Estimating regional forest cover in East Texas using Enhanced Thematic Mapper (ETM+) data

https://doi.org/10.1016/j.foreco.2005.08.033Get rights and content

Abstract

The USDA Forest Service, through its Forest Inventory and Analysis (FIA) program, periodically estimates forest/non-forest area at the county level using aerial photographs. Satellite-based remotely sensed data and digital image processing techniques could substantially reduce the time required to generate this information. Satellites collect data on a repeat basis and with higher frequency than the aerial photos that are currently used for this purpose. In addition to the forest cover estimates, the USDA could use satellite data to generate maps depicting the spatial distribution of forest cover. However, few studies have tested the utility of medium-resolution satellite data for FIA purposes. We tested the potential for using LANDSAT satellite data to obtain forest cover estimates for a six-county region in East Texas. Satellite data were processed using a combination of image classification techniques that could be repeated in other regions of the USA. Results were compared with the results of traditional photo-based estimation techniques and were comparable within a 95% confidence interval. Based on this study we recommend that medium-resolution satellite data can be used for obtaining county-level forest cover estimates.

Introduction

The US Department of Agriculture-Forest Service (USFS) has periodically estimated and published the extent of forest cover and timber resources in the United States as part of its Forest Inventory and Analysis (FIA) program (USFS, 1992, Frayer and Furnival, 1999, Reams and van Deusen, 1999). This information, published since the 1930s, is used by state forest agencies, private timber companies and individual foresters for planning and decision making. In addition to this, FIA results are used for assessing sustainability of forest management practices and predicting the effects of global change (USFS, 2004). The FIA program uses a variation of the double sampling method for collecting data about forest resources. In the first phase, points are placed on aerial photographs and are classified as either forest or non-forest. In the second phase, detailed information about forests is collected by visiting a predefined number of photo-points on the ground. The estimates obtained in the first phase are refined based on the ground information and standard errors are computed (McWilliams and Bertelson, 1986, Kelly et al., 1992, Hansen and Wendt, 1999, Reams and van Deusen, 1999). This information is used for periodically publishing statistical estimates of forest cover at the county level (Wynne et al., 2000, McRoberts et al., 2002).

Wayman et al. (2001) and McRoberts et al. (2002) summarize the limitations of aerial photographs for FIA purposes. Interpretation of the photos is a labor-intensive and time-consuming process. Photographs are expensive and cumbersome to handle, store and transfer. Also, obtaining current aerial photographs for FIA purposes is often difficult. Reams and van Deusen (1999) identified the inability to produce maps using county-level estimates. County-level estimates cannot be used to capture the spatial variability of forest cover within each county. In addition, it takes approximately 8 years for the FIA program to update estimates for the southern region (13 states, Puerto Rico and the Virgin Islands) of the US.

Satellite-based remotely sensed data in combination with semi-automated digital processing could reduce the time required to generate forest and non-forest estimates (Lannom et al., 1995, Cooke, 1999, Czaplewski, 1999, Wayman et al., 2001). Orbiting satellites collect data more frequently and regularly on a global basis than do aerial photography programs. Currently, satellites from the US (LANDSAT-http://landsat.usgs.gov), France (SPOT-http://www.spotimage.fr) and India (IRS-http://www.nrsa.gov.in, http://www.antrix.org) provide medium-resolution multi-spectral data. These satellites collect information in the green, red and infrared regions of the electromagnetic spectrum that is useful for discriminating vegetation. The current LANDSAT satellite (ETM+) developed an anomaly (Scan Line Corrector malfunction) in May 2003 that reduces its utility, but data from an earlier LANDSAT satellite (LANDSAT 5) are nearly identical in character and are still available for use. Plans are underway to include the next LANDSAT sensor in a NOAA satellite scheduled for launch in 2009. Should LANDSAT TM5 fail before 2009, data collected by the IRS and SPOT satellites could be used for forest estimation using methods similar to those described here.

In addition to forest and non-forest estimates, information about forests could also be produced from satellite data in a variety of formats including maps (Dymond et al., 2002). Among other applications, these maps could be produced at regular time intervals and would improve the spatial accuracy and precision of forest cover estimates, provide spatially explicit estimates of changes in forest cover and condition, fuel availability and wildlife habitat among others (Beaubien, 1994, Wayman et al., 2001).

The 1998 Farm bill recommended that the USFS and NASA work together to integrate satellite-based remotely sensed data for the forest inventory program. This bill also mandated that the USFS sample 20% of the plots in a state every year, a substantial increase in sampling density (Wayman et al., 2001). In addition, the FIA was one of several federal government programs reviewed by a study commissioned by the Office of Science and Technology Program (OSTP). One of the recommendations made by this study was to incorporate satellite data in general and LANDSAT Thematic Mapper data in particular into this process to reduce the dependency on aerial photographs for FIA purposes (Peterson et al., 1999).

In order for the USFS to incorporate satellite-based estimates into its FIA program, additional research is required to address the following issues: transferability of image processing and classification methods for other regions, sources of misclassification related to landscape pattern, and precision of the estimated area for each thematic class that incorporates uncertainty. Only comprehensive studies will enable the USFS to evaluate the usefulness of satellite-based estimates in comparison to traditional photo-based estimates. This paper describes such a study where the utility of LANDSAT data to map and estimate forest resources was tested in East Texas.

Satellite image processing and classification of forest resources involves assigning the pixels in the image to predefined forest types. Methods such as unsupervised or supervised classification or a combination of these two are available for grouping pixels into forest or non-forest classes (Lillesand and Kiefer, 2000, Jensen, 2000). Numerous advances have recently been made in image classification algorithms such as fuzzy logic (Liu and Samal, 2002) and neural networks to derive information from satellite images. These advances coupled with developments in computing, have significantly increased the amount of data that can be processed in a given time and the quality of the results.

After classifying a satellite image, an analyst assesses how accurately the image was classified by using verification data, which is usually collected in the field or from high resolution aerial photographs. Classification accuracy is typically reported in an error matrix (Congalton, 1991) consisting of an equal number of rows and columns representing the number of classes, with mapped types on one axis of the table and reference classes on the other. If most of the elements of this matrix fall along the diagonal, there is relatively high agreement between mapped types and ground reference data. Deviations from the diagonal indicate mismatches between the two. The kappa agreement index can also be computed for the error matrix and indicates the level of agreement between the classified image and verification data (Congalton, 1991).

A review of several published studies by Holmgren and Thuresson (1998) found that most studies used randomly distributed verification sites to assess the accuracy of classified images. The number of verification sites used depends on the variance in accuracy among the mapped sites and the statistical precision required. This review also found that in several studies the overall accuracy of the classified image was inflated because the analyst included relatively more verification data corresponding to easily identifiable features in a satellite image. One of the recommendations of this review is that future studies should use some form of systematic sampling to select verification data, and that the total number of verification sites must be based on statistical principles of sampling. Systematic sampling procedures will minimize the bias in the number of data points assigned to different classes.

Inadequate information about the precision of area estimates derived from maps limits the user's ability to understand the uncertainty associated with these estimates. Classified images have errors that can be expressed as overall or individual class accuracy. However, area estimates obtained from any classified image are often reported as a single number, such as 3600 ha of coniferous forests (Wynne et al., 2000). Card (1982) developed a method for incorporating classification errors into area estimates obtained from satellite images. However, this method has been incorporated in few studies (Wynne et al., 2000, Wayman et al., 2001).

Several studies in the US assessed the utility of satellite imagery such as from the LANDSAT multi-spectral scanner (MSS) (Dodge and Bryant, 1976, Fox et al., 1983, Moore and Bauer, 1990), Thematic Mapper (Moore and Bauer, 1990, Bauer et al., 1994, Wayman et al., 2001), and the AVHRR (Iverson et al., 1989, Nelson, 1989, Teuber, 1990, Zhu and Evans, 1992, Zhu and Evans, 1994) for mapping forest cover, and in certain instances for obtaining FIA estimates (Teuber, 1990, Zhu and Evans, 1992, Zhu and Evans, 1994, Hansen and Wendt, 1999, Franco-Lopez et al., 2001). Attempts have been made to use products developed from other projects such as the Gap Analysis Program (Hansen and Wendt, 1999) or the National Land Cover Dataset (McRoberts et al., 2002), to estimate forest cover. However, most of these studies have not addressed all of the issues related to systematic sampling and precision.

The primary objective of this study was to develop a methodology using LAND SAT ETM+ imagery to obtain forest and non-forest estimates comparable to those obtained from aerial photos for East Texas. A secondary objective of this study was to generate maps of forest and non-forest classes with sufficient thematic accuracy to be useful for further stratification and analyses. It is hypothesized that “error-corrected” area estimates obtained from satellite imagery would not be different from the estimates obtained from aerial photographs at the 95% confidence level and that the sources of error in the classified images can be attributed to a limited number of land cover or land use classes. If this hypothesis is supported, LANDSAT imagery may be a cost effective and robust alternative to air photo-interpretation for making FIA estimates.

Section snippets

Study area

Angelina, Nacogdoches, Panola, Rusk, San Augustine, and Shelby Counties in East Texas (31°43′N, 94°24′W) were chosen for this study (Fig. 1). This region receives an average of 119.2 cm of rainfall every year but precipitation varies on an average monthly basis from 5.5 cm in July and 11.64 cm in May. Average annual minimum and maximum temperature vary between 12.8 °C and 25.5 °C. Average summer maximum can reach 35 °C.

The USFS estimated that about 8.4 million hectares (67.5%) of the total area (12.44 

Overall accuracy and class agreement assessment

The overall accuracy for all the counties, when the image was compared to photo-point data, was 85% for the study area and varied from 78% for Rusk County to 96% for San Augustine County (Table 2). Kappa values (Table 2) for the study area (0.67 or 67%) were lower than the overall accuracy, because they incorporated off-diagonal elements of the error matrix, thus providing a more comprehensive view of agreement than the overall accuracy measure. San Augustine County had the highest kappa

Image classification

This study used satellite-based unsupervised classification techniques and successfully matched forest area estimates obtained from more labor intensive traditional photo-interpretation. This suggests that satellite data may be an economical alternative for the USFS FIA process. Other researchers have obtained comparable results, but with more complex iterative classification methodologies (Holmgren and Thuresson, 1998, Wayman et al., 2001). The method described in this paper is simple, robust

Conclusions

Satellite data can provide FIA phase I estimates of forest area which are comparable in precision to those obtained using the traditional photo-estimation method. The method described in this paper allows routine classification of satellite images with minimal training required for USFS personnel. As a result, the USFS can obtain and publish area estimates faster than is possible using traditional photo-interpretation because the time required for classifying a LANDSAT or similar satellite

Acknowledgements

This study was conducted as part of the first author's doctoral dissertation research and was financially supported by the Texas Agricultural Experiment Station and the Heep Foundation. Support provided by the Texas Forest Service personnel, Mr. Thomas Spencer and Mr. Curt Stripling in particular, for conducting field work and aerial photo-interpretation are gratefully acknowledged. The authors thank Texas view for providing LANDSAT images used in this study. Contributions of Dr. William Cooke

References (45)

  • R.G. Congalton et al.

    Assessing landsat classification accuracy using discrete multivariate analysis statistical techniques

    Photogramm. Eng. Remote Sens.

    (1983)
  • R.G. Congalton et al.

    A quantitative method to test for consistency and correctness in photo interpretation

    Photogramm. Eng. Remote Sens.

    (1983)
  • Cooke, W.H., 1999. Forest/non-forest stratification in Georgia with LANDSAT TM data. In: Proceedings of the First...
  • R.L. Czaplewski

    Multistage remote sensing: toward an annual national inventory

    J. For.

    (1999)
  • A.G. Dodge et al.

    Forest type mapping with satellite data

    J. For.

    (1976)
  • Field Guide

    (1996)
  • K. Fitzpatrick-Lins

    Comparison of sampling procedures and data analysis for a land-use and land-cover map

    Photogramm. Eng. Remote Sens.

    (1981)
  • L. Fox et al.

    Classification of forest resources with LANDSAT data

    J. For.

    (1983)
  • W.E. Frayer et al.

    Forest survey sampling designs: a history

    J. For.

    (1999)
  • Hansen, M.H., Wendt, D.G., 1999. Using classified LANDSAT thematic mapper data for stratification in a statewide forest...
  • P. Holmgren et al.

    Satellite remote sensing for forestry planning—a review

    Scand. J. For. Res.

    (1998)
  • W.D. Hudson et al.

    Correct formulation of the kappa coefficient of agreement

    Photogramm. Eng. Remote Sens.

    (1987)
  • Cited by (0)

    View full text