ES 5053 Remote Sensing Final ProjectNeal Simpson

Title

Estimating forest structure in wetlands using multi-temporal SAR

Journal

Remote Sensing of the Environment 79 (2002) 288-304

Author

Philip A. Townsend

Focus Categories

Multi-temporal radar, biophysical characteristics, forested wetlands, basal area, canopy height

Abstract

Data from 202 forested plots on the Roanoke River floodplain in North Carolina were used to asses the capabilities of multi-temporal radar imagery for estimating biophysical characteristics of forested wetlands. The research was designed to determine the potential for using data from current data from satellite SAR sensors to study forests over a wide geographic areas and different environmental backgrounds. Data from Radarsat, ERS-1, and JERS-1 was used. Sites were compared on common flood status areas (flooded and non-flooded). The SAR imagery was compared to Landsat TM imagery to determine forest biophysical properties, and if it was comparable for reliable results. Statistical analysis including regression models were run to determine if the possibility of the multi-temporal, multi-sensor data would be comparable to field tests and optical imagery types collected. If this analysis is proven to be accurate, then this could be used for future research for biophysical forest properties over vast geographic regions with complex environmental gradients.

Introduction

Broad-scale estimation of the bio-physical properties of forests is a hot topic in remote sensing. This type of research provides important information on global change and scientific basis for regional scale forest assessment. Use of synthetic aperture radar (SAR) has been interesting to detect forest properties. There are benefits and limitations to the use of SAR sensors. The benefits are non-attenuation in the atmosphere and SAR backscatter is responsive to multiple structural elements of forest canopies. Use of SAR has shown promising results in analyzing biomass (Dobson et al., 1992), basal area (Dobson et al., 1995), tree height (Dobson et al., 1995), tree diameter and density (Hyyppa et al, 1997), and forest cover class (Dobson et al., 1995). SAR limitations though fall into three categories: 1) Data that has been used in most studies is not widely available or geographically extensive. 2) Widely available data that has been used is produced from single-band, single-polarization imagery. This type of imagery has been demonstrated that multi-channel, cross polarized data exhibit the strongest relationships with forest bio-physical properties. 3) Variations in environmental conditions affect backscatter from forests. Some surface properties sometimes exert a stronger influence on backscatter than forest characteristics. This is especially dramatic in flooded forests, where inundated areas exhibit higher backscatter than non-flooded areas due to increase in double-bounce scattering. Effects of differences in soil and surface properties beneath non-flooded forests are more subtle and are therefore are difficult to address without detailed in-situ data. For broad scale analyses, detailed soil/surface properties are not typically available.

Objectives

Goal was to evaluate the capabilities of multi-temporal SAR imagery from Radarsat, ERS-1, and JERS-1 for estimating biophysical properties of forested wetlands in the Lower Roanoke River floodplain in North Carolina. They attempted to address several issues with using SAR data to study forest structure with the following:

1. How does the sensitivity of SAR imagery to forest biophysical properties differ for flooded and non-flooded forests?

2. Can multi-temporal SAR imagery be used to estimate forest biophysical characteristics accurately?

3. Does the integration of multi-spectral optical imagery with SAR data substantially improve the ability to detect forest properties?

4. What effect do other forest and surface properties, such as species composition and soil characteristics, have on radar backscatter from forested wetlands?

Materials and Methods

The study site for this project was the Roanoke River floodplain in North Carolina. The area consists of wide range of forested wetlands, with various species of hardwood trees and swamped forests from the North Carolina-Virginia Border down to the Bay leading out to the Atlantic Ocean. The species of trees changes drastically from higher elevation areas to flooded areas in the wetlands obviously due to the water content in the soils. 202, 90 x 90m test plot sites located along the Roanoke River had field data collected during 1995-6 time frame, including density, basal area, composition data using the Bitterlich variable plot method. In the 90m plots five points were sampled for basal areas of the various species. Also, canopy closure was assessed in almost all plots for all species. In 39 of the plots forest canopies were measured for top and bottom of canopy to determine canopy depth In 51 of the plots pictures were taken of the forest canopy using a fish eye lens to measure leaf area index, using a video digitizing system the leaf area index was calculated from the negatives. All 202 plots had their coordinates calculated by using differentially corrected GPS. A second data set was collected from 116 forest plots in which cover, density, and basal area were calculated, as well as the top 10cm of the soil underneath the litter of the trees. The soil samples were analyzed for organic matter content, and percentage of sand, silt, and clay. All field data was integrated with an image based vegetation classification of the study area.

SAR data of the field sites was acquired from ERS-1, Radarsat, and JERS-1 satellites. All images were slant-to-ground range corrected and geo-referenced to UTM coordinates. For all of the images, the 12.5m pixel size was used for the spatial resolution, but the actual resolution was closer to 30m after analysis. For all of the three satellites, imagery was collected at different times of the season, and different stages of flooding. With the different times of the season, different stages of flooding, and differences in phrenology, statistical analysis had to take these differences into account. Imagery was collected between 1993 and 1998, but none of the sampled plots had any changes (cutting or diseases), so they had valid use for the imagery and field data. Soil moisture data was not available for the non-flooded areas, which may have substantially influenced data in radar scattering.

Hydrologic state of each plot is stratified for each image acquisition date. For JERS-1 and Radarsat imagery forested areas were classified as flooded or non-flooded after speckle reduction was calculated. Flood classification was validated by wells located throughout the plot areas. For ERS-1 imagery flood classification was not used due to ERS-1 imagery inability to reliably to detect inundation beneath forest canopies, but was indirectly incorporated into flood status.

Landsat TM imagery of seven of the plots was also used to develop the forest cover classification, and compared to SAR data to determine the forest biophysical properties. Landsat imagery was collected from Mar – Aug of 1993 and represents phenological changes in the plots. The images were geo-referenced to UTM coordinates and have a 30m spatial resolution. The imagery had dark object corrections conducted and atmospheric scattering through haze removal. These images were used to determine the NDVI of bands 3 and 4.

Multivariate statistical analysis was used to test hypotheses that multi-temporal, multi-sensor SAR could be used to predict biophysical properties of the forests in the study areas. Multi-temporal backscatter measurements were used as independent variables, while the stand characteristics were use as the dependent variables in the statistical analysis. Forest properties used in the tests were basal area, percent cover, leaf area index, top and bottom height of canopy, and canopy depth. Major differences in radar scattering from these forests are due to differences in flood inundation, so the statistical analyses of the plot data were stratified by flood status. So, three categories of analysis were derived: 1) analysis based on plots that were all flooded on the same dates, 2) analysis based on non-flooded plots on the same dates, and 3) analysis based on plots that are flooded on the same dates and not flooded on others.

These three categories allowed complete coverage of the study area to attain the model forest biophysical attributes spatially. Basal area and cover were split randomly into 136 plots for model development, and 66 plots for model validation. For leaf area index and canopy heights, the stratification of analysis meant that there were not enough plots to split the sample into both test and validation sets, so cross validation was accomplished by jackknifing the data. Normality of forest structure was accomplished by the Shapiro-Wilk statistics. Collinearity was analyzed between forest properties and radar backscatter, and to examine relationships between soil properties and radar backscatter. Simple and multiple linear regressions were used to test the relationships between forest properties and radar backscatter. Ratios between the calibrated scenes were calculated as interaction variables in the models. Multiple linear regressions were used to predict forest stand characteristics as a function of the multi-temporal, multi-sensor data. Separate regression models were used for high basal area stands (BA > 55m2/ha), due to sensitivity of radar at higher biomass levels. Generalized linear models were used to test whether differences in radar backscatter could be attributed to differences in vegetation community composition for the Landsat TM imagery.

Results

There were two main things analyzed in this experiment, 1) relationships between forest structure and radar scattering, and 2) relationships between backscatter and other variables. The relationships between forest structure and radar scattering was broke down into smaller categories to determine which part of forest structures were able to be determined by using SAR sensors. Stratification of data between flooded imagery and non-flooded imagery was used to determine if there would be any difference between the two data sets. The data sets were stratified in flooded plots, non-flooded plots, and for all plots. There was apparent relationship among stratification of the data sets. Basal area showed strong correlations between backscatter and forest structure in all plots, but were stronger in the flooded plots due to the double-bounce backscatter. Tupelo-Cypress swamps showed the best correlations with basal area, because they are typically flooded swamps. Other forest variables appear strongest for plots classified as flooded, especially basal area and height, with the strongest correlations with backscatter. These can be explained by the presence of continuous flooding which minimizes backscatter differences, and the increased responsiveness of SAR to structural components in flooded forests from the proportionally higher level of trunk-ground and crown-ground interactions that occur as consequence of double-bounce scattering from a flooded surface.

Leaf-on versus leaf-off correlations were run to determine if seasonal changes could be used to determine forest structural properties by comparing them against the satellite backscatter. Basal area showed a strong correlation with forest structure properties (especially basal area) for both leaf-on and leaf-off conditions, which suggests differences between scattering during seasons could be used for predicting forest structures by using SAR sensors.

They also looked at what affect the wavelength, polarization, and incidence angle may have in determining forest structure properties. Incidence angle had only slight differences to detect forest properties, although the smallest incidence angle exhibited the lowest sensitivity, while the steepest incidence angle had the strongest sensitivity. Both C and L bands and CVV and CHH polarizations appeared to be useful in detecting forest properties because backscatter was responsive to the forest variables. Since all of these sensors exhibited association with forest properties indicates potential for using multitemporal, multisensor approach to estimate forest structure. Cost for these images is the problem and may be cost prohibitive to do this type of analysis.

How does multitemporal, multisensor SAR compare in estimating forest structure properties. Leaf area index, canopy depth, and crown closure was determined to be better analyzed with optical imagery (Landsat TM), so analysis with SAR was not conducted for these. Basal area was the strongest of correlations, especially in flooded forest plots. The highest correlation model was for all plots flooded and non-flooded, which indicates both plot areas can be used to determine basal area for forests. Leaf on images showed correlations for a few of the plots with basal area, showing the C-band could determine canopy foliage. The ERS-1 CVV image though showed it was useful in the early spring images, but not in the late summer images from attenuation of foliage. The JERS-1 L-band penetrated the canopy and could estimate the sub-canopy forest structure, especially basal area in high basal area plots. This shows the longer wavelength SAR for measuring structure under high biomass areas. The ground truth basal area was run against SAR determined an error of about 5m2/ha, suggesting good results but needs improvement.

Forest height was also responsive for the SAR sensors, especially in flooded areas. For non-flooded forests summer images were the best results, while for flooded areas spring and fall images were the best correlated. Non-flooded forests only were useful in determining height to the bottom of the canopy, which was unexpected.

Once the SAR imagery was analyzed, it was wondered if integration of optical imagery could improve the SAR only models. NDVI was used because it’s strong relationship with vegetation properties. For basal area, the early spring images showed the greatest improvement for the models, due to the small leaves on the trees which are not incorporated into the basal area. Neither leaf-on nor leaf-off scenes improved the model substantially. An improvement to the overall model was less than 0.1, therefore showing integration of optical imagery did not help with analysis of high basal areas. It was not expected that the Landsat TM data would improve the overall model for tree height variables, although surprisingly some of the optical imagery slightly improved canopy height for non-flooded areas. So, although improvements of the overall models for basal area and tree height were slight, the use of optical imagery as supplemental data can be helpful.

Once the forest structure properties were analyzed in the SAR images, what was the effect on backscatter against other variables? Land cover type was the first variable that was examined. Forest biophysical properties were not stratified by forest types since most of the sites were strongly dominated by deciduous forests. Tree types may produce important differences in backscatter between the different types of forests. Vegetation classification was determined using NDVI into 20 different classes to see if backscatter would differ between the forest class types. The only classes which backscatter differed were expected to be flooded on each date, and those that were not flooded, which shows the need for stratification of flooded and non-flooded dates. Evenwhen classes were stratified the correlations were weak, except for Tupelo-Cypress swamps. The Tupelo-Cypress swamps have distinctly different backscatter measurements to the other forest types, probably due to the high basal area. Analysis showed that the reason for the difference in Tupelo-Cypress swamps was from the basal area and not from the backscatter. Differences between forest types are more closely related to the average environmental conditions of the type rather than differences in composition.

Modeling studies have indicated that environmental factors other than flood inundation and vegetation structure substantially affect backscatter response from forests.So, they also analyzed the soil properties of the sample sites to determine what affect they would have on backscatter of the SAR sensors. Previous studies have shown the sensitivity of SAR to surface parameters is most pronounced for co-polarized data, especially for L-band backscatter. No a lot of surface data was available for the soils at the sample sites, except for soil texture and organic matter content in the soils. Only non-flooded sites were used in this analysis because the flooded plots were thought to have little interactions between the soil properties and backscatter. Soil moisture is highly correlated with soil texture and organic matter content, so correlations were ran against the soil properties and backscatter. Few correlations were found to be significant between the soil properties and backscatter. The most frequent correlations were found between backscatter and organic matter and clay content, due to the ability of these to collect water strongly. The correlations were weak overall, but it shows that not only empirical studies control the backscatter response of SAR sensors in forests. Field tests would be important to cross validate the results of the SAR backscatter against physical properties of the soils.