ESA Earth Home Missions Data Products Resources Applications
EO Data Access
How to Apply
How to Access
Spatio-temporal analysis of SAR image series from the Brazilian Pantanal (Geoffrey M. Henebry Hermann J.H. Kux)
Spatio-temporal analysis of SAR image series from the Brazilian Pantanal
Spatio-temporal analysis of SAR image series fro
Site Map
Frequently asked questions
Terms of use
Contact us


Spatio-temporal analysis of SAR image series from the Brazilian Pantanal 

Geoffrey M. Henebry   Department of Biological Sciences, Rutgers University, 07102 Newark, NJ, USA 
Hermann J.H. Kux   Instituto Nacional de Pequisas Espaciais (INPE), 12227-010 São José, dos Campos, SP, Brazil


Flood monitoring in the Pantanal Matogrossense is complicated by several environmental factors that interact to generate landscapes that appear highly heterogeneous in time and space. In light of this heterogeneity, change detection and quantification become difficult: How best to measure spatial heterogeneity through time? What constitutes an appropriate baseline? Our solution uses various metrics of lacunarity to quantify land cover dynamics. We illustrate this approach using a seven-date SAR image series from 1992-93 in which a significant climatic drought follows typical seasonal inundation. Lacunarity analyses on three representative landscapes from the Nhecolândia region reveal distinct spatio-temporal trajectories that include both cyclicity and perturbation in spatial heterogeneity and anisotropy. These features correspond to the flooding cycle and the unusual drydown event. 



Keywords: landscape dynamics; flooding; wetlands; Pantanal Matogrossense; lacunarity; land cover change 


The utility of Synthetic Aperture Radar (SAR) for flood monitoring has been amply demonstrated in the past few years (e.g., ESA, 1995). This application has focused on the occasional catastrophe that affects human settlements when hydraulic loading on a regulated river system exceeds design specifications. Land cover change accompanying such a flood is obvious albeit ephemeral. In the absence of human intervention, periodic inundation of riparian lands is required to sustain the patterns and processes that characterize river-floodplain ecosystems (Junk et al., 1989). Operational flood monitoring of an unregulated river system presents a set of analytical challenges to remote sensing: How best to summarize the spatio-temporal data of an image time series? What constitutes a baseline? How to quantify spatio-temporal patterns to enable comparisons and predictions? These questions require an analytical approach that can quantify temporal development in the spatial structures that compose an imaged scene. This kind of explicit spatio-temporal analysis contrasts with the implicit analysis found in a sequence of classified images. Indeed, the dynamics found in biogeophysical fields are ill-represented, both conceptually and practically, by map series.

One promising approach to identification and portrayal of patterns latent in spatio-temporal data is lacunarity analysis, a multi-scale procedure based on fractal geometry. Here we introduce new lacunarity metrics and a neutral model that facilitates comparisons of lacunarity decay curves. We apply these techniques to a seven-date ERS-1 SAR image series to discover spatio-temporal landscape trajectories found within a very dynamic tropical floodplain. This study extends our previous work on lacunarity analysis of SAR imagery (Henebry and Kux, 1996; Henebry and Kux, 1995; Kux and Henebry, 1994a; Kux and Henebry, 1994b).

Lacunarity Analysis

Lacunarity indices use multi-scale windowing to measure the scale dependency of spatial heterogeneity and anisotropy in binary maps in terms of departures from translational and rotational invariance (Plotnick et al., 1996; Henebry and Kux, 1995; Plotnick et al., 1993). The indices are sensitive to map density and aggregation. Higher lacunarity indicates a more sparse, more clumped distribution within the map. Random maps exhibit a lack of persistent spatial structure under multi-scale windowing (i.e., correlation length approaches zero) and thus achieve low lacunarity scores. Conversely, maps containing larger aggregates maintain high lacunarity scores until the size of the sampling window exceeds the size of the aggregates.

Interval-scaled imagery must be converted into binary maps for lacunarity calculations. Slicing the image histogram into even quantiles controls map density, thereby making lacunarity sensitive only to the scale dependency of aggregation. For image time series it is useful to track lacunarity using a constant window size. Choosing the optimal window size, however, can be tricky: sampling too large or too small an extent can miss significant changes in spatial structure. It is thus prudent to conduct preliminary analyses to determine the shape of the lacunarity decay with occurs with increasing sampling extent. Comparing decays curves is rather difficult, whether it be quantiles formed from the same image or comparable quantiles across different image. We develop here a simple neutral model to facilitate comparisons.

Assume that the image histogram is sliced into quartiles and four corresponding binary maps are formed. No pixels are lost in this conversion: their values are simply collapsed to one bit. A convenient property of the basic lacunarity index is that its value at a window size equal to the image grain is the reciprocal of the map density. Therefore, the lacunarity indices sum across the quartiles to unity, when evaluated at window size of 1. This additivity forms the basis for a neutral model. We can pose an expectation that this uniform partitioning of lacunarity across the quartiles will be preserved with increasing sampling extent. At any sampling window size, we expect that the lacunarity index for any particular quartile will contribute 25 percent to the total lacunarity. In other words, we are positing a scale-invariant partitioning of lacunarity as the neutral expectation. Percent deviation from this fractal geometric expectation provides a metric that is more sensitive to change in spatial structure and easier to assess than raw decay curves. Furthermore, since there is "conservation of spatial order", i.e., no loss of spatial elements, total deviation across quartiles must sum to unity and the quartile-specific partitioning of deviation provides a means to distinguish between the textural aspects of scene object backscattering and speckle noise.

Study Area

The Pantanal is the largest wetland habitat on the planet: an immense assemblage of alluvial fans formed during the Pleistocene, it covers 139,000 km2 in Brazilian states of Mato Grosso and Mato Grosso do Sul (Klammer, 1982; Rizzini et al., 1988). The Pantanal is alsoo ne of the more radiometrically dynamic landscape in the tropics due to extensive seasonal flooding by the Paraguay River and its tributaries. An ecotonal landscape that developed during the Holocene, the Pantanal is a complex mosaic of shallow lakes, periodically inundated grasslands, and islands and elevated corridors of forest, which together support an abundant and diverse fauna of birds, fish, reptiles, and mammals, including four million head of cattle (Alho et al., 1988).

The Pantanal has a tropical semihumid climate with mean annual temperature of 25o C and mean annual precipitation of 1100 mm concentrated into a rainy season from October to March (Alho et al., 1988; Rizzini et al., 1988). Altitude in the alluvial plains range from 100-200 m asl. The Pantanal can be broadly classified into three subregions according to the degree and duration of flooding as determined by local topography: (1) the Alto Pantanal, the relatively higher elevations where about 20 percent of the area floods to depths of 30-40 cm for two to three months per year; (2) the Médio Pantanal, a transitional zone where more extensive flooding last from three to four months; and (3) the Baixo Pantanal, the low-lying areas where little topographic relief translates into almost complete inundation to depths of 3-4 m during the rainy season (Adámoli, 1986; da Silva, 1986; Alho et al., 1988; Rizzini et al., 1988).

The Pantanal is remarkable for its reduced declivity (2.5-5.0 cm/km). Local topographic features (2-4 m above the surrounding lands), resulting from either ancient Aeolian sandfields (Klammer, 1982) or termite activity (Ponce and da Cunha, 1993) are critical for determining habitat for both flora and fauna. Forested ribbons of higher ground known as cordilheiras are never flooded and serve as seasonal refuge for terrestrial animals. Where the water is deep, hydrophytes predominant; in areas with sufficient water flow, productive grasslands emerge (Rizzini et al., 1988). Highly permeable soils in the Pantanal lead to a substantial seasonal drydown (June to September) that favors xeric species on elevated soils (Prance and Schaller, 1982).

The study area lies in the region of the Pantanal called Nhecolândia, located along the southern tier of the Rio Taquari alluvial fan. Nhecolândia is remarkable for the hundreds of freshwater and saline lakes that punctuate the landscape. We analyzed three typical landscapes: a large quasi-perennial wetland with bordering woodlands (Figure 1), a mosaic of lakes ringed by trees and interspersed among grasslands, (Figure 2), and a well formed channel with riparian forest (Figure 3).

Figure 1: Multi-date composites for flooded wetland landscape. Red (11/93) and blue (12/92) are constant for each image; from top to bottom, green is 2/93, 5/93, 6/93, 8/93, and 9/93. Images are © ESA-R 1992,1993.

Figure 2: Multi-date composites for lake mosaic landscape. Band assignments as in Figure 1. Images are © ESA-R 1992,1993.

Figure 3: Multi-date composites for riparian forest landscape. Band assignments as in Figure 1. Images are © ESA-R 1992,1993.


Our image time series spanned seven dates (Table 1). These acquisitions were well positioned to sample a single flooding cycle. Figure 4 provides a glimpse at the broader hydro-climatological context. The SAR images were georeferenced, ground-range projected, real-valued, 3-look digital data processed by INPE. The nominal ground resolution of these data was 25 m with a pixel spacing of 12.5 m in both range and azimuth. For each landscape type the same scene (1024x1024 pixels = 164 km2) was extracted at each date and coregistered to the corresponding December 1992 scene using linear offsets. The resulting misregistration was minimal (<2 pixels) and lacunarity is robust to misregistration errors when image extent is large relative to resolution.

The quartiles (Q1, Q2, Q3, Q4) of the histograms were calculated and four binary images were thereby generated for each image subregion. In contrast to the random resampling approach of our earlier work, we calculated lacunarity using exhaustive subsampling separately for window shapes of w(2j) and w(j2), where j ranged from 4 to 1024, yielding sampling window areas from 0.125 ha to 32 ha. (For more detail on calculation of the lacunarity index, see Henebry and Kux, 1995; Plotnick et al., 1993). Lacunarity values for Q1 and Q4 were combined into one scaled index: SLI=2-(1/Q1LI+1/Q4LI). We also calculated the combined deviation in percent for Q1 and Q4 from the neutral prediction.

Date Orbit Number Center Latitude Center Longitude
1992 12 12 7369 S19o 12' 04" E304o 02' 06"
1993 02 20 8371 S19o 12' 40" E304o 01' 26"
1993 05 01 9373 S19o 15' 00" E304o 01' 19"
1993 06 05 9874 S19o 11' 42" E304o 02' 10"
1993 08 14 10876 S19o 11' 38" E304o 01' 52"
1993 09 18 11377 S19o 12' 14" E304o 01' 19"
1993 11 27 12379 S19o 11' 56" E304o 01' 55"
Table 1: ERS-1 SAR data from track 210 & frame 3987.
Figure 4: Monthly stage height of Paraguay River observed at Ladario, Mato Grosso do Sul from 1/84-12/94. Each X indicates a date in the ERS-1 SAR image series. 


We have noted previously (Henebry and Kux, 1995) that most of the spatial heterogeneity in these scenes was located in Q1 and Q4, corresponding to the lowest and highest backscattering values, while the middle 50 percent of the histogram was dominated by spatially random speckle noise. This was again the case for each scene we analyzed here; thus, we focus only on Q1 and Q4. Further, we present only results obtained from the w(j2) window shape. Although results are comparable from w(2j), the regional flooding and drainage pattern generates a stronger N-S gradient.

If we plot the Q1+Q4 scaled lacunarity index against composite Q1+Q4 percent deviation for different window sizes, we can obtain spatio-temporal patterns that captures some essential elements of the landcover dynamics (Fig. 5, Fig. 6, Fig. 7). In each landscape, there is (1) quasi-periodicity evident in the spatial arrangement of backscattering, (2) more definite cyclicity with larger sampling area, and (3) clearly anomalous positions for the late 1993 acquisitions. These anomalies likely arise from an extreme drought that struck the Pantanal during the later half of 1993. How do we interpret these trajectories? Note that deviations increase as scaled lacunarity decreases with increasing sampling window area. This pattern indicates persistent aggregates in the extreme quartiles, which correspond to forested patches for Q4 and open water or wetlands for Q1. Large range in deviation values (e.g., Fig. 5) indicates a lot of spatial rearrangement, suggesting dramatic shifts in the backscattering characteristics of large patches. Finally, are these trajectories statistically significant? This is subtle question because: (1) lacunarity index values were derived from exhaustive and non-independent sampling of the image; (2) the index is based on the first two moments of the sampling distribution weighted for mass; (3) the index does not have, as yet, analytically defined moments; (4) permutation tests would be prohibitively expensive. More significant is the question of whether these trajectories are representative of landscape dynamics from an inter-annual perspective; only additional image series can provide an answer to this question of baselines.

Figure 5: Landscape trajectory for flooded wetland. Lacunarity metrics calculated for 1st and 4th quartiles. Number labels indicate month of image acquistion. Observation sequence in months is 12/92-2/93-5-6-8-9-11/93. 
Figure 6: Landscape trajectory for lake mosaic. Lacunarity metrics calculated for 1st and 4th quartiles. Number labels indicate month of image acquistion. Observation sequence in months is 12/92-2/93-5-6-8-9-11/93. 
Figure 7: Landscape trajectory for riparian forest. Lacunarity metrics calculated for 1st and 4th quartiles. Number labels indicate month of image acquistion. Observation sequence in months is 12/92-2/93-5-6-8-9-11/93. 


The spatio-temporal patterns revealed in this study are provocative but not yet conclusive. We have illustrated the utility in extracting spatio-temporal patterns for environmental monitoring, specifically, for definition of nominal behavior and assessment of disturbance impacts. Our long term goal is the predictive modeling of landcover dynamics using landscape trajectories extracted from image time series derived from multiple sensors.


G.M.H. acknowledges support from NSF grant DEB-9696229 and a Fulbright Senior Research Fellowship at INPE during 1993-94. H.J.H.K. acknowledges ESA for support from an ERS-1 Pilot Project. Both authors acknowledge further support from ESA for an ERS-2 follow-on project. The ERS-1 SAR data were processed by and acquired through INPE.


Adámoli, J., 1986:
A dinámica das inundações no Pantanal. In: Anais do 1o Simpósio sobre Recursos Naturais e Sócio-Econômicos do Pantanal , EMBRAPA-DDT, Brasilia, pp. 51-61.
Alho, C.J.R., et al., 1988:
Environmental degradation in the Pantanal ecosystem. BioScience, 38, pp. 164-171.
ESA, 1995:
New Views of the Earth: Scientific Achievements of ERS-1. ESTEC, Noordwijk, The Netherlands.
Henebry, G.M., 1993:
Detecting change in grasslands using measures of spatial dependence with Landsat TM data Remote Sens. Environ., 46, pp. 223-234.
Henebry, G.M., and Su, H., 1993:
Using landscape trajectories to assess the effects of radiometric rectification Int. J. Remote Sens., 14, pp. 2147-2423.
Henebry, G.M., and Kux, H.J.H., 1996:
Spatio-temporal land cover dynamics in the Pantanal assessed using lacunarity analysis on an ERS-1 SAR image time series. Proc. IGARSS '96, pp. 195-197.
Henebry, G.M., and Kux, H.J.H., 1995:
Lacunarity as a texture measure for SAR imagery. Int. J. Remote Sens., 16, pp. 565-571.
Junk, W.J., et al., 1989:
The flood pulse concept in river-floodplain systems. In: Proceedings of the International Large River Symposium. Can. Spec. Publ. Fish Aquat. Sci. 106, pp. 110-127.
Klammer, G., 1982:
Die paläowuste des Pantanal von Mato Grosso und die pleistozäne Klimageschichte der brasilianischen Randtropen. Zeitschift für Geomorphologie N.F., 26, pp. 393-416.
Kux, H.J.H., and Henebry, G.M., 1994a:
Multi-scale texture in SAR imagery: landscape dynamics of the Pantanal, Brazil. Proc. IGARSS '94, pp. 1359-1364.
Kux, H.J.H., and Henebry, G.M., 1994b:
Evaluating anisotropy in SAR imagery using lacunarity functions. Int. Arch. Photogram. Remote Sens., 30(7A), pp. 141-145.
Plotnick, R.E., et. al., 1993:
Lacunarity indices as measures of landscape texture. Landscape Ecology, 8, pp. 201-211.
Plotnick, R.E., et. al., 1996:
Lacunarity analysis: a general technique for the analysis of spatial patterns. Physical Review E, 53(5), pp. 5461-5468.
Ponce, V.M., and da Cunha, C.N., 1993:
Vegetated earthmounds in tropical savannas of Central Brazil: a synthesis with special reference to the Pantanal do Mato Grosso. J. Biogeog., 20, pp. 219-225.
Prance, G.T., and Schaller, G.B., 1982:
Preliminary study of some vegetation types of the Pantanal, Mato Grosso, Brazil. Brittonia, 34(2), pp. 228-251.
Rizzini, C.T, et. al., 1988:
Ecossitemas Brasileiros/Brazilian Ecosystems. Index Editora, Rio de Janeiro.
da Silva, T.C., 1986:
Contribuição do geomorfologia para o conhecimento e valorização do Pantanal. In: Anais do 1o Simpósio sobre Recursos Naturais e Sócio-Econômicos do Pantanal , EMBRAPA-DDT, Brasilia, pp. 77-90.

Keywords: ESA European Space Agency - Agence spatiale europeenne, observation de la terre, earth observation, satellite remote sensing, teledetection, geophysique, altimetrie, radar, chimique atmospherique, geophysics, altimetry, radar, atmospheric chemistry