Monitoring land cover in Acre State, western Brazilian Amazonia, using multitemporal remote sensing data

Transcription

1 Monitoring land cover in Acre State, western Brazilian Amazonia, using multitemporal remote sensing data Yosio E. Shimabukuro Valdete Duarte Egidio Arai Ramon M. Freitas Paulo R. Martini André Lima Instituto Nacional de Pesquisas Espaciais (INPE) Av. dos Astronautas, 1758 CEP , São José dos Campos, SP, Brazil { yosio, valdete, egidio, ramon, martini, andre }@dsr.inpe.br Abstract This paper presents the use of multitemporal remote sensing data for monitoring land cover changes in Acre State, western Brazilian Amazonia. For this study, the 2000 Landsat ETM+, the 1990 Landsat TM, and 1980 Landsat MSS were used. The 2005 and 2007 Terra MODIS images were also used to map deforestation occurred during the recent years and to map burned areas occurred in the 2005 dry year. The ETM+, TM, MSS and MODIS images were converted to vegetation, soil, and shade fraction images using Linear Spectral Mixing Model. Then land cover maps were produced by digital classification of these fraction images. The results showed that deforestation increased 7,114 km 2 from 1980 to 1990, 4,900 km 2 from 1990 to 2000, and 3,258 km 2 from 2000 to About 2,815 km 2 of vegetation regrowth areas were observed in the 2000 ETM+ images. The analysis of 2005 MODIS images showed that 3,700 km 2 of deforested areas and 2,800 km 2 of forested areas were burned in Acre State in This information is critical for regional and global environmental studies and for efforts to control such burning and deforestation activities in the future. Index Terms Landsat data, land cover map, multisensor data, multitemporal analysis, digital classification, Amazonia region. 1. INTRODUCTION The updating of land use land cover (LULC) change dynamics is a main purpose of the remote sensing science community. Remote sensing provides an important record

2 for studying multitemporal changes in the terrain surface. In addition to being the only data source available for many inaccessible and large regions, remote sensing provides the opportunity to acquire data with the spatial coverage and temporal frequency adequately for studying and monitoring vegetation at a relatively low cost. Several studies using remote sensing for change detection demonstrate that remote sensing technologies are becoming increasingly important for studying the forest monitoring at regional and global scales (Hansen et al. 2000; Achard et al. 2001; Friedl et al. 2002; Eva et al. 2004). The Landsat Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), and Terra/Aqua MODIS data collect information in the major portions of visible, nearinfrared, and mid-infrared of the electromagnetic spectrum, allowing for detecting and characterizing the land cover at regional and global scales. In addition, the global GeoCover orthorectified Landsat image mosaics provided by United States Geological Service (USGS) are a free remote sensing data traditionally used for LULC change studies. The GeoCover image mosaics were generated for the years 1990 (TM) and 2000 (ETM+) allowing to evaluate the land cover during this 10 years time period. The first satellite images were provided by MSS sensors that were in operation onboard the Landsat 1, 2 and 3. The National Institute for Space Research (INPE) data distribution policy made the entire archive of Brazilian Landsat imagery free for the public users. Then the historic time-series image dataset can be produced for LULC change dynamic studies. One of the best examples of operational monitoring for quantifying land cover change is the work of the National Institute for Space Research (INPE) in mapping deforestation within the Legal Amazon. Since 70 INPE has been performing the evaluation of deforested areas in the Legal Amazon. During this time it was utilized data from MSS (Multispectral Scanner System) onboard the Landsat satellite. It was performed two estimates using MSS images acquired during the 1973 to 1975 and 1975 to 1978 time periods. In 1988 INPE started to estimate the deforested areas in an annual base. At this time the MSS data were replaced by TM data, which improved the map accuracy due to the better TM spatial (30 m) and spectral (6 bands) resolutions when compared to MSS (80 m; 4 bands). Until the end of 90 decade, the methodology of mapping deforested areas was based on manual interpretation followed by map digitizing to calculate the deforested areas. This procedure did not allow to produce a digital product for PRODES. Then the information was available only in a Table format. In 1997 was

3 created a first digital version based on methodology presented in Shimabukuro et al. (1998). However PRODES maps only the converted areas occurred in the forested areas. The deforested areas mapped since the beginning of the project have not been analyzed in the following years. Thus the dynamic of land use in these areas are not observed. Significant part of these areas has been abandoned immediately after the clearing or after few years of use for cattle raising (Steininger 1996). Then these areas start the processing of natural succession of vegetation cover. The vegetation regrowth areas have higher photosynthetic, evapotranspiration, and respiration rates and higher capacity of carbon stock than agricultural areas as mentioned in Goreau and Mello Those factors show the important role of the secondary forest for the carbon balance and then the necessity to monitor these areas. Using medium, moderate or low spatial resolution sensor data we have the so-called mixture problem, i.e., the pixel value is a mixture of reflectance from different targets within each pixel. Several techniques, such as modeling and empirical estimations, have been applied to depict subpixel heterogeneity in land cover from remotely sensed data (DeFries et al. 2000). Fraction images derived from different remote sensing data have provided consistent results for monitoring deforestation (Shimabukuro et al. 1998), land cover change (Carreiras et al. 2002), vegetation classification (DeFries et al. 2000), and burned areas mapping (Shimabukuro et al. 2009). Fraction images, derived from a linear spectral mixing model, constitute synthetic bands with information on end-member proportions. The generation of these images is an alternative approach to reduce the dimensionality of image data and enhancing specific information for digital interpretation (Aguiar et al. 1999). In this context, the objective of this paper is to use multitemporal remote sensing data for monitoring land cover changes in Acre State located in the western Brazilian Amazonia. 2. STUDY AREA Acre State, located in the western region of Brazilian Amazonia (Figure 1), served as the site for this study. According to IBGE vegetation maps, the study area is primarily covered by moist tropical forest ( Floresta Ombrófila Aberta ) that has been partially deforested during the last decades. The climate in Acre is classified as Am in the Koppen system. Average monthly temperatures range from 24 to 27 degrees Celsius. Yearly rain is about 2,100 mm, with a dry season on June to August. In 2005, the

4 western Amazonia region suffered a drought that provoked a large number of fires, including fires that penetrated into standing forests in the study area (Brown et al. 2006). FIGURE 1 3. MATERIAL AND METHODS 3.1 Remote Sensing Data The Landsat Program is a series of Earth-observing satellite missions jointly managed by the U.S. Geological Survey and NASA. Since 1972, Landsat satellites have collected information about Earth from space using several sensors. MultiSpectral Scanner (MSS) sensor provided a historical record of the Earth's land surface from early 1970's to early 1990's. The MSS scenes have a pixel size of 80 meters with 4 spectral bands ( um, um, um, and um). The GeoCover mosaic is a global set of regional images mosaicked from the Landsat GeoCover data set. This project was designed at NASA Stennis Space Center ( The 1990 GeoCover mosaic images were produced using 3 spectral TM bands (Green, NIR and MIR) resampled for 28.5 meters of spatial resolution with 50 meters of absolute positional accuracy. The 2000 GeoCover mosaic images were produced using 3 spectral ETM+ bands (Green, NIR and MIR) resampled for meters of spatial resolution using Cubic Convolution Interpolation Method. The Terra/MODIS MOD09 products with seven surface reflectance bands (1-Red, 2- NIR, 3-Blue, 4-Green, 5-, 6-, 7-MIR), acquired in 2005 and 2007, were also used to map land use land cover changes and burned areas. For this, the red (band 1, centered at 640 nm), NIR (band 2, centered at 858 nm), and MIR (band 6, centered at 1640 nm) surface reflectance bands with 250m of spatial resolution were used. The band 6 has 500 m spatial resolution resampled to 250 m. Figure 2 shows the image mosaics used in this study: MSS, TM, ETM+, and MODIS. FIGURE Methodological Approach Figure 3 shows the digital classification procedure using multisensor data. The images dataset used in this study are: a) the Landsat Multispectral Scanner (MSS) image centered on 1980, with pixel size of 80 m; b) the GeoCover orthorectified Landsat

5 Thematic Mapper (TM) mosaic image centered on 1990, with pixel size of 28.5 m resampled to 100 m; c) the GeoCover orthorectified Landsat Enhanced Thematic Mapper Plus (ETM+) mosaic image centered on 2000, with pixel size of m resampled to 100 m; and MODIS acquired in 2005 and FIGURE 3 The MSS, TM, ETM+ datasets were used for analyzing the land use and land cover changes during the 20 years ( ) time period. In addition, 2005 and 2007 MODIS images, with 250 m of spatial resolution, were used to map deforestation occurred during the recent years and also to map burned areas occurred in the 2005 dry year in the study site. The Landsat-MSS, -TM and ETM+, and Terra-MODIS images were converted to vegetation, soil, and shade fraction images in order to enhance the characteristics of land cover, expressed as different mixtures of these few number of terrain components. The Linear Spectral Mixing Model (LSMM) has the objective to estimate the amount of soil, vegetation and shade for each pixel from the spectral response in the sensor bands, generating the corresponding soil, vegetation and shade fraction images (Shimabukuro and Smith, 1991). The LSMM can be written as: r i =a*vege i +b*soil i +c*shade i +e i, where r i is the response for the pixel in band i of TM image; a, b, and c are the proportion of vegetation, soil, and shade in each pixel; vege i, soil i and shade i correspond to the spectral responses of each components; e i is the error term for each band i. A sensor (e.g. TM) bands are used to form a linear equation system that can be solved by any developed algorithm. The unmixing methods available in several software packages estimate the proportion of each component inside the pixel by minimizing the sum of squares of the errors (e.g., Constrained Least Squares, Weighted Leat Squares). The resulting fraction images contain specific information: soil fraction image highlights mainly non-vegetated areas (clear cuts, bare soil, etc.); the vegetation fraction image shows the vegetation cover condition similar to the well known normalized difference vegetation index (NDVI); and shade fraction image enhances water bodies, vegetation cover structure, and burned areas.

6 The generation of these images is an alternative approach to reduce the dimensionality of image data and enhancing specific information for digital interpretation. Then land cover maps were obtained by digital classification of these fraction images, following a procedure based on image segmentation, unsupervised classification, and postclassification edition (Shimabukuro et al. 1998). In this study, shade and soil fraction images (1980 MSS, 1990 TM, 2000 ETM+, and 2007 MODIS) were used to map deforested areas, while vegetation fraction images were used to map regrowth areas. For mapping burned areas in the 2005 MODIS data, shade fraction image was the primary source of information. The increase of shade proportion present in a forested pixel provides the information to detect the effects of fire on the forest canopy. Similarly, the increase of shade proportion in a grassland pixel indicates the effects of fire over deforested areas. Image segmentation is a technique to group the data, in which only contiguous regions and similar spectral characteristics can be joined. The image segmentation approach used in this study is based on a region growing technique. Two threshold parameters have to be set by the analyst to define segments (regions) that will be used in the subsequent classification procedure: (a) similarity threshold (the Euclidean distance between the mean digital number of two regions, under which they will be grouped together); and (b) an area threshold (minimum area to be considered as a region, set by the number of pixels) (Bins et al. 1993). Segmented images were classified using ISOSEG (Bins et al. 1993), a region classifier algorithm based on clustering techniques. This unsupervised algorithm uses the covariance matrix and the mean of the regions to estimate the centers of the classes. The analyst defines an acceptance threshold, the maximum allowed Mahalanobis distance that a mean digital number may be from the center of a class, to be considered as belonging to that class. After the classification process, some classes may be regrouped to express more faithfully terrain features. After the unsupervised classification, it is necessary to check the resulting maps. This task is done by the interpreter using image edition. For this, the original color composites are used for comparison. This task minimizes the omission and commission errors normally produced by any digital classifier (Almeida Filho and Shimabukuro 2002). These resulting products allowed to estimate the interchanges in the land use land cover classes (such as forest, deforestation, regrowth, burned forest, and burned grassland

7 areas) over the considered time period as well as the increment of deforested areas from one period to another. 4. RESULTS AND DISCUSSION The Linear Spectral Mixing Model was applied to multisensor remote sensing data in order to highlights the land cover types in the study area. As an example, Figure 4 shows the vegetation, soil, and shade fractions images derived from 2007 MODIS mosaic. FIGURE 4 Figure 5 shows the land cover classification for the years 1980, 1990, 2000 and It also shows the land cover changes during this time period. The legend of the map produced includes: deforestation 1980, deforestation 1990, deforestation 2000, deforestation 2007, forest 2007, regrowth 2000 (deforestation 1980 and 1990), regrowth 2000 (deforestation 1990 and forest 1990), regrowth 2000 (forest 1980 and 1990). FIGURE 5 The multitemporal analysis of Landsat datasets corresponding to 1980 (MSS), 1990 (TM) and 2000 (ETM+) showed that the deforestation areas increased 7,114 km 2 from 1980 to 1990, 4,900 km 2 from 1990 to 2000, and 3,258 km 2 from 2000 to 2007 time periods. It also showed that about 2,815 km 2 of regrowth areas were observed in the 2000 ETM+ images (Figure 6) FIGURE 6 The Terra/MODIS product MOD09, acquired on September 05 and 15, 2005 and on October 12 and 21, 2005, were used to evaluate the burned areas occurred in this dry year in the Acre State. Figure 7 shows the burned areas mapped in the 2005 MODIS time series. FIGURE 7

8 The analysis of MODIS images showed that 6,500 km 2 of land surface were burned in Acre State in Of this, 3,700 km 2 corresponded to the previously deforested areas and 2,800 km 2 corresponded to the forested areas (Figure 8) (Shimabukuro et al. 2009). FIGURE 8 The derived information about deforestation, regrowth and burned areas are critical for regional and global environmental studies and for efforts to control such burning and deforestation in the future. 5. CONCLUSIONS The method described in this paper can be used to digitally classify land cover changes using multisensor images of Acre State. The results showed that deforestation increased 7,114 km 2 from 1980 to 1990, 4,900 km 2 from 1990 to 2000, and 3,258 km 2 from 2000 to About 2,815 km 2 of vegetation regrowth areas were observed in the 2000 ETM+ images. The analysis of 2005 MODIS images showed that 3,700 km 2 of deforested areas and 2,800 km 2 of forested areas were burned in Acre State in This information is critical for regional and global environmental studies and for efforts to control such burning and deforestation activities in the future. The results demonstrate that multisensor data are important sources of information for mapping and monitoring land cover changes and can be used at the regional level in Brazilian Amazonia. The next step of this research is to apply the proposed method for the entire Amazonia as part of DETER (Shimabukuro et al. 2006) (Detection of Deforested Areas in Real Time, and PANAMAZONIA ( panamazon.htm) operational projects developed at the National Institute for Space Research (INPE). 6. REFERENCES Achard, F., Eva, H.D. and Mayaux, P., Tropical forest mapping from coarse spatial resolution satellite data: production and accuracy assessment issues, International Journal of Remote Sensing, 22(14):

9 Aguiar, A.P.D., Shimabukuro, Y.E. and Mascarenhas, N.D.A., Use of synthetic bands derived from mixing models in the multispectral classification of remote sensing images, International Journal of Remote Sensing, 20(4): Almeida Filho, R. and Shimabukuro, Y.E., Digital processing of a Landsat-TM timeseries for mapping and monitoring degraded areas caused by independent gold miners, Roraima State, Brazilian Amazon. Remote Sensing of Environment, New York, v. 79, p Bins, L.S., Erthal, G.J. and Fonseca, L.M.G., Um método de classificação nãosupervisionada por regiões. In: Brazilian Symposium on Graphic Computational and Image Processing, 6., 1993, Recife. Proceedings, Rio de Janeiro: Gráfica Wagner, p Brown, I.F., Schroeder, W., Setzer, A., Maldonado, M., Pantoja, N., Duarte, A. and Marengo, J., Monitoring Fires in Southwestern Amazonian Rain Forests, EOS Transactions, 87: Carreiras, J.M.B., Shimabukuro, Y.E. and Pereira, J.M.C., Fraction images derived from SPOT-4 VEGETATION data to assess land-cover change over the State of Mato Grosso, Brazil, International Journal of Remote Sensing, 23, DeFries, R.S., Hansen, M.C. and Townshend, J.R.G., Global continuous fields of vegetation characteristics: a linear mixture model applied to multi-year 8 km AVHRR data, International Journal of Remote Sensing, 21, , Eva, H.D., Belward, A.S., Miranda E.E., Di Bella, C.M., Gonds, V., Huber, O., Jones, S., Sgrenzaroli, M. and Fritz, S., A landcover map of South America, Global Change Biology, 10(5), Friedl, M.A., McIver, D.K., Hodges, J.C.F., Zhang, X.Y., Muchoney, D., Strahler, A.H., Woodcock, C.E., Gopal, S., Schneider, A., Cooper, A. Baccini, A., Gao, F. and Schaaf, C., Global landcover mapping from MODIS: algorithms and early results, Remote Sensing of Environment, 83, Goreau, T.J., Mello, W.Z., Tropical deforestation: some effects on atmospheric chemistry, Ambio, 17 (1): Hansen, M., DeFries, R., Townshend, J.R.G. and Sohlberg, R., Global land-cover classification at 1 km resolution using a decision tree classifier, International Journal of Remote Sensing, 21(6/7):

10 Shimabukuro, Y.E. and Smith, J.A., The least-squares mixing models to generate fraction images derived from Remote Sensing multispectral data, IEEE Transactions on Geoscience and Remote Sensing, 29 (1): Shimabukuro, Y.E., Batista, G.T., Mello, E.M.K., Moreira, J.C. and Duarte, V., Using shade fraction image segmentation to evaluate deforestation in Landsat Thematic Mapper images of the Amazon region, International Journal of Remote Sensing, 19(3): Shimabukuro, Y.E., Duarte, V., Anderson, L.O., Valeriano, D.M., Arai, E., Freitas, R., Rudorff, B.F.T. and Moreira, M.A., Near real time detection of deforestation in the Brazilian Amazon using MODIS imagery, Revista Ambi-Água, 1, 1, Shimabukuro, Y.E., Duarte, V., Arai, E., Freitas, R.M., Lima, A., Valeriano, D.M., Brown, F. and Maldonado, M.L.R., Fraction images derived from Terra MODIS data for mapping burned areas in Brazilian Amazonia, International Journal of Remote Sensing, 30(6): Steininger, M.K., Tropical secondary regrowth in the Amazon: age, area and change estimation with Thematic Mapper data, International Journal of Remote Sensing, 17 (1):9-27.

11 List of Figures Figure 1: Location of the Study area: Acre State. Figure 2: Image mosaics of Acre State: 1980 MSS, 1990 TM, 2000 ETM+, and 2007 MODIS. Figure 3: Digital classification procedure. Figure 4: Fraction images derived from 2007 MODIS mosaic of Acre State: a) color composite (R6 G2 B1), b) vegetation fraction, c) soil fraction, and d) shade fraction images. Figure 5: Classification of multisensor (MSS 1980, TM 1990, ETM and MODIS 2007) data of Acre State. Regr. means regrowth areas and Def. means deforested areas. Figure 6: Estimate of land cover types in Acre State. Figure 7: Classification of burned areas using 2005 MODIS data of Acre State: (a) MODIS color composite (band 6(R) band 2(G) band 1(B)) based on burned pixels; and (b) classification of burned areas over MODIS vegetation fraction image. Figure 8: Classification of burned areas using 2005 MODIS multitemporal data of Acre State, showing the following classes: forest, deforestation, burned in forest, and burned in deforestation.

12 Figure 1: Location of the Study area: Acre State.

13 Figure 2: Image mosaics of Acre State: 1980 MSS, 1990 TM, 2000 ETM+, and 2007 MODIS.

14 Figure 3: Digital classification procedure.

15 Figure 4: Fraction images derived from 2007 MODIS mosaic of Acre State: a) color composite (R6 G2 B1), b) vegetation fraction, c) soil fraction, and d) shade fraction images.

16 Figure 5: Classification of multisensor (MSS 1980, TM 1990, ETM and MODIS 2007) data of Acre State. Regr. means regrowth areas and Def. means deforested areas.

17 Figure 6: Estimate of land cover types in Acre State.

18 Figure 7: Classification of burned areas using 2005 MODIS data of Acre State: (a) MODIS color composite (band 6(R) band 2(G) band 1(B)) based on burned pixels; and (b) classification of burned areas over MODIS vegetation fraction image.

19 Figure 8: Classification of burned areas using 2005 MODIS multitemporal data of Acre State, showing the following classes: forest, deforestation, burned in forest, and burned in deforestation.