> Research

Comparing Integrated LIDAR and Multispectral Image Data with Field Measurements in Bottomland Hardwood Stands


Forest resource management has changed dramatically in the past 10 years, and in no discipline is this as evident as it is in remote sensing. Traditional photogrammetric techniques such as area delineation and stereoscopic measurements have given way to advanced spectral classification and spatial measurement of forest health and structure. As the toolbox continues to grow, another technology has entered the picture which, when combined with already proven techniques, has potential to improve efficiency with respect to fieldwork, time, and overall costs. This technology is LIDAR, and through cooperation with the USDA Forest Service, this project is exploring its use, with other complimentary technologies, for assessment of Southern bottomland hardwoods.

The specific LIDAR (Light Detection and Ranging) configuration used in this project provided multi-return data with a relatively small footprint (~3.14 ft2 or 0.3 m2). The combined technology involved is a four camera multispectral (MS) system that yielded a spatial resolution of approximately 0.3 meters (~1 ft). LIDAR provides valuable spatial accuracy, particularly in the height domain, while the MS data contributes the spectral information needed for species group classification.

Project Objectives

While the end result is hoped to be improved estimation of timber volumes from these types of remotely sensed data, this project does not directly embrace volume estimation. Instead, an exploration is made into the first steps in making volume prediction; estimation of forest and individual tree parameters. The particular parameters of interest here include tree height, crown area, density, and species group classification, characteristics that are also valuable in other environmental assessments.


The project is following three major stages aside from sensor data collection, which was contracted out. These steps involve the collection of field data, the extraction of information from the remotely sensed data, and the testing and compilation of results.

Field data were collected in two hardwood bottoms in East Mississippi with circular fifth acre plots situated in a systematic pattern with at least two dominant and/or codominant trees located in different quadrants (dictated by cardinal azimuth degrees) in each plot. If a plot did not meet this criterion, an alternate was chosen in the same relative position as the original with respect to the distance from LIDAR swath centerline. This allows for future testing of parameter estimation as it varies moving away from LIDAR swath center. The actual field data included; tree species, DBH (Diameter at Breast Height), total height, merchantable height, height to foliar live crown, height to branched live crown, and crown classification (i.e., dominant, codominant, intermediate, and suppressed). In addition to these records, a dominant or codominant tree was chosen from at least two different quadrants and their crown radii in eight directions (45 degree increments starting at north) were measured along with direction and distance of the boles from plot center. These measures, in conjunction with the location of the plot centers that were established with real-time DGPS (Differential Global Positioning System), allowed the construction of vector data layers of tree crowns (Figure 1) and plots in a GIS (Geographic Information System) environment. In this format, crown spatial features could be cross-referenced with LIDAR and MS data.

LIDAR data were surfaced into two datasets. The first was the crown layer derived by from the first return points interpolated to a raster surface (Figure 1). A DTM (Digital Terrain Model) was then generated from point output using the contracted LIDAR provider’s proprietary ground point extraction software (Figure 2).

The MS imagery required some reflective normalization across individual frames as some illumination inconsistencies, mainly caused by bidirectional reflectance, plagued the initial attempts to perform image classification. The problem encountered in “correcting” this phenomenon was lack of information that is considered key in adjusting for such variation. An empirical approach was designed and implemented which only required an average reflectance value per band over several images and the x and y coordinates of these mean reflective values away from image ”hotspot” (point where sensor-to-target-to-sun angle equals zero). The results greatly increased the visual quality of the images (Figures 3 and 4) and made the classification of them on a mosaic level homogeneous with respect to within-frame general reflectance.

After minimizing within-frame illumination discrepancies, the imagery was orthorectified to the first return (canopy) LIDAR surface. The images from the study area were then mosaicked and loaded, along with a vegetation height raster dataset (defined by subtracting the first return surface from the DTM surface) into an object-oriented image classification software package to isolate individual tree crowns and classify them by species groups. The classification result will be a vector dataset with species or species group classification semantics, polygon/crown area, and the maximum value from the vegetation height LIDAR surface attributed to each polygon (representing crowns). These attributed estimates, along with a simple crowns-per-area (density) calculation, will then be tested against field measurements to achieve the project’s objectives.

Expected Results and Preliminary Discussion

It is expected that the resulting vector dataset will not show a significant difference to any of the variables outlined for testing in the objectives of the project. While confidence is high in this statement, it is also expected that in the event that a difference does occur, it can be modeled so that a regression adjustment will yield more accurate results and also depict general relationships between LIDAR and field conditions. The importance in the eventual acceptance of these results is that these measurements, while simple, demonstrate the first step in a new generation of active sensor utilization in forest resource management. These new forest measurement techniques do not provide all the estimates needed for any timber volume estimation, but this project, even at this inconclusive stage, shows positive results. 

This technology has been formed into a concise inventory protocol in a double sampling format here in the Department of Forestry, at Mississippi State University. New studies at MSU are examining relationships of intensity and distributions of returns with respect to species identification and understory assessments. With these technologies, we hope to approach the point where detailed and accurate timber inventories can be made on the parcel level with minimal field labor. This would definitely provide a revelation in remotely sensed data that traditional spectral information has never afforded. But for now, the forest resource management community continues to view this tool with great enthusiasm as one to improve efficiency in labor, time, and money.

Curtis A. Collins

For additional questions on this topic you may contact Dr. David Evans.