THE USE OF ERS DATA FOR MONITORING OF LAND COVER CHANGE IN SAHELIAN REGIONS THE EXAMPLE OF THE DELTA OF THE NIGER RIVER (MALI)

Transcription

1 THE USE OF ERS DATA FOR MONITORING OF LAND COVER CHANGE IN SAHELIAN REGIONS THE EXAMPLE OF THE DELTA OF THE NIGER RIVER (MALI) Catherine Mering (1), Yveline Poncet (2), Siegried Hess (1), Elmar Claplovics (3) (1) University of Paris VII, UFR GHSS, 2 Place Jussieu, Case 7001, Paris Cedex 05, France mering@lgs.jussieu.fr (2) Laboratoire ERMES, IRD, 5 rue Carbone, OrlŽans Yveline.Poncet@orleans.ird.fr (3) Department of Earth Sciences, University of Dreden, Mommsen Strasse 13, D Dresden, Germany INTRODUCTION Vegetation analysis in semi arid regions is in high demand and remote sensing applications are well established. In the field of change detection of Sahelian landscapes, one has to inventory states of vegetation at different recording dates. The Inland Niger Delta (Mali) is an azonal hydrographical landscape complex within the semi arid West African zone of the Sahel with water available through all the year. The changes in the land cover is due to pre-flood and post flood hydrological conditions of the Niger, which depends to seasonal and annual pluviometric events. Our purpose is to study different ERS images on the same area, in order to survey changes in the landscape between different seasons and different years. One of the advantages of using ERS data upon Landsat and Spot images for studying the inter seasonal and interannual changes over this large wetland, is that the view of the ground is no more obstructed by the frequent clouds and mist as it is usually on passive remote sensing images. In order to observe with accuracy annual changes as well as seasonal one, we have selected a set of ERS images in the region of Mopti, taken in the pre-flood and post flood periods between 1992 and We first have reduced the speckle on ERS images by means of appropriate filters in order get a more accurate vision of change. This changes are detected with colour making images by RGB techniques from three dates images. On the resulting images, major variations in the back scattering of landscape units will appear as coloured flat zones. With the help of ground truth and aerial photographs of some wellknown areas, the coloured zones are interpreted whether as flood, or as changes in the vegetation cover. PRE PROCESSING OF ERS IMAGES The change detection techniques are based on visual or numerical comparison of radiometric values of the same spatial entities on several images. On such images, grey tone regions have then to be sufficiently homogeneous to be compared. One has then to reduce the speckle of each image, in order to be able to interpret the change of radiometric value as a change in the surface state. Therefore, in order to analyse changes of the landscape in the Inland Delta of Niger from ERS images taken at various dates, one has first to reduce the speckle on each image. On radar images, the speckle noise is one of the main obstacle for digital processing. These images do not contain zones with spatially uniform reflectivity : for a given grey level threshold, the pixels are generally not contiguous. In order to reduce the speckle, we assume that the speckle effect is produced by isolated pixels having a grey tone value drastically different from their neighbours. There are many ways to build filters that reduce the speckle effect by smoothing the image in a given neighbourhood whatever is the value of the central pixel. Smoothing techniques are applied here upon an extract of the ERS scene of May 96, in the south of the Delta (Fig. 1). One well known method for local smoothing is the Median filter. Local smoothing can also be obtained by Morphological filters which are based on morphological transformations of grey tone functions, that is, Morphological Openings and Closings with a given structuring element. If utilised alone, a Morphological Opening (resp. Closing) smoothes only white (resp. dark) details. However, on grey-tone images such as radar ones, the noise may be composed by both dark or white details. Smoothing filters must then act the same way on high and low values. Such filters are called self dual filters. The median filter is self-dual but it is not idempotent; it

2 means that its efficiency is proportional to the number of iterations, and that moreover it does not always converge. We have used here self dual and idempotent filters built from well-defined defined sequences of Openings and Closings Given a Morphological Opening γ with a convex structural element B and the dual Morphological Closing ϕ, it has been proved that f= ϕγϕ is a -overfilter and g = γϕγ is a - underfilter filter. One thus can calculate the center β of f and g. (Fig.2). This filter is a self dual and idempotent filter [1]. Moreover, contours and connectivity of the grey tone regions have to be preserved for an available comparison of a chronological set of images. Connected filters on greytone images product grey-tone flat zones, and restore therefore the connectivity of greys-tone objects. This property have been used for image segmentation [2]. From the basic Erosion and Dilation, one can deduce Connected Filters such as Closings and Openings by Reconstruction. A Closing (resp. Opening) by Reconstruction is obtained by iteration until idempotence of an Erosion (resp a Dilation) of size one from the initial Dilation (resp Erosion) of the original image. It is then possible to obtain - overfilters and - underfilters from inf and sup composition of Openings and Closings by Reconstruction. We have used this property here to build a connected center filter βc [3] from the definition of an overfilter fc and an underfilter gc deduced from the Closing by Reconstruction ϕc and the Opening by Reconstruction γc. Given the two filters γc and ϕc which are respectively an Opening and a Closing by Reconstruction, we applied to the original radar image a filter which is the centre βc between fc and gc defined as follows: β c = ( I f c ) g c where f c = ϕ c γ c ϕ c and g c = γ c ϕ c γ c One can see on Fig. 2 that the speckle has been partially removed and that the connectivity of the grey-tone structures have been restored. Fig. 1 Extract from the ERS scene of 16th June 1992

3 IMAGE ANALYSIS FROM GROUND KNOWLEDGE Fig. 2 Connected Center Filter on image of Fig. 1 Our purpose is to study different ERS images on the same area, in order to survey changes in the landscape between different seasons and different years. In this Sahelian region, drastic changes in wetness occur between the high water season (October) and the low water season (may) in this seasonally large flooded area with an extent of square kilometres from south of Djenne to Timbuktu [4]. In order to establish a set of interpretation elements of the Inner Delta images, we have established a link between of natural and anthropic landscape features and seasonal changes. Months Hydrology Weather Productions April-May-June Low and very low waters Very dry and hot season Grazing and fishing July-August-September Arriving of the flood from the South (upstream), the flood invades all the plains and low areas Rainy season (sometimes rain begins in June) vegetation October-November- December January-February-March Low decreasing of the flood, letting wet the low areas Low decreasing of the flood, every part of the delta is drying Dry and warm season (sometimes rain ends in October) Dry and cold season Table 1 Seasonal major events in the Inland Delta of Niger The rice cropping is beginning (end of tillage, sowing), growing of every Full period of herding, yielding, fishing Full period of herding, yielding, fishing We have then selected images taken during three hydrologic moments : the very beginning of decreasing water (November-December), the two third of decreasing water (February) and the lower waters (April-May). The hydrologic and climatic year is ordinarily divided, in the delta, in four season. The production systems (herding, rice cropping, fishing) may be simplified and described according to theses four seasons (Table 1). The hydrologic year (for measurements and comparisons) begins the first of May. One has to notice that in 1994 occurred an important hydrologic and climatic event in the whole West Africa : the succession of dry years (known as

4 The Drought) ended in June, when happened a spectacular rainy season, accompanied with high and long-lasting flood. In spite of a slight decreasing of pluviometric and hydrologic resources, the following years have been said to be «normal» considering the sequence (dry) and the ante 1973 means («normal»). The physiographic entries of Table 1 are selected according to the ground knowledge: the homogeneity of the landscape objects and the changes of features are observed before and after the rainy season of As a matter of fact, this peculiar year is considered as a "threshold year" between generally dry years and generally «humid» or rather «normal» years. Three hydrologic moments are selected: the very beginning of decreasing water (November- December), the (approximately) two third of decreasing water (February) and the lower water (April-may). However, a lot of features or objects are not very well identified with regard to the sigma 0 values : their size, their shape and orientation must be more precisely studied in the future, to establish more certain similarity and difference, i.e. more precise basics of interpretation. Moreover, due to the lack of identifiable catchments in this flat and flooded country, two very important variables are not known nor mastered, neither on the field nor on the available maps : the soil humidity (quantity and depth) in relation with its granulometry ; the features of the very low relief (no detailed orographic maps in this area). We have to take into account that the whole area, flooded or nor flooded, is a very complicated mosaic of various soil, water, vegetation and land use patterns, with complex evolutions from season to season and from year to year under natural and anthropic actions. For this reasons, an interpretation of the ERS images of the Inner Delta is attempted for small scales only, i.e. corresponding with large «regional» areas and no «local» details. In a first approach, we study the images without any enhancement or classification [5]. We have made a visual interpretation, selecting first the directly observable objects (tone values, recognisable position and shape). Lakes, channels, large ponds (i.e. clear water) belong to this category which is not really useful since we already have precise maps of water features. The interesting part is the different tones in the lake Debo, indicating the large sandbanks in its northern part, and their different sizes on the different images. We think that highly enhanced images would show the sandbanks lying in the Niger channel (which are a nuisance for navigation) and (providing a large set of images is available) their shifting movement upstream. All the features taking place in the (more or less) dry lands (that is to say the less low areas in altitude) are much more difficult to identify and, subsequently, to interpret : the large scale mosaic of their different physiographic features seems to be enhanced in the ERS images, relatively to the aerial photographs. We can see that the low values of reflectance correspond with interesting precision with the wetter grounds : the whole southern delta as delimited on available maps and the central and lower parts of this delta on dates images. More precise examination, however, shows unexpected tones, limits and features. We are particularly interested in the permanent features (as seen on every image) and their changing limits (see rough map). For most of them, we can only make hypothesis. - A large permanent feature (but on the date image), in the shape of a large crescent ( south-west of lake Debo, in the clear tones (high values of reflection). This implies rough and/or dry surfaces. The crescent is stangely positioned, inside the regularly flooded perimeter. From ground survey, we know that it is grassland scattered with sand dunes, these covered by millet crops from July to October. During the other months, only remain stubble, then small furrows. CHANGE DETECTION FROM RADAR IMAGE ANALYSIS As we possess different ERS images recorded at two different seasons of one same year (1992) and different ERS images recorded before may 1994 and after may 1994, we are allowed to compare the physiographic features taking place : - during the dry season and the wet season of a generally dry same year (1992) : June 16, 1992 (Fig. 3); November 3, 1992 (Fig. 4); - during two different hydro-pluviometric years : (a dry one) and (a «normal» one, much wetter than 1994) : May 5, 1994 (Fig. 5); February 29, 1996 (Fig. 6). It happens that we can compare the tones and the forms in some quadrants of these images with observations made on the field at approximately convenient dates : Approximate correspondence between dates of ERS images and observations on the field. Some observations on the field [6] are directly useful because of the coherence with season and year ; others are useful only by comparing two or three different situations, before and after the recording of the image for instance (Table 2).

5 Fig. 3 ERS Scene on the Delta (16th June 1992) Images Field Field (comparative) Places June 16, 1992 (Fig. 3) March-April 1993 June 1996 The whole delta, Lakes Korientze, Debo and Walado November 3, 1992 October 1992 The whole delta, (Fig. 4) May 5, 1994 (Fig. 5) February 29, 1996 (Fig. 6) March-April 1993 April-May 1995 The whole delta, Korientze, Walado, The northern Diaka, The Batamani area March 1996 Lakes Korientze, Debo and Walado, The northern Diaka Table 2. Correspondance between ERS images and field observations in the Inland Delta of Niger Comparing the two images of 1992 (Fig. 3 and Fig. 4), we look for : - elements of interpretation of the tones of the images, and explanations for the anomalies, - particularly interesting geographic objects, in order to study their changes from date to date or to find new informations about them on the images. The June image is not very contrasted (Fig. 3). The clear water features (River Niger, lake Debo, Lake Korientze, a very small lake Walado) are recognisable by morphology and position, in the lighter tones. We can hazard that these water are very low and turbid at that moment. The whole terrestrial area is dry, dark on the images, with white small rounded areas which are ponds, either retaining water (permanent ponds) or constituted of the smooth and fine clay of their dry beds. The darkest features are not easy to interpret : - a. most of them seem to show close to the lakes (large areas at the limit of lake Debo, a small area in the middle of lake Walado),

6 - b. others may be clearly distinguished as straight lines in the northern part, the sand dunes, - c. a lot of very dark spots are scattered in small more or less contiguous groups along channels, and often seen close to small white areas. Fig. 4 ERS Scene on the Delta (3rd November 1992) Hypothesis, according to the year and the season : - a are flat, smooth and dry sand banks or clay banks, without vegetation, or with very scarce and low grass (overgrazing is common in June). - b are the «dunes» themselves : bare surfaces ready, at the moment, for receiving the seeds of pluvial crops (sorghum and millet). - c are likeable bare, dry and newly tilled areas, ready to be sown in rice («peasant rice», sown in dry soil before the arriving of rain and flood). The middle tones of grey draw large features underlining different parts of the active delta. These parts are more clearly visible, with almost the same limits, in the image recorded in November of the same year (Fig. 4). a. North-West and North : dry and almost dry soils, vegetation and crops : poor. b. Centre of the image and Diaka surroundings : dry and almost dry soils, vegetation and crops : poor or nil. This has been observed on the field, particularly around the DembŽ river. This area is constituted of alluvial banks of different ages and kinds of material (ranging from clay to sand) with more or less dense shrubby vegetation. In this area, the Rogonta sand alluvial dune and the «Ferou de Dialloube» may be easily distinguished and recognisable as test areas. c. A clearly lighter area between a and b seems to be the more «humid» area, but the explanation of this relative wetness is not easy : low areas (flood plains), with rice crops not yet yielded. In November, the flood (even a poor flood, in 1992) is not yet completely retired. But we can notice some anomalies : The lake Korientze and the lake Debo (which limits are blurred) are black. The channels (Niger, Mayo Raneo, Dembe, Diaka) are «medium grey». Large white and somewhat blurred areas appear neatly, with no clear explanation available. The clear crescent west of the lake DŽbo, particularly, is present on every image : it has not been noticed in the landscape as a relief or a soil feature, but it could be compared with the knowledge of local fishermen who have reported «ancient channels» and the reappearing of (infiltrated?) water in this area (field inquest, Poncet, 1996). The «Walado crescent» described above is one of them. Another interesting shape is the different tones of lake Debo, featuring different types of flooded or previously flooded areas. On an size enhanced image, 16/06/1992 one can detect a large expand of clear water (light tones), connected with

7 tributary channels, particularly East, West and South of the lake. All around this clear water, the darker and homogenous areas of dry and bare sand or mud. In June 1992, before the seasonal flood and before the arrival of the first seasonal storms, the lake is near the lower level of water ever seen. North of these sand and mud banks, a middle grey homogenous area figures a shallow water area, covered with burgu (which is similar to a reeds covered landscape). This middle grey can be found south of the lake, in less homogenous areas. Fig. 5 ERS Scene on the Delta (5th May 1994) On the whole image, the channels seem more easily readable than on other remote sensing images, particularly, of course, the narrow channels more or less covered with dense or scarce (depending on season and hydrological maximum of the year) vegetation. The relief of their banks underlines them, even when their is no water. But this fact can bring confusion between two sorts of «banks» : the functional ones of the regularly functional channels, the ancient ones of Pleistocene channels, as in the case of the Ferou de Dialloube (Fig. 3). We attempt to get synthetic results about both interseasonal and interannual changes. For this purpose we have computed a Karhunen Loewe Transformation from three images: These three images were chosen by the following way: The two images of 1992 (June and November) were chosen to detect major seasonal changes during this "dry year". As already said above (Table 2), June is at the end of the dry season, and thence of the lower flood, as November is at the end of the wet season, when the flood is decreasing letting wet the low areas. In order to observe on the same image interannual variation, we have selected the third image as the February image of 1996, that is an image of the dry season of a rather "normal" year. By analysis of the contribution of each component: the following results are found: the image of November 92 contributes for 0, 90% to the first component. the image of February 96 contributes for 0, 90% to the second component. the image of June 1992 contributes for roughly 0,80% to the third component. In order to read from one single coloured image the synthesis of these results, we allocate to each component blue, green and red colour according of their original order (Fig. 7). The blue colour corresponds here to the flooded part of the Delta during the end of the wet season. On Fig. 7, one can see that the central part of the Delta as well as the western link of the Niger appears mainly in blue: it shows clearly that, even during a "dry year", this central part of the Delta is flooded. The green colour corresponds to important changes between 1996 and the two other dates of 1992: there probably have been more changes in the landscape between both dry seasons of and of 1996, than between the humid and the dry

8 seasons of the As shown on Fig. 7, the green colour is prevailing on the western part of the Delta, which could mean that the landscape in this peculiar region has changed a lot after Fig. 6 ERS Scene on the Delta (29 th February 1996) Finally one can observe that the red colour prevails on the northern part of the Delta: the regions coloured in red (mostly regions around the Lake Debo) are probably the one where there were drastic changes in 1992 between the end of the wet season in November and the dry season in June: the northern part of the Delta must have completely changed from flooded state to a completely dry state. This preliminary results leads to general conclusions about interseasonal and interannual changes. In some extent, one may say that climatic changes observed during 1994 may have some consequences on the transformations of the landscape in the Delta. In order to produce more precise results, one has however to get ERS images on the same date for the three years studied, and for present years. Hypothesis about seasonal and interannual changed could then be more accurately be studied, and pattern recognition methods could be employed not only to localise the changes in the landscape, but also to understand their origin. REFERENCES [1] J. Serra, Mathematical Morphology, Vol2, Academic Press, London, 412p, [2] J. Crespo, J. Serra, R.W. Schafer, «Image Segmentation using Connected Filters», Mathematical Morphology and its Applications to Signal Processing, pp , May [3] Mering C., Parrot J.F. : "Radar image analysis using morphological filters", Mathematical Morphology and its application to Image Processing, J. Serra and P. Soille Eds, pp , [4] S. Hess, J.C. Bergs "Quantification of wood cover dynamics by computer-assisted analysis of aerial photographs in the Internal Niger Delta (Mali)", Zbl Geol. Palašnt Teil I, pp 1-18, Stuttgart, Mai [5] Cslaplovics E. "GIS for local to regional scale desertification monitoring and assessment in the West African Sahel. Proceedings on the conference Africa GIS'97, Gaborone, 1997, pp [6] Quensire J., Poncet Y."La džcentralisation Malienne : quelle prise en compte de l'organisation Žcologique et sociale de la pche artisanale", Synthse Finale de l'a.i DURR 21p. (to be published).

9 Fig. 7 Coloured Composition from three decorrelated images (3/11/ 92, 16/06/92, 29/02/96)