Samuel E. Cox

Image Analysis Compared with Other Methods for Measuring Ground Cover

ABSTRACT Ground cover is a key indicator of rangeland health but conventional methods for measuring ground cover are labor intensive. Analysis of digital images has the potential to reduce ground-cover-measurement labor requirements. We compared cover measurements by image analyses of digital images (sensor resolution = 0.97 mm/pixel ground sample distance) with measurements derived from a laser point frame, and from two transect methods. We found there was low agreement in plot-to-plot comparisons but results were usually not different when averaged over a large number of plots or transects. We conclude that image analysis of large numbers of samples (images) produce mean values not different from conventional field methods, and, that image analysis is a superior choice for detecting relative change, since it facilitates greater data collection, reduces human bias by limiting human judgments, and provides a permanent record in images that can be retained for future scrutiny.

The Accuracy of Ground-Cover Measurements

Abstract Ground cover is a key indicator of rangeland condition and influences rangeland management decisions, yet there have been few advances in ground-cover measurement methods. The advent of digital photography and automated image processing promise a revolution in the way ground cover is measured. To assess the potential for automation we compared conventional and automated methods for measuring ground cover against known artificial populations. The known populations were created from 20 nadir images of a Wyoming big sagebrush (Artemisia tridentata Nutt. ssp. wyomingensis Beetle & Young) vegetation type acquired with a 5-megapixel Olympus E20 digital single lens reflex camera mounted on an aluminum camera frame at 2 m above ground level. The images were converted to color, 2-dimensional images that no longer represented real-world conditions but had known cover values and conserved a simplified form of the pattern and spatial context of the plant community. These images were then printed at 1:1 scale to a 1 × 1-m poster. Posters were evaluated for color cover under laboratory conditions using the conventional techniques of steel-point frame, laser-point frame, line-point intercept, ocular estimation, and line intercept. Photographs of the posters were measured for color cover using standard and custom-created algorithms within the VegMeasure image analysis framework, and using the Digital Grid Overlay method. Results indicate that conventional techniques had significantly greater correlation (≥ 92% agreement of measured to known) than measurements from the algorithms used in the VegMeasure analysis (70%). The critical factor influencing accuracy of point-sampling methods was the area of the contact point for the given method. These findings provide an important measure of relative accuracy among methods for land managers and for researchers seeking to improve rangeland monitoring methods.

Image-based monitoring to measure ecological change in rangeland

High-resolution, image-based methods can increase the speed and accuracy of ecological monitoring while reducing monitoring costs. We evaluated the efficacy of systematic aerial and ground sampling protocols to detect stocking-rate differences across 130 ha of shortgrass prairie. Manual (SamplePoint) and automated (spectral) image-analysis methods were compared for both aerial and ground data. Vegetative cover changes due to grazing were detectable from 1-mm ground sample distance (GSD, a measure of resolution) digital aerial photography with as few as 30 samples yielding enough data to predict bare ground within ± 5%. We found poor agreement between automated and manual image-analysis methods, but good agreement between manual analyses of imagery from the air (100 m above ground level [AGL]) and from the ground (2 m AGL). We conclude that cover measurements made using SamplePoint from 1-mm GSD images (from 2 or 100 m AGL) can detect ecologically important changes in key indicators such as bare ground. The costs of ground and aerial methods differ markedly, and we suggest that aerial imagery is most cost effective for areas larger than 200 ha.

Detection-Threshold Calibration and Other Factors Influencing Digital Measurements of Ground Cover

Abstract New methods of image acquisition and analysis are advancing rangeland assessment techniques. Most image-analysis programs require users to adjust detection thresholds for color or object classification, a subjective process we postulated would be influenced by human error and variation. We developed a ground-cover–measurement calibration procedure, the digital grid overlay (DGO), which is similar to image point sampling (dot grid) advanced by earlier researchers. We asked 21 rangeland professionals to measure ground cover using 2 subjective visual-estimate methods (threshold adjustment process, or TAP, and external [to the software] visual estimate, or EVE) and the DGO on 5 different nadir-view images of rangeland. We also compared cover measurements made by DGO-calibrated software in automated batch processing against DGO manual-only measurements. We found an unacceptable range of variation among rangeland professionals using TAP. The DGO and EVE values were more closely aligned. We discovered an age-related bias in bare-ground measurements: all users over 50 years of age classified more bare ground than did all users under 50 years of age when using TAP. One explanation for this bias is age-related yellowing of the eye lens. Manual DGO measurements required up to 15 minutes per image compared to about 1 second per image for automated computer analysis after software calibration. The greatest bare-ground difference between the DGO-calibrated software and manual DGO measurements for the data sets analyzed was 5.6% and the correlations imply that reasonably accurate automated measurements can be used for bare-ground measurements from digital-image data sets. The exception is where the software cannot adequately separate litter and bare ground. The digital methods we tested need improvement. However, external calibration (DGO or EVE) of current-generation image-analysis algorithms bring economical, statistically adequate monitoring of extensive land areas within the realm of practical application.

Very Large Scale Aerial Photography for Rangeland Monitoring

Abstract Ecological assessment of ground cover by conventional on‐the‐ground point sampling is labor intensive, expensive, and biased by access. Historically, motion blur has prevented high‐resolution aerial photography from being used for ground cover measurements. To reduce motion blur we used a fixed‐wing, 225‐kg (empty weight) airplane flown at 72 km/hr ground speed 100 m above ground level with both a modified Hulcher 70 mm camera (500 mm lens, Kodak Aerocolor HS SO‐846 film, 1/4,000‐second shutter speed), and a Canon 11. 1 ‐megapixel, automatic, digital single lens reflex, color camera (420 mm focal‐length lens). The resulting very‐large scale aerial (VLSA) photography had resolutions of 5.0 and 2.1 mm GSD (Ground Sample Distance) for the scanned film and digital‐camera images, respectively. Motion blur was minimal. The cost for obtaining 450 VLSA photographs over a 70,800‐ha watershed was

Technical Note: Lightweight Camera Stand for Close -to -Earth Remote Sensing

0.07 ha. We found our methods well adapted for extensive aerial surveys to monitor the ecological condition of rangeland watersheds.

Grazing Intensity and Spatial Heterogeneity in Bare Soil in a Grazing-Resistant Grassland

Abstract Digital photography and subsequent image analysis for ground-cover measurements can increase sampling rate and measurement speed and probably can increase measurement accuracy. Reduced monitoring time (labor cost) can increase monitoring precision by allowing for increased sample numbers. Multiple platforms have been developed for close-to-earth remote sensing. Here we outline a new, 5.8-kg aluminum camera stand for acquiring stereo imagery from 2 m above ground level. The stand is easily transported to, from, and within study sites owing to its low weight, excellent balance, and break-down multipiece construction.

Large-scale Aerial Images Capture Details of Invasive Plant Populations

Abstract We used very large scale aerial (VLSA) photography to quantify spatial patterns in bare soil in the northeastern Colorado shortgrass steppe. Using three pairs of pastures stocked at moderate (0.6 animal unit months [AUM] · ha−1) versus very heavy (1.2 AUM · ha−1) rates, we detected greater bare soil under very heavy (mean  =  22.5%) versus moderate stocking (mean  =  13.5%; P  =  0.053) and a lower coefficient of variation across pastures under very heavy (0.48) versus moderate stocking (0.75; P  =  0.032). Bare soil exhibited significant positive spatial autocorrelation across distances of 60–120 m under moderate stocking (Morans I  =  0.14), while patchiness at this scale was eliminated under very heavy grazing (I  =  −0.05). Across distances of 120–480 m, we observed no spatial autocorrelation with either stocking rate. Spatial autocorrelation was greatest at a separation distance of 2 m (I  =  0.48–0.58) but was unaffected by stocking rate at this scale. Thus, very heavy grazing did not increase spatial autocorrelation in bare soil across scales of 2–480 m. Means and variability in the distribution of bare soil were not influenced by ecological site. Bare soil increased primarily at the scale of individual plant clusters through both increases in the density of small (2–20 cm) bare patch intercepts and increases in the frequency of bare patch intercepts of 20–60 cm (rather than < 20 cm). Our approach demonstrates the utility of VLSA for analyzing interactions between grazing and other landscape features and highlights the importance of spatially explicit sampling across broad scales (pastures) while testing for potential shifts in patchiness of bare soil at the scale of plant interspaces.

Dual-camera, high-resolution aerial assessment of pipeline revegetation

Abstract Satellite and high-altitude aerial remote sensing have been used to measure dense infestations of invasive weeds over very large areas but have limited resolution and cannot be used to detect sparsely distributed weeds. Ground-based methods have provided detailed measurements of invasive weeds but can measure only limited areas. Here we test a novel approach that uses a lightweight airplane, flying at 72 km ṡ h−1 and 100-m altitude, to rapidly collect high-resolution images over relatively large areas. We obtained 1 987 images, each representing 48.5 m2 of mixed-grass prairie with 2-mm resolution (ground sample distance). From these images we were able to reliably measure small patches and even individual plants of the invasive forb Dalmatian toadflax (Linaria dalmatica [L.] P. Mill.). Ground-based measurements of aboveground toadflax biomass were highly correlated (R2 > 0.93) with point-intercept and visual-estimate cover measurements from aerial images. The time required to analyze images ranged from 4 to 45 seconds for presence/absence data and from 1 to 6 minutes for cover data. Toadflax was present in 795 of 1 987 images but exceeded 1% cover in only 99 images. Given the observed variation among images in toadflax cover, at least 400 images were needed to precisely estimate the mean toadflax cover of 0.2%. These results suggest that such high-resolution aerial imagery could be used to obtain detailed measurements of many invasive weed populations. It may be most useful for identifying incipient weed infestations and expanding the scale at which population-level attributes of weed populations can be effectively measured.

Postfire Downy Brome (Bromus tectorum) Invasion at High Elevations in Wyoming

Energy-extraction results in significant disturbance to rangelands in Wyoming and other western US states. Although reclamation is required by law, US General Accounting Office reports from 1999 and 2005 are clear that affected government agencies have—over much of the past decade—had difficulty accomplishing mandated environmental monitoring of extraction-related disturbance. We evaluated two pipeline rights of way (ROW) using nested images (1- or 2- with 13- or 20-mm ground sample distance (GSD)) acquired during Very-Large Scale Aerial (VLSA) surveys. Aerial monitoring allowed for the collection of large numbers of geocoded samples, and for subsequent cover measurements using methods with demonstrated accuracy equal to that of conventional ground-based methods. Both pipelines had vegetative-cover deficiencies relative to their Plan of Development (POD) requirements. Using bare ground and ground-cover measurements from the higher-resolution imagery, we present a spatial representation of each pipeline ROW that allows quick identification of sections of the ROW that may need further reclamation action to meet POD standards. We also present aerial monitoring costs. We recommend VLSA pipeline surveys as a means for facilitating required environmental monitoring and for addressing the monitoring backlog that has developed with increased energy-extraction activity.