Land Cover Map of West Asia
Using 1-Km AVHRR Data Hussein Harahsheh and
Ryutaro Tateishi Abstract Center for Environmental Remote Sensing (CEReS), Chiba University 1-33 Yayoi-Cho, Inage-ku, chiba 263, Japan Fax: +81 -43 290-3857 E-mail :tateishi@rsirc.cr.chiba-u.ac.jp, Hussein@rsirc.cr.chiba-u.ac.jp NOAA AVHRR data are regarded as one of the important tools that have been used in the exploration of natural resources especially vegetation resource at the regional and global levels. This paper addresses the fact that NOAA AVHR (30 second) could be used to develop land cover maps at regional level, which would lead to a balanced development in arid and semi-arid regions. This study falls into the following parts. (1) Cluster analysis of the NDVI 10 days composite data set for one year in total 36 data set. (3) Histograms analysis. (3) Ground truth collection and analysis. (4) Applying a supervised classification rules in form of three structure. As a result three of maps were produced. (1) Ground truth map contains all of the collected ground truth data. (2) Ground truth information sources map with text describing the source of each ground truth. (3) As a final result a 30 second Land Cover map of west Asia was produced, this map will be necessary as input data for a future desertification studies. Introduction The main purpose of this study is to produce land cover map of west Asia and demonstrate the role of NOAA AVHRR (NDVI 10-days composite data set for one year in total 36 data set and visible channels 1&2) data for vegetation monitoring and land cover classification at regional scale in arid and semi-arid environments like west Asia region. West Asia generally characterized by arid climate, the southern part is extremely arid. However high precipitation occurs in coastal mountain ranges and in the extreme north and north east parts. The eastern Mediterranean countries are influenced Mediterranean frontal depressions, the average precipitation in this region ranges from 1500 mm to 70 mm. Egypt climate is extremely arid and average precipitation is about 10 mm. Arabian peninsula generally characterized by a ht dry climate. The morphology general of the study area is flat, though relatively narrow mountain ranges extend along the coastlines of the Red sea, the Mediterranean sea and the Gulf of Oman. Mountains and plateaus are found mainly in the north, they are mostly low relief features. West Asia region is intensive use of ground water which leads to land salinity of soil. Most of marginal lands in west Asia are permanent pastures and 85% of them are consider in danger to desertification. These marginal land are susceptible to inappropriate land use practices, such as overgrazing, fuel cutting and inadequate cultivation. Vegetation degradation is widely found serious in these dry marginal lands. This study falls into the following parts: 1-Cluster analysis of the NDVI 10 days composite data set for one year in total 36 data set. 2-Histograms analysis. 3-Ground troth collection and analysis. 4-Applying a supervised classification rules in form of tree structure. Data and Materials The base data used in this study are the International Geosphere Biosphere Program (IGBP) 1- km NOAA AVHRR 10-days composite data sets for April 1992 through march 1993. While the 10-days maximum normalized difference vegetation index (NDVI) composite data are sued to classify the vegetation land cover features, the 10-days composite data set of the period 1 to 10 of June 1992, channel 1 and channel 2 are used t approve the discrimination of non-vegetation land cover features. Resurrs -1 data with it's wide swath width (600 km x 600 km) and medium spatial resolution (170m) is bridging the gap between AVHRR 1m resolution and Landsat TM. This data can be used for environmental monitoring of large areas and for agricultural classification. In our case the available images cover just a small part of the study area, so we used this data to defined ground truth data and in the visual interpreation at the post -classification level. And we use the several maps as resources information and ground truth collection data such as Land use map of Jordan, scale1:600,000. and Land cover map of Syria, scale 1:1,000,000. These data and others wee scanned and registered as image to image to the raw NDIV AVHRR 1 km data to fit the same geographic corrdinate of AVHRR data. Pre-Processing of AVHRR data The global 1 km 10-days composites AVHR data processed on Interrupted Goode Homolosine map projection (36 NDVI 10-days composites bands) were transformed into plate carray projection (latitude and longitude coordinates system). This transformation will locate easily the ground truth data and the usage of latitude and longitude corrdinates system is very practice and clear to view. We used this corredinates system to extract sub-images (of the 36 NDVI 10-days composites bands) representing the study area, the corrdinates of upper left corner are 40 degree northern latitude and 25 degree eastern longitude, the corrdinates of lower right corner are 10 of June 1992, channel 1 and channel 2 were extracted through the home page of the US geological Survey's (USGS), then these two channels were registered to the extracted NDVI data to fit their corrdinates system. As much noise is involved in NOAA AVHRR data, noise free NDIV compiled on monthly base should be used. This requires that the 36 NDVI 10-day composites data sets extracted for the study area are recomposed into 12 NDIV monthly composites data sets. This recomposing was generated using an algorithm based on the maximum NDVI value of the month. Pre-Classification In order to study the realationshp between the different values of NDVI through the four season of the year, four histograms of NDVI monthly composites data representing the four seasons were created, other four histograms were created using the vegetation cover mask. We conclude that January has lowest amount of vegetation. April period has an important vegetation activities in the north of the study area, while high vegetation activities appears in southern part during the period of October, which is justified by the sub-tropical climate, the period of July has a medium vegetation activities especially in southern and northern parts, and few or non vegetation activities in middle part of the study area. Knowing that the study area cover a relatively large surface 4200 km X 3600km, there is insufficient knowledge exist about its, and spectral signatures show high variability, so the unsuperivised classification or clustering is a very useful tool in this case. To achieve this classification, we used the K-Means algorithm. We obtained about 15 clusters which should be interpreted in terms of land cover classes. For his step of inerpretation we used the preliminary labeling to provide a general understanding of the characteristics of each cluster and to determine which classes have two or more disparate land cover classes represented within their spatial distribution. Table (1) describes the result of labeling. This result gives us a primary idea about the land cover types and their characteristics, this will be taken in consideration in the supervised classification.
Ground truth data collection and analysis As the collection of ground truth data is a time consuming as expensive, it is decided to collect the ground truth data from the existing land cover maps described in previous paragraph, especially we consider that land cover categories do not change so much at regional level. From the thematic data and Resurs-1 imagery we collected about57 ground truth region which represent thirteen land cover classes. To keep these data with their geographic location (latitude and longitude), a "ground truth map" was created containing thirteen classes of land cover in form of ground truth. The total ground truth regions were coded by numbers, then "ground truth information sources map" was created according to these assigned codes. A test file containing information about the sources of ground truth regions was attached to this map. It is worth while to be noted that this map will be of great utility as reference information. The analysis of ground truth data consist of creation of diagrams representing the variations of NDVI values through the year for each ground truth data separately. So thirteen diagrams were created to analyze the data. Table (2) summarize the analysis. From this table we got a lot of information. The maximum NDVI values of vegetation ranges from around of 0.13 to 0.33, the other types of land cover ranges from 0.05 to 0.1, sabkhat land cover has a very small value of NDVI around 0.01. We think that the maximum NDVI value related to sand dune (0.1) is due to reflectance forest have the same range of maximum > 0.30, this is reflect the defect of NDVI to separate the intensive cultivated area from the forest areas, this will produce some concussion in the classification. In general the main period of vegetation activities is April to July, except for irrigation, there are tow periods of vegetation activities, one from early March to end of May, and the other one from early July to end to August. The grass and shrubs land cover class has a low NDVI value comparing with other types of vegetation, we think this is due to the type of shrubs in such dry area, the maximum activity of this class is during October. The semi-dry land covered by basalt and cherts stones have small peaks especially during April, after a rainfall period, these small peaks are targets of the Bedouin people moving in the desert and looking for the bushes for their animals.
In general all land cover types are of the same range of minimum NDVI values below 0.1 except for irrigated areas, which is independent of the rainfall and continue through the year. The period of the minimum vegetation activities changes in function of land cover type, but generally from November to February. The diagrams of channel 1 and channel 2 shows a good variation between ground truth data, especially the non vegetation land cover features, so these two channels were selected as additional data to help in discrimination on non vegetation land cover classes. Land cover map of west Asia In the frame of preparation of supervised classification a yearly vegetation cover map (Vegetation cover mask) should be created. To attain this purpose our methodology consist of obtaining the vgetation cover of each month of the year using the monthly NDVI data sets, then adding in an accumulative manner the 12 monthly vegetation cover data. We called this summation of vegetation cover, because it represent the minimum NDVI value without confusion between vegetation and other non-vegetation classes especially sand dune feature (see table 2), so this value include all kind of vegetation activities from the low density range land to the dense forest. The collected ground truth data representing the land cover classes were used to generate signatures, each one of these signature correspond to one class. The generation of supervised classification using NOAA AVHRR data, may have a problem of confusion of classes because of the low resolution of such data, so we try to overcome this problem, increase the separability between classes, assure an precise mapping accuracy and make the work more easy in adopting the tree structure methodoloty, which consist of dividing the work in form of tree beginning with one class and finish with all classes at the top of tree. On the first level we separated the water bodies and sea from the land using the monthly NDVI data set of January because this month appears the maximum water bodies before the evaportansporation, then the land class was divided into tow classes, vegetation class and non vegetation class, this separation was done using the vegetation cover mask, at the third level we used the supervised maximum likelihood classification to isolate the forests from the other vegetation types, at the same level we isolated the sabkhat which are kind of wet land, from the non vegetation class to find out the bare land class this achieved by visual interpretation of NOAAA AVHRR visible channels and the thematic data. In the fourth level we applied the supervised classification on the non forest mask to establish the folliwing classes : 1-range land 2-grass crop 3- irrigated areas 3-grass and shrubs. In the same manner we applied the supervised classification on the barren land class and found out the following classes; 1- sand dunes a 2- sand b 3- semi-dry land-basalt stones cover 4-semi-dry land-chert cover 5-dry land(desert area) 6- rocks area (granite,…). As final result we have the following land cover classes with the system codes adopted by the working group " 1-km land cover data base of Asia". Table (3)
Conclusion This work illustrates the potential of NOAA AVHRR data and other types of land cover satellite derived information for regional land cover assessment and mapping, and demonstrate the value of this data as a major source of information for mapping "Hot Spot" areas which should be identified for further in investigation with higher resolution data and field verification. We think that NOAA AVHRR data and other types data provide invaluable, very timely, cost-effective and objective information which is to be used for natural resources management and planning, and without such satellite data this type of work would not be possible. The availability of other sources of information like Landsat and Resources -1 data are utmost importance to improve the classification result. |