Deforestation in the tropical rainforest belt is one of the main global change processes which exerts large influence on climate and on global biochemical cycles. Monitoring is needed to estimate yearly deforestation rates and to register "hot spots" of deforestation but also to follow dynamics in deforested areas concerning vegetation conversion, regeneration and possible degradation processes. Thus, several rainforest study sites have been defined within the SIR-C/X-SAR missions, especially in the Brazilian Amazon, for investigation of multifrequency and multipolarization radar data on these subjects.
In co-operation between DLR (German Aerospace Research Establishment) and INPE (National Brazilian Institute of Space Research) two study sites in South-western Amazon (in the states of Acre and Rondonia) are under investigation using multifrequency SIR-C/X-SAR data, in order to characterise land cover and land cover changes. In Acre, especially around the capital Rio Branco, deforestation is mainly pushed by extensive cattle farming but also by colonisation processes. Clearing by burning is a common procedure for the farmers who use burning also for cleaning the pastures. In Rondonia around Ji Parana, fishbone patterns of deforestation have been formed by the colonisation scheme of PAD´s (Projeto Assentamento Dirigido = Project of Directed Settlement)(Walker,1996). Location of the study sites of Rio Branco and Sena Madureira (Acre) as well as the sites of Ji Parana and Jaru (Rondonia) are shown in Fig. 1.
Objectives in both project parts are to evaluate
the potential of multifrequency and multipolarimetric SIR-C/X-SAR
data for
In the Acre testsite the five April data takes of the Sena Madureira site cover the area from Rio Branco along the road towards Sena Madureira in the NW, two of them are multipolarimetric. Three SIR-C/X-SAR data takes have been registered in October. Main evaluations have been performed with one of the multipolarimetric datasets (April 16) and with the dual-pol. dataset of April 15 and Oct. 4. Landsat TM data of August 1992 and July 1994 were used as reference. Several ERS SAR datasets of the years 1992, 1993, 1995, and 1996 were available for change detection purposes within an ERS-1 pilot project. During a field survey in June/July 1994, additional reference information was registered assisted by FUNTAC (Technological Foundation of Acre) and the University of Acre. An overview of the entire area was obtained by an overflight during which 35 mm airphotos were taken as well. Several vegetation profiles, arboreal and bush individuals along sections of 60 m by 2 m size were drawn , mainly in regrowth, to estimate biomass.
For the northern test site, in the department of Jaru, three SIR-C/X-SAR datasets of the Pantanal site have been used: one full-polarimetric data take of April 9 and one dual-polarised data take of April 10 and October 1, respectively. For the southern test area around the city of Ji-Parana the full-polarimetric data take of April 10 and one dual-polarised of October 1 were used. All data are from descending passes.
In Ji-Parana, the southern test site, only the dual-polarimetric dataset has been evaluated for classification purpose, because of its better ground resolution. Work on the full-polarimetric data to evaluate the possibilities to estimate biomass content of regenerating areas is continuing. Additionally, a set of five ERS-1 images has been evaluated. One TM-scene recorded on July 15, 1994, was used as reference data. Field surveys including an overflight have been performed in co-operation between INPE and DLR in Oct.'95 and Nov.'96. The second field survey had the objective to gather vegetation profiles for the allometric estimation of biomass in different succession stages and forest stands.
For investigation of SIR-C/X-SAR data mainly standard polarisations were extracted, that is HH-, VV- and HV-polarisation besides total power (TP) in L- and C-band. In addition, circular polarisation states (LL-pol., left-hand circular, and RR-pol., right-hand circular) were used. Extracted polarisations were transferred to backscatter coefficient values 0 in dB to compare different data takes and regions. For these comparisons calibrated products were very important. In order to test a maximum likelihood approach for supervised classification, the data was MAP-filtered applying window sizes 3 by 3 and 5 by 5. For texture classification unfiltered data was used.
Supervised classification of different landcover classes was one of the main objectives of the study. For preparation of landcover classification, e.g. to select bands and training sets, backscatter signatures of various polarisation of the selected reference sites have been studied. Two different classification approaches were used: conventional Maximum Likelihood classification, performed on pre-filtered SIR-C/X-SAR band combinations, and the EBIS texture classifier ("Evidence Based Interpretation of Satellite data") by Lohmann (1994).
The window based classifier EBIS takes into account the speckle and textural information of radar data within a pixel environment. Window sizes between 5 by 5 and 15 by 15 can be chosen. In one option, local histograms in this window environment are evaluated for their class contribution, based on multinomial distributions and criteria of evidence, according to Dempster-Shafer decision rules (Lohmann 1991). Greyvalue intervals between 6 and 16 can be chosen for the description of local histograms. In another option, textures can be classified based on several co-occurence feature vectors, also modelled by multinomial density functions (Lohmann, 1994). The co-occurence matrix of direct neighbours can be calculated by using the directions horizontal, vertical, left-diagonal, and right-diagonal.
Deforestation Mapping
Separation of forest and non-forest areas was best using L-band data, because of the longer penetration depth of this band, a fact known from many studies. In C- and X-band separation depends more on the vegetation season, with a better separation at the end of the dry season (October) than in the ending wet season in April (Keil et al, 1996). Especially for beginning regrowth, L-band data is superior. An amazingly good result was reached for deforestation mapping by using an X-band dataset of October 4 for textural classification (Keil et al, 1995). The high geometric resolution of this data take is connected with a high information content on texture of the canopies in the EBIS classification. A window of 13 by 13 and horizontal and vertical co-occurence data was used. For the main study site North-west of Rio Branco, accuracies of 99% for forest and 93% for non-forest were reached by this approach. Old regrowth, older than about 12 years, had to be incorporated in the forest class. The forest/non-forest classification result was used in a change detection map, comparing with evaluations of ERS-1 data of 1992 and 1993 (Keil et al, 1996).
The dynamics of deforestation in the clearing period between April and October is demonstrated by two corresponding L-band datasets in Fig. 2. An area along the road between Sena Madureira and Rio Branco is shown in the combination of L-HH, L-HV and L-TP with the rainforest in bright green. Newly cleared areas in the October scene appear in pink, as L-HH reacts on remaining trunks and wood by a larger portion of double bounce scattering. In April, L-HH marks the alluvial forest zones along the rivers by double bounce scattering, also appearing in pink. Thus, detection of subcanopy-flooded forest areas is demonstrated, where water or at least water saturated soils are responsible for the high trunk-ground interaction.
Mapping of Landuse Intensity and Regeneration States
Besides the question concerning rates and mechanisms of forest conversion to agricultural landuse, another main question for Amazon is: at what rate are abandoned lands converted to secondary forests, and what is the fate of abandoned lands, regarding e.g. degradation and erosion processes (LBA Science Planning Group, 1996)?
The following land cover classes on deforested areas are of interest in the context of conversion (dos Santos et al, 1996):
The regenerating sites were often quite inhomogeneous
and difficult to delineate. Also, the high dynamics of landcover
changes by burning partly complicated the comparison of SIR-C/X-SAR
data with ground surveys and TM data of July 94. An example for
Intermediate Regrowth is shown in a diagram of a
vegetation transect in Fig. 3, recorded N of Rio Branco.
Fig. 2:SIR-C Images of an Area NW of Rio Branco, area 37.5 km by 40 km.
Composite of L-band Data (L-HV,L-HH, L-TP in R,G,B)
above: Scene from April 16, end of wet season, below: Scene from Oct. 4, end of dry season..
Polarimetric signatures of these classes show larger overlaps on all polarimetric and frequency states, but L-HV polarisation can be addressed as the most important band for the conversion classes (Fig. 4). In addition, pre-filtering is essential for pixel-wise classification, for EBIS texture classification pre-filtering is not necessary
|
| ||||
PERCENTAGE OF VERIFIED AREA
Table 1: Classification Accuracy for SIR-C Data of April
A result of landcover classification by EBIS is shown
in Fig. 5. Here, a set of L-HV, L-HH, L-TP of April has
been used as classification input. Five classes were evaluated,
Overgrown pastures and Regrowth had to be combined.
Under the conditions of quite large extended pasture lands and
regeneration units north-west of Rio Branco, EBIS proved to be
a successful classification tool. According to a classification
estimate, class accuracies lay between 60.7% and 97.6% except
for the Sub-canopy Flooding class, which was not detected
in full extension (Tab. 1). Regrowth classification was
not very clear against pasture and forest, possibly because of
high vegetation period at the end of the rainy season.
Fig. 5: EBIS classification Result in the Area NW of Rio Branco (37.5 x 40 km)
Signature analysis was also applied for the Rondonia test site to evaluate the best combination of frequencies and polarisations to optimise the classification. A first guess on separability of the was made using the results of signature analysis, shown here for unfiltered data.
Fig. 6 shows the rather low separability of the classes High Biomass Pasture (partly overgrown) and Initial Regrowth in all C-Band polarisations, because, relative to this wavelength, the classes High Biomass Pasture and Initial Regrowth have a similar distribution of scattering elements. Separability is somewhat better in L-HH, VV and L-band total power. Separability of Low Biomass Pasture against High Biomass Pasture is best in L-HV.
Fig. 7 shows
the difficulties in separating landuse classes containing more
biomass by using higher frequency data. Due to penetration depth
of C-band SAR in vegetation a similar distribution of vertically
and horizontally orientated scatterers in the upper crown layer
of rainforest and early succession stages reduces separability.
The fast increasing leaf biomass of young regrowth stands provides
for this phenomenon. Separability is best in L-HV, as shown above.
The separability of old regrowth and rainforest is low even in
L-band, because of the high biomass levels and stand heights in
these classes. Depending on the succeeding plant society regrowth
stands of 6 years or older, having a stand height of 20 metres
and more (Cecropia spec.), were observed. Discrimination of rainforest
stands and various types of pastures was quite straightforward
in L-Band.
A combination of C-HV, LHH and L-HV proved best for
classification of the October data, whereas for the April scene
a combination of C-HH, L-HH and L-HV was chosen, in order to enhance
the detection of sub-canopy floodings. The data were MAP-filtered
twice with 3*3 and 5*5 moving windows and then classified with
the Maximum-Likelihood classifier. Also, several different EBIS
approaches were applied to the SIR-C/X-SAR data but were not successful.
One reason for this failure lies in the special pattern of the
fazendas in this region. Problems occur when small-scaled patterns,
as the heterogeneous fishbone patterned fazendas in Rondonia,
lead to several landcover borders within the moving window's pixel
environment. Discrimination of the classes by maximum-likelihood
was faster, more accurate and more diversified in the Rondonia
study site.
Classification results for April and October data sets are shown in Fig. 8 to 9. In the larger area of the October classification in Fig. 9, the fishbone pattern structure of colonisation is shown.
|
| ||||
PERCENTAGE OF VERIFIED AREA
Table 2: Classification Accuracy for SIR-C Data of April
In an approach for accuracy assessment, class limits of Regrowth proved more diffuse in April, which might be a result of higher humidity in the rainy season. The cause of the high overlap of Pasture and Regrowth (Tab. 2) may be seen in the increase of biomass on pastures during rainy season. Class limits between Initial Regrowth and Pastures are unstable. The drop in accuracy of Sub-Canopy Floodings is also due to insufficient ground truth data. Sample areas for the evaluation had to be interpreted from the SAR dataset and could not provide enough data to estimate the error in the areas far from river banks.
| Forest | Degraded
Pasture | New
Clearcut | Pasture | Regrowth | |
| Forest | 96.5 | 0.0 | 1.0 | 2.5 | 0.0 |
| Deg. Past. | 0.0 | 90.5 | 0.0 | 7.5 | 2.0 |
| New Clear. | 1.0 | 0.0 | 86.3 | 0.4 | 12.3 |
| Pasture | 0.5 | 2.3 | 1.7 | 93.7 | 1.8 |
| Regrowth | 4.3 | 1.8 | 2.6 | 30.3 | 61.0 |
PERCENTAGE OF VERIFIED AREA
Table 3: Classification Accuracy for SIR-C Data of October
Classification accuracy increases slightly in October for classes Degraded Pasture and Regrowth (Table 3). In October a new class appears. Newly cleared areas are discriminable visually and by automatical classifiers. Double bounce on remaining trunks provides a high definition of this class in L-HH polarisation.
Evaluation of vegetation transects is going on in
order to find regressions between biomass and L-HV data, possibly
also other band combinations like ratios of L-HCV and C-HV. For
differentiation of pasture and regeneration states, this ratio
was found to be valuable.
Fig 8: SIR-C Images of April and October of the Jaru Test Site and Classification Results
Fig. 9: Result of Classification of October Image, Image Area appr. 80 x 60 km
In L-band SAR discrimination of forest and non-forest classes was found to work especially well. L-HH is sensitive to sub-canopy floodings, which could mostly be discriminated (Fig. 8). Thus, for operational monitoring, the Japanese JERS sensor with its L-HH band should deliver good results in detecting and mapping these areas. The correlation of biomass in HV polarisations is high and increases when data of two frequencies are used. The saturation level seems to be about 100 tons/hectare, which is consistent with the latest publications on this issue. Early to middle stages of regeneration areas, not older then ten years, are detectable. Correlation to biomass isn't that high in L-HH, due to double bounce scattering on tree trunks. Different pasture stages are detectable. The use of biomass ratios improves separability of pasture classes.
Using SIR-C/X-SAR data discrimination of the following classes is possible, uncertainties for transition stages still remain:
primary rainforest
sub-canopy floodings (flooded rainforest along rivers)
initial regrowth(1-4 years of age)
intermediate regrowth (5-15 years)
pastures (partly covered with palmtrees)
degraded pastures (< 40 % ground cover)
bare soils (dirt roads, badlands)
fresh clearcuts (newly burned areas, without vegetation cover)
The transfer of the experience gained from the analysis of SIR-C L-band data to operational data like JERS data has yet to be done. However, SIR-C/X-SAR data with their high information content have to be regarded as transect data. The deduced information may be used for extrapolation to large neighbouring areas covered by sensors like JERS and ERS-1/ERS-2. In addition, SIR-C data, being recorded in the rainy and dry season of 1994, provide a valuable database to understand and interpret signatures in SAR images of these regions. The strips of data over the Amazon can especially be used for mapping and estimation of different regenerating stages. Through the estimation of biomass levels with SIR-C data more current observations over the Amazon can be cross-checked and verified. Thus, SIR-C data deliver a valuable input for upcoming projects to estimate interference of landuse pattern, biochemical cycles, and climate, such as LBA project (The Large Scale Biosphere-Atmosphere Experiment in Amazonia).
Dos Santos, J.R., H.J.H. Kux, M. Keil, M.S.P. Lacruz, D.R. Scales, 1996: Interactive Analysis of Polarimetric SIR-C Data and Landsat Data for the Spectral and Textural Characterization of the Land Cover in SW Amazon, ISPRS-96 Conference, Int. Archives of Photogrammetry and Remote Sensing, Vol. XXXI, Part B7, p. 209 - 213, Vienna 1996.
Keil, M., D.Scales, R.Winter, 1995: Investigation of Forest Areas in Germany and Brazil using SAR Data of the SIR-C/X-SAR and other SAR Missions, Proceedings IGARSS ' 95 Conference, 10-14 July 1995, Vol. II, p. 997-999, Firenze, Italy.
Keil, M., D.Scales, R.Winter, H.Kux, J.R.Dos Santos, 1996: Tropical Rainforest Investigation in Brazil using Multitemporal ERS-1 SAR Data, Proceedings of the Second ERS Application Workshop, London, 6-8 Dec. 1995. (ESA SP-383, February 1996).
LBA Science Planning Group, 1996: The Large Scale Biosphere Atmosphere Experiment in Amazonia (LBA), Integrated Science Plan, Cachoeira Paulista, Brazil, Dec. 1996.
Lohmann, G., 1991: An Evidential Reasoning Approach to the Classification of Satellite Images, DLR- Research Report 91-29, Deutsche Forschungsanstalt für Luft- und Raumfahrt, Köln, 1991.
Lohmann, G, 1994: Co-occurrence based Analysis and Synthesis of Textures, International Conference on Pattern Recognition, Oct. 94, Jerusalem, Israel.
Walker, R. (editor), 1996: Land Use Dynamics in the
Brazilian Amazon, in Ecological Economics, Special Issue No. 18,
1996, Elsevier.