Comparison of Atmospheric Correction Methods in

Download PDF
2 downloads 0 Views 253KB Size Report
Comment
low-intensity field sampling to obtain a general view of ... satellites, cloud-free images for creating multi-image mosaics are ... effects from the satellite data; including image-based ... multitemporal Landsat images after atmospheric correction.
05-060

1/11/06

3:08 AM

Page 155

Comparison of Atmospheric Correction Methods in Mapping Timber Volume with Multitemporal Landsat Images in Kainuu, Finland I. Norjamäki and T. Tokola

Abstract Using remote sensing to monitor large forest areas usually requires large field datasets. The need for extensive data collection can be reduced through interpretation of several images simultaneously. This study focused evaluating the accuracy and functionality of stand volume models in overlapping multi-temporal images that could form large areas covering a mosaic of scenes. Various atmospheric correction methods were tested to generalize field information outside the coverage of single images. A dataset consisting of three overlapping Landsat ETM images taken on different dates was used to compare atmospheric correction methods with uncorrected raw data. The methods tested were 6S, SMAC, and DOS. Aerosol data from MODIS were used in retrieving parameters for the 6S algorithm. The coefficient of determination values for the regression models used in estimating the total volume of the standing crop varied from 0.46 to 0.62 and standard error from 57 to 77 m3/ha, depending on the image calibration method used. All the atmospheric correction methods improved the classification of the multitemporal images. In comparison to the uncorrected data, the relative RMSE values for the multitemporal images decreased by an average of 6 percent on with DOS, 14 percent with SMAC, and 15 percent with 6S.

Introduction Material from remote sensing is typically combined with low-intensity field sampling to obtain a general view of forest resources. Earth observation data provide a practical tool for the mapping and frequent monitoring of landcover over large regions. Current optical satellite systems were used in regional forest inventories (Jaakkola and Saukkola, 1979; Jaakkola et al., 1988; Muinonen and Tokola 1990; Tomppo 1993; Bauer et al., 1994). The use of high-resolution data in regional surveys is limited mainly by the cost and difficulty of automatically interpreting detailed and complexly textured information (Hyppänen, 1996). The main methods used for the estimation of forest characteristics have been stratification of digital remotesensing data to homogeneous spectral classes, either according to an unsupervised or supervised scheme (e.g., Poso et al., 1984 and 1987; Horler and Ahern, 1986; Häme, 1991; Brockhaus and Khorram, 1992) and direct estimation of characteristics using regression analysis (e.g., Tomppo, 1987, 1992; Ripple et al., 1991; Ardö, 1992). The non-

parametric weighted kNN-based method (Kilkki and Päivinen, 1986; Muinonen and Tokola, 1990; Tomppo 1993 and 1998; Tokola et al. 1996; Trotter et al. 1997; Nilsson and Ranneby, 1997) was also used for similar purposes. All estimates derived using these methods will eventually be based on field data, while remote-sensing data are generally used to expand the data by interpolation over non-sampled areas. The variables of interest are modeled separately in the regression approach. In the stratification and weighted kNN approaches, however, several variables can be estimated simultaneously. A nonparametric method, such as the kNN approach, needs extensive reference field data and therefore is very expensive for large-scale forest inventories. The accuracy of satellite image-based forest inventories is highly dependent on the quality of the satellite data interpreted. The average time window for acquiring optical images for forest inventory purposes in Finland is under four months. The relatively long repeat time for alternative satellites, cloud-free images for creating multi-image mosaics are frequently not available. When such data exist, they usually consist of images from many different phases of the growing period of the forest, and the spectral characteristics of different optical satellite systems need to be calibrated. Since the cost of Landsat images has recently decreased, there has been growing interest in the use of multitemporal Landsat imagery, and this imagery type was previously studied (Helmer et al., 2000; Lefsky et al., 2001; Oetter et al., 2001; Song and Woodcock, 2002 and 2003; Hadjimitsis, et al., 2004). Often other satellite data are required to fill the gaps between existing Landsat mosaics. There are many factors that cause uncertainty in the use of multitemporal satellite data: aging of the instrument, atmospheric conditions, topography, phenology, distance of the target to the sun, and sun and view angles. However, the problem is normally avoided using relative normalization among images (Olsson, 1993 and 1995; Tokola et al., 1999; Cohen et al., 2001). Another approach, absolute image calibration, is still an attractive alternative, although there are many difficulties in modeling all the physical conditions required. In optical remote sensing, the atmosphere is the primary source of noise preventing the accurate measurement of surface reflectance

Photogrammetric Engineering & Remote Sensing Vol. 73, No. 2, February 2007, pp. 155–163. Department of Forest Resource Management, University of Helsinki, Finland ([email protected]). PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

0099-1112/07/7302–0155/$3.00/0 © 2007 American Society for Photogrammetry and Remote Sensing February 2007

155

05-060

1/11/06

3:08 AM

Page 156

(Song and Woodcock, 2003). These factors are additive or act in combination with multiplicative factors; thus ratiobased techniques give only approximate results (Egan, 2004). An important atmosphere-related factor that reduces radiometric accuracy is the proportion of aerosols and gases present, which directly affects the amount of scattering and absorption registered by the instrument. Another important effect is caused by satellite-sun geometry. Various algorithms were proposed for removing these effects from the satellite data; including image-based algorithms (Chavez, 1988, 1989, and 1996; Liang, 2001 and 2002; Song et al., 2001; Song and Woodcock, 2003) and a number of methods that use radiative transfer codes (RTCs), which require in situ measurements of atmospheric conditions (Kneizys et al., 1988; De Haan et al., 1991; Rahman and Dedieu 1994; Vermote et al., 1997; Hu et al., 2001). In addition, combinatorial methods that use timedependent aerosol measurements and image-based information were developed (Liang et al., 1997; Ouaidrari and Vermote, 1999; Wen et al., 1999). Algorithms were well reviewed in Liang (2004). In Finland, environmental authorities, paper companies, and teleoperators use regional forest maps on scales of 1:50 000–1:100 000 for planning tasks as well as for timber procurement. Large-area forest inventory based on Landsat image interpretation is often a suitable method for producing such maps. Normally, the interpretation of multiple images requires large field datasets from the areas in each scene, which makes interpretation costly. Here, various atmospheric correction methods were tested to create a multitemporal image mosaic covering the target area and apply the same interpretation procedure to the entire area. The method can be used when a large area covering the same field dataset (training data) is used for the entire area of the image mosaic. This study is mainly focused on evaluating the accuracy and functionality of stand volume models in multitemporal Landsat images after atmospheric correction and image calibration. Other variables of interest include height and the proportion of different tree species. The

models predicting the proportion of different tree species are used to estimate the dominant tree species.

Material Tests for the atmospheric correction methods were performed using three Landsat ETM images taken on different dates (Table 1) in the region of Kainuu, northern Finland (Figure 1). About 95 percent of the land area of Kainuu is forested and the dominant tree species is Scotch pine (Pinus sylvestris L.). The average volume for the forests is 73 m3/ha. Dryish mineral soils predominate in the eastern and northern parts of the province, while peatlands are characteristic for the western areas. Most of the forests within the study area are privately owned and relatively intensively managed. The data were distributed in two datasets, located about 150 km apart (Figure 1). The first of the datasets was measured during the summer of 2002 and included 277 sample plots. The remaining sample plots (167) were measured during the summer of 2003. Sampling of the field data was done from U-shaped clusters, each of which contained 8 to 15 sample plots 150 to 200 m apart. Advance information (old classification data and visual interpretation) was used so that the data to be collected would include as much variation in height, basal area, and tree species composition as possible. The sample plots were located using GPS, and the suitability of the plot was evaluated. If the field plot hit a stand in which recent cutting or natural damage had occurred, it was rejected. It was also required that the measured plot be located in a stand with a minimum area of 0.5 ha and that the distance TABLE 1.

LANDSAT

ETM IMAGES

USED

IN THE

STUDY

Path

Row

Acquisition Date

188 188 188

15 15 15

29052002 17082002 26072000

Figure 1. Study area and coverage of the Landsat images used.

156

February 2007

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

05-060

1/11/06

3:08 AM

Page 157

TABLE 2.

MEAN VOLUMES

OF THE

FIELD DATA

Mean Vol. (m3/ha) Mineral soils Seedling stands Young aged stands Middle aged stands Old forest stands Peatlands Seedling stands Young aged stands Middle aged stands Old forest stands

TABLE 3.

AND THE

SHARE

OF

PINE, SPRUCE

AND

DECIDIOUS PLOTS

BY

FOREST CLASSES

No. of Plots

Pine Plots (%)

Spruce Plots (%)

16.44 85.15 154.41 244.58

39 95 135 46

35.90 64.21 62.96 28.26

25.64 10.53 21.48 60.87

38.6 25.26 15.56 10.87

18.14 53.92 84.37 221.15

6 66 49 8

83.33 66.67 77.55 37.50

16.67 19.70 10.20 37.50

0.00 13.64 12.24 25.00

VOLUME CLASS DISTRIBUTION

OF THE

Class 50 m3/ha 50–99 m3/ha 100–149 m3/ha 150–199 m3/ha 200–249 m3/ha 250 m3/ha

FIELD DATA Plots in Class 86 105 116 72 35 30

to the stand border be more than 25 m. The basal area (m2/ha) and mean height were measured by tree species, and the volume was calculated using the volume equations of Nyyssönen (1954). Other variables such as fertility and age were also assessed. Of the 444 sample plots, 315 were situated in mineral soils and 129 on peatlands (Table 2). The mean volume for the sample plots was 130 m3/ha in mineral soils and 104 m3/ha on peatlands. Scotch Pine was the dominant tree species in the field data collected. There was a lack of deciduous stands on peatlands, which was due to the natural tree species composition of the study area. The volume distribution of the field data was also reviewed by classifying the sample plots into volume classes of 50 m3/ha (Table 3). The distribution between classes was fairly even to as much as 20 m3/ha, after which the number of sample plots per class decreased.

Methods Estimation Method for Stand Characteristics The present study was based on multitemporal Landsat images and field data that were used to in supervised estimation. Separate stand volume models based on the field data and spectral values were created for raw images, as well as for atmospherically corrected images, referred to as reference models. The quality of the reference models was tested by comparison with other images. The result was evaluated by classifying the estimates in classes of 50 m3/ha. The proportions of estimates falling into the correct class and neighboring classes were calculated. The 50 m3/ha classes were chosen for practical purposes (Table 4). After the most viable atmospheric correction method was found, new regression models were created to predict the proportion of different tree species (deciduous trees, pine, and spruce). For modeling the volume, an ordinary least squares (OLS) regression was applied, and for modeling the proportions of tree species a seemingly unrelated regression (SUR) method (e.g., Zellner, 1962; Johnston, 1972; Binkley and Nelson, 1988) was used. An SUR can be used for estimating the result of several equations simultaneously when a set of equations is assumed to have cross-equation PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

TABLE 4. Reference Image 26072000 31052002 17082002 26072000 31052002 17082002 26072000 31052002 17082002 26072000 31052002 17082002

KEY STATISTICS

Calibration Method Raw Raw Raw DOS3 DOS3 DOS3 SMAC SMAC SMAC 6S 6S 6S

FOR THE

Decidious Plots (%)

REFERENCE MODELS

Se, m3/ha

r2

Bias, m3/ha

sdb

n

57.21 63.14 62.58 57.27 57.69 62.25 58.83 63.40 63.03 59.60 62.32 61.79

0.602 0.494 0.485 0.623 0.575 0.489 0.517 0.482 0.461 0.596 0.473 0.545

0.35 1.24 1.17 1.06 1.44 1.19 2.46 1.55 1.78 1.44 1.80 1.29

2.71 3.00 2.97 2.72 2.74 2.95 2.79 3.01 2.99 2.83 2.96 2.95

444 444 444 444 444 444 444 444 444 444 444 444

error correlation. If disturbances in the various equations are correlated, joint estimation with SUR is in general more efficient than a separate estimation using OLS (Binkley and Nelson, 1988). The coefficients for the set of models are obtained first with OLS, the covariance matrix between the error terms of the models is calculated, and finally new coefficients for the set of equations are estimated. The estimates of tree species proportions were used to classify the pixels according to the estimated dominant tree species. The kappa value (Rosenfeld and Fitzpatrick-Lins, 1986) was used as a measure of classification accuracy, and the statistical significance of the differences in kappa values before and after correction were tested with the t-test. The t-test values were calculated from the standard error of the kappa estimates (Terry, 1987). The form of the stand growing stock models (m3/ha) was y  e, and for the proportion of species volumes the form of the model was arcsin(sqrt(y))  f(), where   the linear combination of spectral values from Landsat bands 2 through 5 and their transformations to reflectance following atmospheric correction. The ratio, product, logarithm, and square root of the original bands were used in the transformations. Atmospheric Correction Methods Three different methods were tested with image calibration: (a) Dark Object Subtraction (DOS), (b) Simplified Method for Atmospheric Correction (SMAC), and (c) the Second Simulation of the Satellite Signal in the Solar Spectrum (6S) atmospheric model. The first is a nonparametric image-based method that is suitable for areas with dense vegetation. February 2007

157

05-060

1/11/06

3:08 AM

Page 158

When atmospheric scattering optical depths and aerosol parameters are known from external sources, RTCs, such as SMAC and 6S, can be used in atmospheric calibration. The results from atmospherically corrected images were compared with those calculated from raw data with no atmospheric correction applied. DOS-based methods are widely used for atmospheric correction because they are relatively simple to use and utilize only the information derived from the image itself. The DOS-based approach assumes the existence of such objects that have zero or near zero surface reflectance. With this assumption, the minimum sensor signal (DN) value in the histogram is considered to be an effect of the atmosphere and is subtracted from all the pixels (Chavez, 1989). Song et al. (2001) studied the accuracy of four different DOS-based methods with respect to classification and change detection. The best results were achieved using a DOS3 method. This model was adopted for our study with the exception that the diffuse downward radiation at the surface was considered to be zero. Surface reflectance is calculated in the DOS3 model as presented by Kaufman and Sendra (1988) and by using certain simplifying assumptions. The model computes atmospheric transmittance from the target towards the sensor and the transmittance in the illumination direction, assuming the presence of only Rayleigh scattering but no aerosols. The optical thickness for Rayleigh scattering is estimated using the method of Kaufman (1989). Due to atmospheric scattering effects, the path radiance is estimated assuming 1 percent surface reflectance for dark objects (Chavez, 1989 and 1996; Moran et al., 1992; Song et al., 2001). The second technique used here was the SMAC method (Rahman and Dedieu, 1994; Häme et al., 2001), which is a semi-empirical atmospheric correction method based on the 5S model (a Simplified Method for the Atmospheric Correction of Satellite Measurements in the Solar Spectrum). In this model, the raw digital counts are first converted to top-ofatmosphere (TOA) reflectances, using time-dependent calibration coefficients. These reflectances are then converted to atmospherically corrected reflectances. The most important of the input parameters for the algorithm is aerosol optical depth (AOD). The ratio of the TOA reflectances of ETM channels 3 and 7 are used for defining the AOD value. Here, an AOD surface was computed, and the mean AOD value was used in atmospheric correction and calibration. The water and ozone contents were assigned default values. The third method, the 6S model, is based on the radiative transfer theory developed by Chandsarekhar (1950) and takes into account the main atmospheric effects: gaseous absorption by water vapor, oxygen, ozone, and carbon dioxide and scattering caused by aerosols and molecules. The input parameters for the model are the sun-sensor geometry, atmospheric model for gaseous components, aerosol model (type and concentration), AOD, ground reflectance, and spectral band. The model requires several measurements of atmospheric optical properties at the time of image acquisition, which often limits the use of the model. The unavailability of accurate atmospheric data is one of the main reasons why in many applications only such correction algorithms that use the information derived from the image itself are used operationally. Many of the parameters used in the 6S model can be derived using the data from the Moderate-Resolution Imaging Spectroradiometer (MODIS) satellite (e.g., Kaufman et al., 1997). The accuracy of MODIS was studied by comparing the observed data with the data measured using ground-based sun photometers in the Aerosol Robotic Network (AERONET) which can achieve an AOD accuracy of  (0.01  0.02a) (Holben et al., 1998 and 2001). Most of the MODIS aerosol retrievals, according to Chu et al. (2002), are found within the retrieval error of a  158

February 2007

(0.05  0.2a). In present study, the parameters derived from the MODIS data included the altitude, pressure, temperature, and H2O and O3 densities, which were computed from the MODIS Atmosphere Profile product (MOD07_L2) dataset (Menzel et al., 2002). The AOD at 550 nm was computed from the MODIS Atmosphere Profile product (MOD04_L2) dataset (Kaufman and Tanré, 1998). Parameters were calculated as averages from cloud-free pixels in a window of 10 km 10 km. Here, the aerosol model was maintained at constant levels. Prior to running the code, the sensor signal was converted to at-satellite radiance using the relation Lsat  G(DN)  B, where G is the sensor gain, and B the bias. However, the applicability of MODIS data is dependent on the vegetation cover. Estimation of aerosol parameters is based on the DOS approach and requires dense vegetation in the target area. This requirement is fulfilled in widespread areas of Scandinavia.

Results The methods used for retrieving the AOD were different for 6S and SMAC. AOD estimates derived for SMAC were generally larger than for 6S. The mean AOD values computed for SMAC were 0.15 (image 26.7.2000), 0.18 (image 29.5.2002), and 0.21 (image 17.8.2002); the AOD values retrieved from MODIS were 0.08, 0.04, and 0.16, respectively. Variation also existed in AOD within the coverage of the images (Figure 2); the effect of intra image AOD variation was ignored in both cases (6S and SMAC). Since the image acquired in August was relatively cloudy (14 percent), its atmospheric optical properties greatly differed from those of the other two images. The most significant differences were in the AOD value and water vapor content. This increased the uncertainty of image correction, which can also be seen in the volume estimation accuracy (Table 5). The standard error values for the reference models were between 57.21 and 63.40 m3/ha and the bias was between 0.35 and 2.46 m3/ha (Table 3). Atmospheric correction did not significantly affect the accuracy of the reference models. Regardless of the correction method employed, the most accurate models were achieved when

Figure 2. The interpolated AOD surface (29.5.2002) within the study area (100 km 100 km) at 550 nm (MODIS retrieval).

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

05-060

1/11/06

3:08 AM

Page 159

TABLE 5.

EFFECT

OF

ATMOSPHERIC CORRECTION

Reference Image

Compared Image

se, m3/ha

Raw

31052002

26072000 17082002

Raw

26072000

63.14 64.63 66.71 57.21 76.54 71.34 62.58 64.21 58.30 57.69 61.59 67.08 57.27 69.03 64.19 62.25 67.16 59.04 63.40 67.13 66.74 58.83 72.29 74.49 63.03 67.37 59.21 62.32 63.51 63.64 59.60 70.71 76.69 61.79 66.11 58.14

Correction Method

31052002 17082002 Raw

17082002 31052002 26072000

DOS3

31052002 26072000 17082002

DOS3

26072000 31052002 17082002

DOS3

17082002 31052002 26072000

SMAC

31052002 26072000 17082002

SMAC

26072000 31052002 17082002

SMAC

17082002 31052002 26072000

6S

31052002 26072000 17082002

6S

26072000 31052002 17082002

6S

17082002 31052002 26072000

the satellite data from July and the most inaccurate when the satellite data from late May were used; the differences were, however, insignificant. The reference model was applied to two similarly processed images from another point in time, and the statistics were calculated. The result was evaluated by classifying the estimates in classes of 50 m3/ha. The proportions of estimates falling into the correct class and neighboring classes were calculated. In general, the standard error estimates for overlapping images increased to some extent, but the most significant was the increment of bias (Table 5 and Figure 3). When the raw reference models were applied (with no correction procedure) to two other neighboring images, the bias was between 77.21 and 78.50 m3/ha. The amount of bias has a direct deteriorating effect on the classification results. All atmospheric correction methods improved the classification results of the multitemporal images used. When the SMAC and 6S results were compared with the uncorrected data, SMAC decreased the bias by a mean of 67 percent whereas 6S decreased the bias by 61 percent (Figure 3). The biases for the DOS3-corrected images remained quite high decreasing a mean of 12 percent. Then, differences between relative root-meansquare-error (RMSE) values were minor compared to differences in bias values (Table 5). This indicates that atmospheric correction method mainly correct systematic error. The results from the 6S and SMAC corrections were similar. When the reference SMAC models were applied, the biases for the two overlapping images were between 26.08 and 20.47 m3/ha. Of the estimates, 22.97 to 39.41 percent fell into the correct class and 69.37 to 82.88 percent into the correct or neighboring class. The biases for the 6S-corrected PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

RMSE,

ON

%

51.6 52.8 54.5 46.8 62.6 58.3 51.1 52.5 47.7 47.2 50.3 54.8 46.8 56.4 52.5 50.9 54.9 48.3 51.8 54.9 54.5 48.1 59.1 60.9 51.5 55.1 48.4 50.9 51.9 52.0 48.7 57.8 62.7 50.5 54.0 47.5

KEY STATISTICS Bias, m3/ha

Correct Class (%)

Class 1 (%)

1.24 52.55 65.75 0.35 18.14 71.26 1.17 78.50 38.94 1.44 35.04 77.21 1.06 19.17 46.63 1.19 70.52 37.09 1.55 18.36 23.45 2.46 5.67 26.08 1.78 20.47 13.39 1.80 10.40 24.72 1.44 10.58 21.67 1.29 37.64 21.40

37.16 18.24 15.54 40.09 37.39 15.32 37.16 23.87 38.96 36.04 20.05 11.04 42.12 38.51 18.47 37.84 26.58 39.64 36.04 22.97 24.10 40.32 37.16 31.31 37.16 38.29 39.41 34.62 27.38 28.28 38.91 39.09 37.78 37.05 35.23 38.18

80.18 59.68 49.77 84.01 73.65 47.75 79.95 53.60 75.23 79.05 70.95 40.77 84.23 75.45 55.18 79.73 56.98 75.00 79.73 70.50 69.37 81.98 76.58 73.42 81.76 77.70 82.88 80.09 76.47 71.49 81.90 76.59 73.30 78.41 72.73 81.36

datasets were between 24.72 and 37.64 m3/ha. Of the estimates, 27.38 to 39.09 percent fell into the correct class and 71.49 to 81.36 percent into the correct or neighboring class. The classification accuracy decreased when the volume of the sample plot exceeded 200 m3/ha (Figure 4). As generated with the classification method proposed, the total volume of growing stock at the municipal level was overestimated by a mean of 12 percent compared with the statistics from the ninth National Forest Inventory (NFI). The test was done using the statistics from five municipalities around the study area. The effect of atmospheric correction on the enhancement of tree species estimation accuracy was inconsequential, since the tree species models did not perform particularly well in any situation. When uncorrected reference datasets were used, an average of 62 percent of the plots were included in the correct dominant tree species class, and when the compared overlapping datasets were used, the result was 57 percent. The corresponding figures for 6S-corrected images were 62 percent (reference datasets) and 58 percent (compared datasets). When kappa figures were computed for the dominant tree species classification result, the kappa values were between 0.159 and 0.185 for models using uncorrected spectral values, whereas in 6S-corrected models they were between 0.123 and 0.181 (Table 6). When the models were applied to the overlapping images, the kappa values were 0.030 to 0.241 for uncorrected images and 0.027 to 0.240 for 6S-corrected images. Pure deciduous stands (with a proportion of deciduous species of over 95 percent) were recognized accurately in uncorrected reference images and compared images for a mean of 59 percent and 40 percent of February 2007

159

05-060

1/11/06

3:08 AM

Page 160

Figure 3. The root of difference of mean square error (m3/ha) between reference model and calibrated image model shows the overall error caused by different calibration methods (line). The bias (m3/ha) indicates correct level of calibration of different methods for the compared multitemporal images.

the time, respectively. The corresponding figures for 6S corrected images were 53 percent and 49 percent. When the number of species increased, the accuracy of dominant tree species estimation decreased. Distinguishing between pine and spruce plots was rather unreliable. In all cases less than 20 percent of the spruce plots were recognized, since they were dominated by pine plots. Almost all the pine plots were classified correctly, but many of the spruce and deciduous plots were also classified as pine.

Discussion

Figure 4. Timber volume classification accuracy for two multitemporal images corrected with 6S. The reference model is created using the July image.

TABLE 6.

Reference Model May July August

160

COMPARISON

OF TREE SPECIES CLASSIFICATION BETWEEN UNCORRECTED AND 6S SIGNIFICANCY FIGURES ARE ALL SIGNIFICANT AT THE 0.05 LEVEL

Kappa (Uncorrected Image Classification) May 0.185 0.030 0.148

February 2007

July 0.241 0.159 0.173

The standard error for reference models predicting total timber volume varied between 57 and 63 m3/ha, and the coefficient of determination values were 0.47 to 0.62. In Landsat images, the results were similar to those obtained here (Tomppo, 1987; Ardö, 1992; Tokola et al., 1996). Atmospheric correction had little effect on the classification of the reference image, which was also shown in several

August 0.109 0.021 0.172

Kappa (6S Corrected Image Classification) May 0.123 0.027 0.034

July 0.147 0.181 0.218

August 0.240 0.101 0.161

CORRECTED DATASETS.

Sig. of Differences Between Uncorrected and 6S, t May 27.548 2.980 67.942

July 42.133 9.933 20.234

August 54.197 56.253 5.042

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

05-060

1/11/06

3:08 AM

Page 161

previous studies (Fraser et al., 1977; Kawata et al., 1990; Song et al., 2001). As expected, estimation of dominant tree species was quite unreliable. Similar results for speciesspecific estimation were obtained in other studies (e.g., Franklin, 1994; Tokola et al., 1996; Huguenin et al., 1997). Although SUR models resulted in more stable results than separate regression models, the overall accuracy was only slightly better than with separate model groups. Pure deciduous tree plots can be recognized with reasonable accuracy, but separation between pine and spruce stands using the empirical models presented here is prone to extensive bias. The dominant coniferous tree species in the field data also predominated in the estimation. In practical forest inventories it may not be realistic to attempt to separate different coniferous species, but rather classify the desired area into deciduous and coniferous pixels. Using auxiliary information, it may be possible to improve tree species estimation. If the tree species composition is known by sub-area, it could be used to assign weights for the estimation. The selection of the dark object is a crucial step in DOSbased models. The use of a wrong dark object value in computation can lead to a significant increment in bias. Song et al. (2001) studied the accuracy of image-based correction algorithms and showed that the DOS3 model gave the best results with respect to classification and change detection. In another study carried out by Song et al. (2003), the simple DOS3 method was compared with methods that use externally measured aerosol data. They found that DOS3 improved the vegetation index (NDVI) value, but still under-corrected the image. Moreover, in estimating greenness DOS3 did not give as favorable a result as the other two algorithms that utilized aerosol data. In our study DOS3 improved the estimation of forest characteristics to some extent, but not enough for empirical models to work well in classifying multitemporal images. Absolute atmospheric correction methods can convert satellite measurements accurately into surface reflectances (Holm et al., 1989; Moran et al., 1992), and this leads to improved classification results when multitemporal satellite images are used. Time-dependent input parameters are needed for the SMAC and 6S methods. This makes their use more complicated than the use of purely image-based algorithms. The atmospheric degradations are expressed in terms of optical depths, in which an optical depth of unity results in an attenuation of 63 percent (Egan, 2004). The vertical optical depth of mid-latitude aerosols undergoes a factor 2 to 10 seasonal variation, depending on the land-use (Egan, 2004). At present, there are sources such as MODIS from which many of these parameters can be obtained with sufficient accuracy. Cloudiness and haze (when the scene includes semitransparent cloud and aerosol layers) can dilute the quality of MODIS data. As was the case here, atmospherically clear scenes are seldom encountered in Finland. Haze can arise from a variety of atmospheric elements, such as water droplets, ice crystals, or fog/smog particles (Kaufman, 1989). The influence of haze on measured radiance is most significant in the visible spectral region (Zhang et al., 2002). For hazy scenes, ancillary data upon which to base an absolute atmospheric correction are often lacking. The imagebased method presented by Liang et al. (2001) showed good results in separating heterogeneous aerosol scattering effects, especially when small scale-variation was clear and identification of hazy regions was obvious. The MODIS data used in this study were unsuitable for the image acquired in August, which evidently increased the uncertainty of the 6S correction. In the present study, differences of no more than 10 to 15 percent after correction at most were seen in forest area reflectance between the images used. Several factors in Finland’s forests could lead to bias when a large-area field sample is used in a small-area estimaPHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

tion. One significant factor in uncertainty is the variability in surface reflectance due to phenology (Song and Woodcock, 2003). Even if the leaf spectral properties and the amount of leaves in the canopy could be assumed to be similar among the images used, the observed reflectance may vary within the season. This is due to the changing sun angle, which causes variability in the amount of shadows cast in the canopy. The images used in this study were acquired in different sections of the growing period. Local variation in the development of different plants always occurs especially at the beginning of the growing period. If forest canopies have varying phenology on different acquisition dates, atmospheric correction may not always allow empirical models from one date to be applied to another image. Due to the relatively warm summer of 2002, the leaves of deciduous trees were already full-grown by late May within the study area, although the phenology difference from May to August was quite significant. On the other hand, the phenology between the July and August datasets was hardly notable. Differences in forest stand structure between forest ownership groups are another reason for local variation in forested areas. Soil fertility factors also vary locally within the vegetation zones. Even though the sample plots used in this study (Figure 1) were located within the same plant ecological zone, variability was shown in the soil factors. In the conditions prevailing in Finland, soil type affects the spectral responses received from forested areas. If digital ancillary data are available, e.g., forest or soil type maps concerning the target area, the reliability of sub-area estimation can be improved and more representative field samples chosen based on a priori information (Tokola and Heikkilä, 1997). Geographic distance between sample areas can as easily cause bias in estimates. The best interpretation results using satellite images are achieved when the field data are collected within a 20 km search radius (Tokola, 2000; Katila and Tomppo, 2001; Lappi, 2001). Even though the distance between the remotest plots used in this study was more than 100 km, the distribution of the error terms in the models did not appear to be geographically dependent. The reliability of estimates is also dependent on the size of the forest stand (Poso et al., 1987). The western part of the study area is located in a region with no significant changes in elevation, whereas the eastern part is hilly but slopes gently. Topographic variation may cause bias in satellite image interpretation if no normalization is applied. The primary topographic effect is the change in direct solar radiation on a sloping surface, due to the changing incidence angle between the sun and surface normal. In more mountainous areas, the state of the atmosphere and vegetation is also highly influenced by the relief (Ekstrand, 1996). Song et al. (2003) showed that the NDVI and wetness indices are more resistant to topographic influence than brightness and greenness. Topographic normalization of Landsat TM image pixels improved the interpretation of results of forest site fertility classes (Tomppo, 1992). Here, no topographic normalization was applied, because the sample data were collected only from gently sloping or totally flat surfaces. Still, topographic variation caused uncertainty in classification of other parts of the image. In comparison to the uncorrected data, the relative RMSE values for the multitemporal images decreased by a mean of 6 percent with DOS, 14 percent with SMAC, and 15 percent with 6S. Atmospheric correction methods mainly correct systematic error of timber volume estimation. Although all the atmospheric correction methods improved the classification of the multitemporal images, the importance of ground truth is obvious. There are several reasons, why similar spectral responses are recorded in the satellite from different objects. February 2007

161

05-060

1/11/06

3:08 AM

Page 162

Still, the MODIS data can provide a good data source for estimating optical variation within Landsat scenes. For regional planning and large-scale inventories, the importance of unbiased estimates can be more important than precision pixel estimates. Estimates could be calibrated regionally to a more reliable level, through the use of statistically accurate background information. In Finland, for example, NFI municipal estimates or forest center statistics could be used as calibration references and consistency with different data sources could be achieved.

References Ardö, J., 1992. Volume quantification of coniferous forest compartments using spectral radiance recorded by Landsat Thematic Mapper, International Journal of Remote Sensing, 13: 1779–1786. Bauer, M.E., T.E. Burk, A.R. Ek, P.R. Coppin, S.D. Lime, T.A. Walsh, D.K. Walters, W. Befort, and D.F. Heinzen, 1994. Satellite inventory of Minnesota forest resources, Photogrammetric Engineering & Remote Sensing, 60:287–298. Binkley, J.K., and C.H. Nelson, 1988. A note on the efficiency of seemingly unrelated regression, American Statistician, 42:137–139. Brockhaus, J.A., and S. Khorram, 1992. A comparison of SPOT and Landsat TM data for use in conducting inventories of forest resources, International Journal of Remote Sensing, 13: 3035–3043. Chandrasekhar, S. 1950. Radiative Transfer, Oxford University Press, U.K., 393 p. Chavez, P.S., Jr., 1988. An improved dark-object subtraction technique for atmospheric scattering correction of multispectral data, Remote Sensing of Environment, 24:459–479. Chavez, P.S., Jr., 1989. Radiometric calibration of Landsat Thematic Mapper multispectral images, Photogrammetric Engineering & Remote Sensing, 55:1285–1294. Chavez, P.S., Jr., 1996. Image-based atmospheric corrections – Revisited and improved, Photogrammetric Engineering & Remote Sensing, 62:1025–1036. Chu, D.A., Y.J. Kaufman, C. Ichoku, L.A. Remer, D. Tanré, and B.N. Holben, 2002. Validation of MODIS aerosol optical depth retrieval overland, Geophysical Research Letters, 29(12). Cohen, W.B., T.K. Maiersperger, T.A. Spies, and D.R. Oetter, 2001. Modeling forest cover attributes as continuous variables in a regional context with Thematic Mapper data, International Journal of Remote Sensing, 22(12):2279–2310. De Haan, J.F., J.W. Hovenier, J.M.M. Kokke, and H.T.C. Stokkom, 1991. Removal of atmospheric influences on satellite-borne imagery: A radiative transfer approach, Remote Sensing of Environment, 37:1–21. Egan, W.G., 2004. Optical Remote Sensing, Science and Technology, Marcel Dekker, Inc., New York, 506 p. Ekstrand, S. 1996. Landsat -TM based forest damage assessment: Correction for topographic effects, Photogrammetric Engineering & Remote Sensing, 62:151–161. Franklin, S.E., 1994. Discrimination of subalpine forest species and canopy density using digital CASI, SPOT PLA, and Landsat TM data, Photogrammetric Engineering & Remote Sensing, 60:1233–1241. Fraser, R.S., O.P. Bahethi, and A.H. Al-Abbas, 1977. The effect of atmosphere on the classification of satellite observation to identify surface features, Remote Sensing of Environment, 6:229–249. Hadjimitsis, D.G., C.R.I. Clayton, and V.S. Hope, 2004. An assessment of the effectiveness of atmospheric correction algorithms through the remote sensing of some reservoirs, International Journal of Remote Sensing, 25:3651–3675. Häme, T. 1991. Spectral interpretation of changes in forest using satellite scanner images, Helsinki, Acta Forestalia Fennica, 222, 111 p. Häme, T., P. Stenberg, K. Andersson, Y. Rauste, P. Kennedy, S. Folving, and J. Sarkeala, 2001. AVHRR-based forest proportion 162

February 2007

map of the Pan-European area, Remote Sensing of Environment, 77:76–91. Helmer, E.H., S. Brown, and W.B. Cohen, 2000. Mapping montane tropical forest successional stage and land use with multi-date Landsat imagery, International Journal of Remote Sensing, 21:2163–2183. Holben, B.N., T.F. Eck, I. Slutsker, D. Tanré, J.P. Buis, A. Setzer, E. Vermote, J.A. Reagan, Y.J. Kaufman, T. Nakajima, F. Lavenu, I. Jankowiak, and A. Smirnov, 1998. AERONET – A federated instrument network and data archive for aerosol characterization, Remote Sensing of Environment, 66:1–16. Holben, B.N., D. Tanré, A. Smirnov, T.F. Eck, I. Slutsker, N. Abuhassan, W.W. Newcomb, J.S. Schafer, B. Chatenet, F. Lavenu, Y.J. Kaufman, J. Vande Castle, A. Setzer, B. Markham, D. Clark, R. Frouin, R. Halthore, A. Karnieli, N.T. O’Neill, C. Pietras, R.T. Pinker, K. Voss, and G. Zibordi, 2001. An emerging groundbased aerosol climatology: Aerosol optical depth from AERONET, Journal of Geophysical Research, 106:12067–12097. Holm, R.G., R.D. Jackson, B. Yuan, M.S. Moran, P.N. Slater, and S.F. Bigger, 1989. Surface reflectance factor retrieval from Thematic Mapper data, Remote Sensing of Environment, 27:47–57. Horler, D.N.H., and F.J. Ahern, 1986. Forestry information content of Thematic Mapper data, International Journal of Remote Sensing, 7:405–428. Hu, C., F.E. Muller-Karger, S. Andrefouet, and K.L. Carder, 2001. Atmospheric correction and cross-calibration of LANDSAT-7/ETM imagery over aquatic environments: A multiplatform approach using SeaWiFS/MODIS, Remote Sensing of Environment, 78: 99–107. Huguenin, R.L., M.A. Karaska, D. Van Blariconi, and J.R. Jensen, 1997. Subpixel classification of bald cypress and tupelo gum trees in Thematic Mapper imagery, Photogrammetric Engineering & Remote Sensing, 63:717–725. Hyppänen, H., 1996. Spatial autocorrelation and optimal spatial resolution of optical remote sensing data in boreal forest environment, International Journal of Remote Sensing, 17: 3441–3452. Jaakkola, S., and P. Saukkola, 1979. Timber volume estimation and cutting opportunity mapping using multispectral remote sensing techniques, The Photogrammetric Journal of Finland, 8(1). Jaakkola, S., S. Poso, and G. Skråmo, 1988. Satellite remote sensing for forest inventory - Experiences in the Nordic countries, Scandinavian Journal of Forest Research, 3:545–567. Johnston, J. 1972. Econometric Methods, 2nd edition. McGraw-Hill. 437 p. Katila, M., and E. Tomppo, 2001. Selecting estimation parameters for the Finnish multisource National Forest Inventory, Remote Sensing of Environment, 76:16–32. Kaufman, Y.J., and C. Sendra, 1988. Algorithm for automatic atmospheric corrections to visible and near-IR satellite imagery, International Journal of Remote Sensing, 9:1357–1381. Kaufman, Y.J., 1989. The atmospheric effect on remote sensing and its corrections, Theory and Applications of Optical Remote Sensing (G. Asrar, editor), Wiley & Sons, Inc., New York, pp. 336–428. Kaufman, Y.J., A.E. Wald, L.A. Remer, Bo-Cai Gao, Rong-Rong Li, and L. Flynn, 1997. The MODIS 2.1-um channel – Correlation with visible reflectance for use in remote sensing of aerosol, IEEE Transactions in Geoscience and Remote Sensing, 35: 1286–1298. Kaufman, Y.J., and D. Tanré, 1998. Algorithm for Remote Sensing of Tropospheric Aerosol from MODIS – Algorithm Theoretical Basis Document, Products: MOD04_L2, MOD08_D3, MOD08_E3, MOD08_M3, ATBD Reference Number: ATBD-MOD-02, URL: http://modis.gsfc.nasa.gov/data/atbd/atbd_mod02.pdf (last date accessed: 13 November 2006). Kawata, Y., A. Ohtani, T. Kusaka, and S. Ueno, 1990. Classification accuracy for the MOS-1 MESSR data before and after the atmospheric correction, IEEE Transactions in Geoscience and Remote Sensing, 28:755–760. Kilkki, P., and R. Päivinen, 1986. Reference sample plots to combine field measurements and satellite data in forest inventory, Remote Sensing-aided Forest Inventory, Proceedings of seminars organized by SNS, Hyytiälä, Finland, 10–12 December, University PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

05-060

1/11/06

3:08 AM

Page 163

of Helsinki, Department of Forest Mensuration and Management, Research Notes, No. 19:209–215. Kneizys, F.X., E.P. Shettle, W.O. Gallery, J.H. Chetwynd, L.W. Abreu, J.E.A. Selby, S.A. Clough, and R.W. Fenn, 1988. Atmospheric Transmittance/Radiance: Computer Code LOWTRAN-7, Air Force Geophysics Lab, Hanscom AFB, Massachusetts, AFGL-TR88-0177. Lappi, J., 2001. Forest inventory of small areas combining the calibration estimator and a spatial model, Canadian Journal of Forest Research, 31:1551–1560. Lefsky, M.A., W.B. Cohen, and T.A. Spies, 2001. An evaluation of alternate remote sensing products for forest inventory, monitoring, and mapping of Douglas-fir forests in western Oregon, Canadian Journal of Forest Research, 31:78–87. Liang, S., H. Fallah-Adl, S. Kalluri, J. JaJa, Y.J. Kaufman, and J.R.G. Townshend, 1997. An operational atmospheric correction algorithm for Landsat Thematic Mapper imagery over the land, Journal of Geophysical Research – Atmospheres, 102(D14), 17173–17186. Liang, S., H. Fang, and M. Chen, 2001. Atmospheric correction of Landsat ETM land surface imagery – Part I: Methods, IEEE Transactions on Geoscience and Remote Sensing, 39:2490–2498. Liang, S., H. Fang, J. Morisette, M. Chen, C. Walthall, C. Daughtry, and C. Shuey, 2002. Atmospheric correction of Landsat ETM land surface imagery – Part II: Validation and applications, IEEE Transactions on Geoscience and Remote Sensing, 40(12): 2736–2746. Liang, S. 2004. Quantitative Remote Sensing of Land Surfaces, John Wiley and Sons, Inc., 534 p. Menzel, W.P., S.W. Seemann, J. Li, and L.E. Gumley, 2002. MODIS Atmospheric Profile Retrieval – Algorithm Theoretical Basis Document. Products: MOD07_L2, MOD08_D3, MOD08_E3, MOD08_M3, ATBD Reference Number: ATBD-MOD-07, URL: http://modis-atmos.gsfc.nasa.gov/_docs/atbd_mod07.pdf (last date accessed: 13 November 2006). Moran, M.S., R.D. Jackson, P.N. Slater, and P.M. Teillet, 1992. Evaluation of simplified procedures for retrieval of land surface reflectance factors from satellite sensor output, Remote Sensing of Environment, 41:169–184. Muinonen, E., and T. Tokola, 1990. An application of remote sensing for communal forest inventory, Proceedings from SNS/IUFRO Workshop in Umeå, 26–28 February, Remote Sensing Laboratory, Swedish University of Agricultural Sciences, Report 4: 35–42. Nilsson, M., and B. Ranneby, 1997. Estimation of Wood Volume Using Satellite Spectral Data – A simulation Study, Estimation of Forest Variables Using Satellite Image Data and Airborne Lidar (Ph.D. dissertation by M. Nilsson), Swedish University of Agricultural Sciences, Umeå. Nyyssönen, A., 1954. Metsikön Kuutiomäärän Arvioiminen Relaskoopin Avulla (English title: Estimation of Stand Volume by Means of the Relascope), Communicationes Instituti Forestalis Fenniae, 44:30. Oetter, D.R., W.B. Cohen, M. Berterretche, T.K. Maiersperger, and R.E. Kennedy, 2001. Land cover mapping in an agricultural setting using multiseasonal Thematic Mapper data, Remote Sensing of Environment, 76:139–155. Olsson, H., 1993. Regression functions for multitemporal relative calibration of Thematic Mapper data over boreal forest, Remote Sensing of Environment, 46:89–102. Olsson, H., 1995. Reflectance calibration of Thematic Mapper data for forest change detection, International Journal of Remote Sensing, 16:81–96. Ouaidrari, H., and E.F. Vermote, 1999. Operational atmospheric correction of Landsat data, Remote Sensing of Environment, 70:4–15. Poso, S., T. Häme, and R. Paananen, 1984. A method for estimating the stand characteristics of a forest compartment using satellite imagery, Silva Fennica, 18:261–292. Poso, S., R. Paananen, and M. Similä, 1987. Forest inventory by compartments using satellite imagery, Silva Fennica, 21:69–94. Rahman, H., and G. Dedieu, 1994. SMAC: A simplified method for the atmospheric correction of satellite measurements in the

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

solar spectrum, International Journal of Remote Sensing, 15: 123–143. Ripple, W.J., S. Wang, D.L. Isaacson, and D.P. Paine, 1991. A preliminary comparison of Landsat Thematic Mapper and SPOT-1 HRV multispectral data for estimating coniferous forest volume, International Journal of Remote Sensing, 12:1971–1977. Rosenfeld, G.H., and K. Fitzpatrick-Lins, 1986. A coefficient of agreement as a measure of thematic classification accuracy, Photogrammetric Engineering & Remote Sensing, 52:223–227. Song, C., C.E. Woodcock, K.C. Seto, M.P. Lenney, and S.A. Macomber, 2001. Classification and change detection using Landsat TM data: When and how to correct atmospheric effects?, Remote Sensing of Environment, 75:230–244. Song, C., C.E. Woodcock, and X. Li, 2002. The spectral/temporal manifestation of forest succession in optical imagery: The potential of multitemporal imagery, Remote Sensing of Environment, 82:286–303. Song, C., and C.E. Woodcock, 2003. Monitoring forest succession with multitemporal Landsat images: Factors of uncertainty, IEEE Transactions on Geoscience and Remote Sensing, 41:2557–2567. Teillet, P.M., and G. Fedosejevs, 1995. On the dark target approach to atmospheric correction of remotely sensed data, Canadian Journal of Remote Sensing, 21:373–387. Terry, R.A. 1987. Generating kappa statistics and testing useful hypothesis with PROC CATMOD, Proceedings of the Twelfth Annual SAS Users Group International Conference, 08–11 February, Dallas, Texas, SAS Technical Note TS-188, p. 1149. Tokola, T., J. Pitkänen, S. Partinen, and E. Muinonen, 1996. Point accuracy of a non-parametric method in estimation of forest characteristics with different satellite materials, International Journal of Remote Sensing, 17:2333–2351. Tokola, T., and J. Heikkilä, 1997. A priori site quality information in satellite image based forest inventory, Silva Fennica, 31(1): 67–78. Tokola, T., S. Löfman, and A. Erkkilä, 1999. Relative calibration of multitemporal Landsat data for forest cover change detection, Remote Sensing of Environment, 68(1):1–11. Tokola, T., 2000. The influence of field sample data location on growing stock volume estimation in Landsat TM-based forest inventory in Eastern Finland, Remote Sensing of Environment, 73:422–431. Tomppo, E., 1987. Stand delineation and estimation of stand variates by means of satellite images, Remote Sensing-Aided Forest Inventory, University of Helsinki, Department of Forest Mensuration and Management, Research Notes No. 19:60–76. Tomppo, E., 1992. Satellite image aided forest site fertility estimation for forest income taxation, Acta Forestalia Fennica, 229, 70 p. Tomppo, E., 1993. Multi-source national forest inventory of Finland, Proceedings of Ilvessalo Symposium on National Forest Inventories, Finland, 17–21 August, The Finnish Forest Research Institute, Research Papers, 444:52–60. Trotter, C.M., J.R. Dymond, and C. Goulding, 1997. Estimation of timber volume in a coniferous plantation forest using Landsat TM, International Journal of Remote Sensing, 18:2209–2223. Vermote, E., D. Tanré, J.L. Deuzé, M. Herman, and J.J. Morcrette, 1997. Second simulation of the satellite signal in solar spectrum: An overview, IEEE Transactions on Geoscience and Remote Sensing, 35:675–686. Wen, G., S. Tsay, R.F. Cahalan, and L. Oreopoulos, 1999. Path radiance technique for retrieving aerosol optical thickness over land, Journal of Geophysical Research, 104(D24):31321–31332. Zellner, A., 1962. An efficient method of estimating seemingly unrelated regressions and tests for aggregation bias, Journal of the American Statistical Association, 57:348–368. Zhang, Y., B. Guindon, and J. Cihlar, 2002. An image transform to characterize and compensate for spatial variations in thin cloud contamination of Landsat images, Remote Sensing of Environment, 82:459–479. (Received 26 April 2005; accepted 25 August 2005; revised 19 October 2005)

February 2007

163