Summary of originally proposed research

Changes in forest ecosystems in response to natural and anthropogenic influences and the resulting feedbacks to climate are of major concern to scientists studying global change.Northeastern China accounts for about 30% of the forest area in China, where has been undergoing dramatic changes during the last several decades due to forest fires, insect infestations, massive logging, agricultural conversion, and afforestation. In this research, forests parameters (such as biomass) are expected to be extracted using ALOS data based on Radiative Transfer models. The relationships between forest physical parameters and L-band radar backscatter will be investigated using model simulation and radar data analysis. And the research will try to test and evaluate the combined inversion technology using ALOS PALSAR date and other satellite data (especially lidar data).

Keywords: ALOS PALSAR, biomass, radiative transfer model, lidar.

1. RESEARCH OBJECTIVES

The main objective of this research is to develop algorithms for forest physical parameters estimation, such as tree height, biomass, species, using ALOS data. Two test sites in Northeastern China, where extensive field data and inventory data are available, were selected for this study. The detailed objectives include as following.

1)　　　 The elevation and vegetation height from two US Missions (SRTM and ICESAT) will be compared with the In-SAR derived surface elevation.

2)　　　 Investigate the relationships between forest physical parameters (such as tree height profile, biomass, species, etc.) and L-band radar backscatter by using extensive model simulation and radar data analysis.

3)　　　 Investigate the possibility of deriving tree height using ALOS PALSAR In-SAR data.

4)　　　 Investigate the data fusion between ALOS data and other satellite data (especially lidar data), and develop algorithms for forest biomass extractionby combined use of PALSAR data and Lidar.

In phase 1, the objectives include forest classification in research region; PALSAR In-SAR data processing; PALSAR and GLAS data fusion for biomass inversion.

2. RESEARCH PLAN

2.1. The original research plan

1)　　　 Data processing of PRISM, AVNIR-2, PALSAR data;

2)　　　 Develop software to derive surface height data from PRISM data;

3)　　　 Multi-sensor remote sensing data fusion;

4)　　　 Forest classification.

2.2. The actual implementation

Because there are not appropriate PRISM and AVNIR-2 archive data in our research area, we only use PALSAR data in the first research stage and the research content related to PRISM and AVNIR-2 has to be adjusted. Now, the actual research contents are as followed.

1)　　　 Forest classification in research region using TM images;

2)　　　 PALSAR In-SAR data processing.

3. RESEARCH IMPLEMENTATION

3.1. Forest classification

The classification process generally was separated into three steps: firstly, image pre-processing, such as Radiometric Correction, Geometric Correction and Clouds/Shadows Mask were performed. Since topographic factors (slope and aspect) causes a high variation in the reflectance response for similar vegetation types: shaded areas show less than expected reflectance, whereas in sunny areas, the effect is the opposite. So the process of terrain correction for the reflectance signal is critical in the rough mountainous regions. Then, a supervised method was applied to the preprocessed TM data.

1) Terrain correction

In this research, we calculated the illumination image, firstly, according to the geometric conditions of solar irradiation and land surface Slope and Aspect. The cosine value of local incident angle was calculated pixel by pixel, considered as the illumination status. Because the Dem data is from SRTM 90m spatial resolution, the spatial scale is changed to correspond with 28.5m LandSat image. Spatial resolution 90m illumination image data was co-registered to the landSat image, with 10-20 ground control points(GCP), and re-sampled to 28.5m per pixel, by bilinear method. The RMSE of the polynomial function should be less then 0.5 pixel. According to our local experiences only pixel records with 0-3000m and slope less than 60 degrees are reserved as valid values and a Flag mask was built to filter the invalid ones in illumination image. The topographic correction model (Formula 1-3) was used to modify the spectral signals to flat-normalized ones in mountainous regions. In out study, statistically-experienced parameter Ck is determined by training samples in primary evergreen needle forest region with different illumination status. Figure1 shows the flow chart of terrain correction.

 (1)

　　 （2）

　　　　  (3)

where is the solar zenith angle;  is the local incident angle, is the reflectance of a horizontal surface;  is the reflectance of an inclined surface. Where , is based on the empirical–statistical method of Teillet that assumes a linear correlation between the reflectance of each band and .  is the slope of the regression line for band k.  is the spectral signals in the image and considered constant for the entire image, being the intercept in the regression equation.

2) Data processing and mapping

TM image is co-registered and re-sampled to spatial resolution 28.5m, and re-projected to UTM projection. And Calibration is needed to convert DN value to Top of atmosphere reflectance, then spectral indices (NDVI and LSWI) is calculated to represent the slope-shape of the spectral curve. Then, all the spectral bands and VIs were stacked together. A cloud/shadow mask was also created by visual interpretation. PCA transform was applied to this data set to make us concentrate on several principle components for training samples analysis. Thematic legend is composed of 12 class types in our research area. For these area all locate in mountainous region, no Crop land is identified. Figure2 is the classification result of Daxing'anling forest area. Here, two LandSat data sets are collected: TM data(path/row: 122/23) observed on 5 July, 2005 with clear sky (cloud<5%) is selected to classify, ETM+ data observed on May 15,2002 is used as auxiliary data and mapped too, to fill in the cloud/shadow masked region.

Figure1 Flow chart of terrain correction

Figure2 Classification results of DXAL

3.2. PALSAR In-SAR data processing

We have ordered 20 PALSAR 1.1 data in our research areas (Daxinanling and Changbai Mountain) since Oct. 2007. Then we use ROI_PAC program to process these data and to obtain the DEM data and coherence map. Figure3 shows the processing flow of ROI_PAC.

Figure3 The processing flow of ROI_PAC

Using the orbit information given in assistant data files, the baseline is reckoned. The master data (figure4a) obtained Aug. 2007 and the slave data (figure4b) obtained July, 2007 are matched according to the baseline information, and then the interferogram and coherence are generated, which are show in Fig5.

Figure4 The HH polarization of PALSAR  level 1.1 data (a)master data (b)slave data

Fig 5 Interferogram and coherence. (a)Interferogram (b) coherence (c)TM image (d) distribution of coherent coefficients

In Fig5 (b), the red triangle is the LUSHUIHE forest bureau, which is one of our test areas. Around the town is mature red pine, which shows higher coherence because of its stable. Fig5 (c) is the TM image of this area. We can see that clear cut areas and bare ground have higher coherence while rivers and young forest show lower coherence. Coherent coefficient is also an index of the quantity of the InSAR Process. Fig5 (d) shows the distribution of coherent coefficients.

In order to get topography information, the part of phase difference generated by flat ground should be removed. Fig 6(a) is the phase difference only due to topography. In order to improve the unwrap quantity; the interferogram was filtered by adaptive method with strength 0.75, as shown in Fig 6(b). Sometimes the flat ground components in interferogram can not be removed thoroughly, which can be checked by comparing the result and module of SRTM-DEM by 150 meters. This certain value can be obtained from the formula 4.

 　　　　　 (4)

It is known that interferogram is  modulo 2π. Here are related to certain InSAR system, which can be obtained from assistant files and baseline file. Fig 6(c) is SRTM-DEM modulo the value. There are some deformations in Fig 6(b) for it is still in slant range coordinate while Fig 6(c) is geo-coded.

Fig 6. (a) The interferogram after taking off the part induced by flat ground. (b) The Interferogram filtered by adaptive method with strength 0.75 (c) the SRTM-DEM modulo the value get from the formula

Then, the interferogram can be unwrapped according to the baseline attained from orbit data, which is shown in Fig 7(a). Through Radar simulation using given DEM, the baseline was re-estimated and the interferogram was unwrapped again, Fig 7(b).

Fig 7 (a) Interferogram unwrapped with orbit data (b) Interferogram unwrapped with re-estimated baseline

3.3. Forest biomass inversion using PALSAR data

A forest radar backscattering database was build through the combination of a forest growth model and the forest radar backscattering model. The database was composed by 175500 records, including three types of forest, nine types of soil, 13 incidence angles and 500 forest stands. The relationship between forest biomass and simulated radar backscattering coefficients in the database was shown in Fig 7.

Fig7. The relationship between forest biomass and simulated radar backscattering coefficients

In order to validate the LUT in the forest biomass retrieval, another 148500 records were simulated. Then the LUT method was used to retrieve forest biomass. The results were shown in Fig 9. It is obvious that the LUT is useful. Then it was used in the biomass retrieval from PALSAR data, which is registered to forest classification map shown in Fig8. Results were compared with forest biomass from forest map as shown in Fig10. We can see that the standard deviation is about 66 ton/ha although the mean value of difference is only 5.5 ton/ha. In order to make further validation, the retrieval results were checked by point survey data as shown in Fig11. We can see that the results are bad. Multiple resolutions are the major problem in the LUT method. In the inversion process only HH and HV backscattering coefficients was used. The number of observation variables is two less. Therefore the problem of multiple resolutions can not be solved. More information was need. Existing literatures showed that InSAR phase center height was highly correlated with tree height and the biomass can be calculated by allometric equations for given forest species. Therefore the relationship between InSAR data (Coherence and difference of phase center height between C and L band) was investigated as shown in Fig12.

Fig8. The composite PALSAR and forest class map for Lushuihe study area. Red:HH, Green:HV, Blue:forest class map

Figure 9. The inversion results of LUT method using simulation data.

(a) 

(b)

Figure10 . Retrieved biomass by LUT and PALSAR data. (a) Biomass map. (b) the distribution of difference between the retrieved biomass and biomass from forest map

Figure 11. Validation of retrieved biomass by point survey data

(a) 

(b) 

Figure 12. The relationship between tree height and InSAR data. (a) Coherence (b) difference of phase center height between L and C band

　 For mature forest (height>10m), both the Coherence and difference of phase center height were positively correlated with forest height. For young forest (height <10 m) Coherence is negatively correlated with tree height. No correlation appears between difference of phase center height and tree height. This is mainly due to the data acquisition time. The information of phase center height at C band was derived from SRTM which was observed on the year 2000 while the PASAR data was taken on 2007. In the past 7 years, young trees grow taller. Therefore the correlation between difference of phase center height and tree height was weak or vanished.

4. PROBLEMS FOR RESEARCH IMPLEMENTATION

In In-SAR DEM research, there is a problem in 3D module of ROI_PAC program and the DEM cannot be generated correctly, which restricts the terrain correction. We consider to using high resolution topographic data to do this work instead of In-SAR DEM.

Through the above description, we can see that InSAR data was correlated with tree height. Therefore InSAR data can be used as additional information in the biomass retrieval. However, for young forest, the relationship between difference of phase center height and tree height was broken by the inconsistent of data acquisition date.

In the future research we will try to generate digital surface model by PRISM data to take place SRTM in the calculation of tree height.

List of data used in the research

Satellites data and in situ data used in the research are listed in table 1.

Table 1. Sample table