TROPICAL LAND COVER CHANGE DETECTION WITH POLARIMETRIC SAR DATA Emerson Luiz Servello1 Tatiana Mora Kuplich2

Yosio Edemir Shimabukuro

1

Instituto Brasileiro do Meio Ambiente e dos Recursos Naturais Renovveis IBAMA. Av. Ludovico da Riva Neto, 2643 - 78.580-000 - Alta Floresta, MT, Brazil.
2

Instituto Nacional de Pesquisas Espaciais Southern Regional Centre (INPE/CRS) Caixa Postal 5021 - 97110-970 - Santa Maria, RS, Brazil Instituto Nacional de Pesquisas Espaciais (INPE) Av. dos Astronautas 1758 - 12227-010 - So Jos dos Campos, SP, Brazil

{servello, tmk, yosio}@dsr.inpe.br

ABSTRACT There is an increasing need for fast and accurate data on tropical land cover status, and a baseline for land cover monitoring. Remotely sensed SAR data are not sensitive to cloud cover and can be useful for such purpose. Polarimetric SAR data are available in orbital systems, such as RADARSAT-2, and still have to be tested for the classification of tropical land cover and the detection of land cover change, particularly forest conversion. This work presents a study of RADARSAT-2 polarimetric images, acquired in two different dates (September 2008 and October 2009), to assess their potential in classifying forest and non-forest classes in Brazilian Amazonia. SAR images were acquired following different orbit and incidence angles, which anticipated varied conditions for images interpretation and classes discrimination. The complex SAR data were classified based on the distance of Wishart, and information from field campaigns was used for the training and test samples. Classification results were compared to evaluate possibilities for change detection in the forest cover. Classification accuracy figures were around 80%. The use of RADARSAT-2 images allowed the mapping of land cover and land cover change, considering forest and non-forest classes. Index TermsSAR data, classification, deforestation, RADARSAT-2, change detection.

1. INTRODUCTION Agriculture expansion and cattle grazing are important drivers of forest degradation and replacement in Para State, Brazilian Amaznia [1]. Beside effects on micro-climate and local hydrological regime, these land cover changes affect regional socioeconomic and environmental structure. The use of orbital remotely sensed data is a fast and efficient way for detecting and monitoring tropical land use and cover change. In Brazil, examples are monitoring systems for Brazilian Amazonia, such as PRODES (Estimate of deforestation rates) and DETER (Real-time deforestation detection) [2], based on optical orbital remotely sensed data. SAR data, not affected by clouds, are important tools for monitoring tropical forest [3] and can be used to support forest monitoring and conservation activities. Polarimetric SAR data are available in airborne and orbital systems, and classification of land cover is one of the main applications of such data [4]. Other SAR data used successfully for land cover dynamics applications were Japanese Earth Resources Satellite-1/SAR [5], Brazilian Multi-Application Purpose SAR [6], Advanced Land Observation System/PALSAR [7, 8] and R99B SAR [9]. For C band SAR data, land applications can be found in [3, 10 and 11]. RADARSAT-2 C band polarimetric SAR is a relatively new sensor in orbit and its data still have to be tested for the classification of tropical land cover and the detection of change [12]. In this study, we analysed RADARSAT-2 polarimetric SAR data for detecting land

conversion and changes from forest to non-forest cover in a tropical forest area, using 2008 and 2009 images The objective of this work was to assess the potential of such data on forest/non-forest mapping, and evaluate possibilities for change detection on the forest cover extent. 2. DATA AND METHODS Data used were RADARSAT-2 (C band 5,6 GHz) polarimetric (HH-HV-VH-VV with associated phase) images, acquired on 24 September 2008 (descending mode) and 13 October 2009 (ascending mode), in Standard mode (25 km swath, 4 looks, 25 m x 28 m spatial resolution). Data were acquired over Tapajs National Forest and surroundings, in Par State, Brazilian Amazonia (figure 1). In the study area tropical dense forest is the main cover, along with crops, pastures and regenerating forests in different successional stages. The Tapajs National Forest is a forest reserve, although small settlements and few management activities are allowed. This area presents a low deforestation rate, if compared to other areas in Brazilian Amazonia, but still, from 1997 to 2005, around 200 ha of mature forest were cleared each year [1].

SAR image processing was performed with PolSARpro 4.01 (a free polarimetric SAR data processing system) and ENVI 4.5. Classification was done using a supervised classification method, which used the distance of Wishart [4, 7], with a 7x7 window in order to enlarge the samples for the classifier and correct the effects caused by noise (speckle). The training and test samples were selected based on information from field campaigns performed in September 2008 and September 2009. Land cover classes were forest and non-forest, the same classes used in the system for mapping and monitoring deforestation in Brazilian Amazonia [2]. Non-forest class corresponded to young regenerating forests, crops, bare soil, pasture, urban area and rivers/streams. 3. RESULTS According to [9], the use of images acquired with different polarizations can produce varied results when classifying forest and non-forest classes. Our work, however, tried to investigate the effects of using complex SAR data, instead of amplitude images, in the classification results. Although the SAR data were acquired following different conditions orbit, incidence angle and rainfall before the 2008 SAR image the results highlighted the possibilities and potential of using temporal polarimetric SAR data for deforestation mapping. The results can help guiding the use of RADARSAT-2 data for monitoring purposes. Comparing classification results for the 2008 and 2009 SAR images, 500 hectares of forest were converted to other uses. Comparing with PRODES data for the same period [2], we found 100 ha of forest converted, although 1600 ha of the area was covered by clouds and possibly not all forest converted areas were accounted for. Figure 2 displays the 2008 and 2009 classified images and the differences in the classes cover.

Figure 1: Study area in South America, Par State (Mosaic of Terra/MODIS and Landsat/Thematic Mapper images) and detail of a RADARSAT-2 scene used in this work.

Days before the acquisition of the scene in September 2008 there was precipitation around 8 mm. There was no precipitation before the 2009 scene acquisition.

Figure 2: Classified 2008 (a) and 2009 (b) RADARSAT-2 images. Colors for the classes are: green is forest and orange non-forest.

When considering the effects of shadowing caused by relief in the 2008 SAR image classification, figure 3 shows an example of the commission errors for the non-forest class. All the shadowed areas were classified as non-forest and had to be manually edited from the final results. The Digital Elevation Model (DEM) helped to locate areas with potential for misclassification due to relief shadowing (figure 3). This problem did not happened for the RADARSAT-2 October 2009, acquired in ascending mode. The occurrence of precipitation and soil humidity before the acquisition of RADARSAT-2 2008 image increased confusion and reduced discrimination between classes. Figure 4 presents an example of the discrimination between classes in the 2 SAR image dates and also in the TM August 2008 image. The darker areas in the 2009 SAR image (figure 4 top right) are non-forest areas and are typical of bare soil/low vegetated areas in dry condition C band images. For the 2008 SAR image, the rainfall reduced the backscattering range and consequently, the discrimination between classes [5]. Final classification accuracy figures were around 71% for the 2008 SAR image and around 89% for the 2009 SAR image. Most of commission and omission errors happened for the 2008 SAR image classification, as different conditions of image acquisition (orbit, incidence angle and rainfall before the image acquisition) prevented a precise discrimination between forest and non-forest classes. Nevertheless, accuracy figures for classification of both RADARSAT-2 SAR images were high and indicated the possibilities of using such data for tropical forest cover change detection.

works with geo-referenced data. The assignment of a geographic reference system to images is what makes the comparison between images and validation tests possible, essential for change detection approaches. Methods used here were based on supervised classification, which results are dependent upon the quality of the training samples. Further tests and analyses are required to fully address the question of using polarimetric SAR data for tropical land cover change detection and monitoring. We suggest the use of further classification approaches, along with polarimetric data processing techniques, to refine and quantify results.

4. FINAL REMARKS RADARSAT-2 images can be used for the detection of forest/non-forest classes and their change, although some limitations were noted. There was rainfall before the acquisition of one of the SAR images used, which prevented a complete interpretation of the land cover classes, using the Wishart supervised classifier. Areas of considerable relief were obscured by shadows in the RADARSAT image from a descending mode. Shadowing also caused loss of information about land covers and confusion between classes. Still, when comparing with the absence of information caused by cloud coverage in optical data, SAR data is more suited for monitoring purposes, as relief areas are known and shadowing problems can be corrected. To redress this problem, classifications were edited based on ground truth and ancillary data (DEM).. The use of RADARSAT-2 images, however, was feasible in the study of land use and land cover change, considering forest and non-forest classes. The PolSARpro software is an important tool for the classification of polarimetric SAR data, but it requires the use of an auxiliary software that

Figure 3: Relief effects on an extract of SAR September 2008 image: (a) the DEM for the study area from Shuttle Radar Topographic Mission data, (b) relief areas and associated shadows classified as non-forest class RADARSAT-2 September 2008 image and (c) Landsat Thematic Mapper August 2008 image, 5R4G3B. Commission errors were highlighted in the box.

[5] Kuplich, T. M.; Curran, P. J. Temporal analysis of JERS-1 SAR images over regenerating forests in Brazilian Amazonia. In: International Geoscience and Remote Sensing Symposium - IGARSS, 1999, Hamburg. Proceedings. Piscataway: IEEE, v. IV. p. 1895-1897, 1999. [6] Mura, J. C.; Paradella, W. R.; Dutra, L.V.; Santos, J. R.; Rudorff, B. F. T.; Miranda, F. P. ; Silva, M. M. Q.; Silva, W. F. MAPSAR Image Simulation Based on L-band Polarimetric Data from the SAR-R99B Airborne Sensor (SIVAM System). Sensors, n.9, p. 102-117, 2009. [7] Jong-Sen Lee, J. S.; Grunes, M. R.; Ainsworth, T. L.; Du, L. J.; Schuler, D. L.; Cloude, S. R. Unsupervised classification using polarimetric decomposition and the complex Wishart classifier. IEEE Transactions on Geoscience and Remote Sensing, vol. 37, n. 5, p. 22492258, 1999. [8] Almeida filho, R.; Rosenqvist, A.; Shimabukuro, Y.; Santos, J. R. Evaluation and perspectives of using multitemporal L-band SAR data to monitor deforestation in Brazilian Amaznia. IEEE Transactions on Geoscience and Remote Sensing Letters, v. 2, n. 4, p. 409-412, 2005. [9] Guerra, J. B.; da Costa Freitas, C.; Mura, J. C. Evaluating the potential of L Band PolSAR data to discriminate deforestation increment areas in Amazon rain forest. IEEE Transactions on Geoscience and Remote Sensing Letters, v. 4, p. 45-48, 2008. [10] Park, N. W.; Chi, K. H. Integration of multitemporal/polarization C-band SAR data sets for landcover classification. International Journal of Remote Sensing, Vol 29 , n. 16, p 4667-4688, 2008 [11] Rignot, E. J. M. e van Zyl, J. Change detection techiques for ERS-1 SAR data. IEEE Transactions on Geoscience and Remote Sensing, vol. 31, n. 4, p. 896-906, 1993. [12] Kuplich, T. M.; Shimabukuro, Y. E. , Servello, E., Sano, E. Polarimetric signatures and classification of different tropical land covers. In: International Geoscience and Remote Sensing Symposium - IGARSS, 2009, Cape Town, South Africa. Proceedings. Piscataway: IEEE, 2009.

Figure 4: Extract of the different images and examples of nonforest class: top left - September 2008 RADARSAT-2 image, top right - September 2009 RADARSAT-2 image and bottom August 2008 Landsat TM image.

5. REFERENCES

[1] Cohenca, D. Evoluo anual de desmatamentos na Floresta Nacional do Tapajs de 1997 a 2005. XIII Simpsio Brasileiro de Sensoriamento Remoto, Florianpolis, Brasil, Proceedings, 21-26 abril 2007, INPE, p. 6653-6660, 2007. [2] PRODES Project Monitoring the Brazilian Amazonia with satellite data, http://www.obt.inpe.br/prodes/, viewed May 2010. [3] Quegan, S.; Le Toan, T.; Yu, J. J.; Ribbes, F. e Floury, N. Multitemporal ERS SAR analysis applied to forest mapping. IEEE Transactions on Geoscience and Remote Sensing, vol. 38, n. 2, p. 741-753, 2000. [4] Zou, T.; Yang, W.; Dai, D.; Sun, H. Polarimetric SAR image classification using multifeature combination and extremely randomized clustering forests. EURASIP Journal on Advances in Signal Processing, article ID 465612, p. 9, 2010.