During the field survey the geographical coordinates of the predominant vegetation types and other land cover classes were marked on the cluster map using GPS instrument.
Fig: 1 False Colour Composite and Vegetation type map of North Andaman
Later image was classified digitally by the technique of supervised classification using maximum likelihood classifier, with appropriate signatures/training sets generated from half of the ground control points collected during field inventory, for corresponding land cover and vegetation classes. A thematic map representing various land cover and vegetation type classes was prepared and random samples plots were generated in each vegetation type for field sampling. The detailed field inventory for phytosociological data collection include laying quadrant of 0.1 ha size in selected location and gather information about trees, climbers, shrubs, herbs, saplings and seedlings encountered within the quadrant. The collected field data was further used for updating map with the minor variations that are observed during the field survey, finally generating vegetation type map of the study area. Accuracy assessment of the prepared map was performed by overlaying the field sample points corresponding to vegetation types as well as remaining ground control points collected during pre classified field inventory over the classified map. The accuracy of the map was found to be 71 % due to spectral mixing and overlapping between vegetation classes (Table-.2;).Hence to achieve for better performance, image was further reclassified using visual interpretation technique. A sequential interpretation method was adopted for separating different classes step by step using field knowledge and digital classified map.
Fig: 2. Visual vs. Digital method comparative forest area statistics
In the first step all the non-forest classes viz. agriculture / settlement, barren, fallow, mudflats, water bodies were masked out. Next in the second step, coastal vegetation including mangroves and littoral forest were separated based on their location and typical characteristic spectral tone. Finally in third step interior forests viz. evergreen, semi-evergreen and moist deciduous types were delineated. Based on the available ancillary data and prior ground knowledge, image classification was done using on screen visual interpretation technique by digitizing and labeling polygons for various classes using Arc View 3.2.1. The multi spectral characteristics of the different forest types i.e. variation in tone, colour, texture, shade, site, of various objects with in the satellite data formed the basis for classification. The altitude as well as aspect maps produced from DEM were also used as supporting data to locate the altitudinal vegetation types and to identify the hill shade regions while interpreting. The interpretation key used for identifying various vegetation classes on image is depicted in Table-1.
Fig: 3. Buffer Zones and Forest Patches
The final vegetation vector layer generated from hybrid approach was used for patch characterization. To identify the spatial distribution of various forest patches through out the north Andaman, five buffer zones were created starting from coast line with an interval of 1500m, proceeding towards interior forest (> 6000 m from coast). Later with in each forest type patches were categorized into six classes with an interval of 50 ha. This kind of analysis gives the scenario of forest spread at zero meters altitude of coast to 732 m high altitude of interior forest as well as location of large size patches that harbor high species richness.
Results and Discussion
The ultimate result of the classification is to distinguish the area into various forest and non-forest categories. Important vegetation types of the study area include evergreen, semi evergreen, moist deciduous, littoral, dense, degraded and open mangrove (Fig: 1). Water class was excluded from the total area statistics. Semi evergreen forest observed as dominant vegetation type of the north Andaman by both the interpretation methods. Visual technique helped in the delineation of additional stunted evergreen / southern hill top evergreen forest class (later merged with evergreen), and various sub classes within mangrove forest based on their species composition as Rhizophora, Brugeria community etc., due to the variation in spectral values and prior knowledge of the area which could not be achieved by digital method.
The accuracy as well as delineation of various classes in visually interpreted map was found to be high (85%) and this was achieved mainly by the supportive information obtained from the digital technique. The hybrid classification approach using both digital and visual methods along with the ground phytosociological data aided in producing better vegetation map of the study area. Accuracy assessment was performed only for the predominant vegetation types, since coastal vegetation (Mangroves and littoral) are easily separable (Table- 2).
Observations
- Overall there was a difference of 35 Sq.km in area between the two methods adopted for classification.
- A comparison of area statistics in visual and digital classification methods showed that the extent of total forest area is nearly similar. But for non-forest classes higher area interpretation was observed (about 40 sq.km) by visual method (Fig: 2; Table-3).
- There is also a wide range difference in the moist deciduous class.
Table-1. Interpretation Key for Visual Interpretation of predominant vegetation types
Table- 2. Classification Accuracy Assessment of predominant forest types using field sample points
The low mapping accuracy in case of digital classification approach compared to visual was due to the mixed pixel problem. In certain areas on the image there was spectral overlap between vegetation classes, which could not demarcate digitally the predominant evergreen class from semi evergreen, and semi evergreen from southern hill top evergreen (high altitude stunted evergreen forest). The reason for increase in moist deciduous forest area in digital classification is overlapping of semi evergreen, littoral and water pixels in few areas, while in visual, experience and knowledge regarding the distribution of the type made it possible to delineate the classes distinctly from each other.
The structure of forest mainly depends on the physiognomy (height) and species composition of forest type. The digital classification technique is primarily based on the spectral reflectance emitted by various representative species groups of forest on ground. In general the semi evergreen forest shows species composition of both evergreen and semi evergreen species. Depending on the topography, rainfall and soil types, the predominant top canopy species varies and either of the species may emerge as top canopy species group. The observed spectral overlap between evergreen and semi evergreen indicates that in those areas of semi evergreen forest the top canopy is formed by evergreen species, which could not be spectrally separable by digital method and similar is case with the moist deciduous forest.
The main reason that could be attributed for the increase in the area in visual interpretation when compared to digital method is a more precise delineation of the classes like Littoral, mud flats, plantations, mangroves and sand at the edges, along the coast and in between the islands. Visual interpretation was carried out using the knowledge of training sets provided by digital technique, field inventory, topographical features, location of forest types, drainage pattern which helped in producing a better vegetation type map of the study area. Both the methods proved to be better in delineating a homogenous core area patches and digital process failed in drawing clear patch boundaries at edges. The visual interpretation technique is advantageous in detecting spatial patterns and in drafting precise boundaries around relatively homogenous area while the digital methods typically operates one pixel at a time (Estes et al., 1983). Although the training areas were refined in digital supervised classification, there was no change in the seperability even with increasing or changing the number of pixels in the training areas resulting in the misclassification of pixels. Perhaps digital classification algorithm exclusively depends on the spectral reflectance of the ground features without considering topographic factors.
The comparison of visual interpretation with digital image classification demonstrates that the visual interpretation method allows the most detailed differentiation of structures and objects as obtained in the present study and thus the can be used best for detection of landscape changes. Improved digital classification algorithms, which use contextual approach and parcel-based classification, may give results comparable to visual technique. As the maximum interpretation accuracy achieved using digital image processing technique has been reported up to 70 to 80% (Jenson, 1996) there is a need of combining visual interpretation with that of digital to achieve high accuracy in near future. Jenson et al (2001) in their neural network based photo interpretation approach emphasized the use of hybrid approach i.e. visual interpretation and digital interpretation. In the present study also the combined knowledge of visual and digital techniques helped in refining the vegetation classes to produce more accurate vegetation type map of the study area.
Patches – Environment Variability – Species Richness – Diversity
The buffer zone analysis showed that littoral and mangrove forest are restricted within 1500 m zone and beyond 1500 m one can find exclusively typical interior forest types such as evergreen, semi evergreen and moist deciduous. Evergreen forest dominated in 2, 3 4 and 5th zones and moist deciduous decreased as proceed towards interior. This is sign of existence of wet evergreen on high altitudes and moist deciduous on low lands and base of hills. Semi evergreen though present in all five buffer zones but its dominance is observed in zone 1 (>1500 m) on par with mangroves which forms important representative community of zone 1. With the changing forest structure species composition also changes paving way for the thriving of different species community within a small area (Fig: 3).
Table – 3. Comparative Total Area Statistics
Large patches of size greater than 1000 ha were recorded more in semi evergreen (6) followed by evergreen (5) and in moist deciduous all patches are of size less than 1000 ha (Table-4). Evergreen patches distributed more towards the southern side and patch number and size decreases as moved towards northern side, whereas for moist deciduous the case is reverse. This type of forest is concentrated more towards the northern side and patch number decreases when moving towards south. In case of semi evergreen, patches are distributed throughout the islands and large patches are concentrated mostly towards northern side of islands. The main reason for the distribution of two forest types i.e. large patches of moist deciduous towards northern side and evergreen patches towards southern side is due to the variation in the rainfall pattern. In these islands the annual rainfall varied from 2200mm to 3500 mm. Areas lying towards the southwest and west of the hill ranges receive a greater amount of rain showing typical evergreen vegetation and as one proceeds towards north the rainfall goes on decreasing, showing typical moist deciduous vegetation.
Table-4. Distribution of forest patches and their area statistics in each forest type
The field collected phytosociological data analysis showed semi evergreen forest as high species rich and diversified community followed by evergreen and moist deciduous (Table-5). Forest patch size and number probably plays an important role in quantifying species richness and diversity parameters. Patch characteristics for each forest type reveals that patches of size less than 50 ha were recorded more in moist deciduous covering an area of 105 sq.km and patches of size more than 200 ha and even the largest patch of 6311 ha with an area of 450 sq.km was recorded in semi-evergreen forest. In general the number of species normally increases with increasing patch size but the kind of species found varies with the patch size. Since high species diversity in a community could be due to the coexistence of large number of species within a homogenous patch, presence of more number of large size patches (>1000 ha) in semi evergreen forest make it highly diversified inhabiting high species number with in the intact large homogenous patch. Absence of large patches of size >1000 ha and presence of more number of small size patches (
