J. Range Manage.
54: A77-A89 March 2001

Paul T. Tueller, Professor of Range Ecology, University of Nevada Reno, 1000 Valley Road, Reno Nevada 89512, email: ptt@equinox.unr.edu

Key words: Rangeland forage production, utilization, remote sensing, image processing

Remote sensing has been used in range management primarily as a tool to map various rangeland ecosystems or plant communities. Efforts have focused on using both photo interpretation and image processing or a combination of the two to accomplish this. The determination of range production and utilization is basically a classification problem. Since range management is a land based science, we are concerned with the distribution of information across landscapes. The need is to identify and map areas of differences in forage production over time to describe the different levels of utilization so that the data might be useful to managers. Production of forage is best defined as the increment of new biomass over time during a growing season. This is a difficult problem because of the need to examine the numerous species including both herbaceous forage and browse, both in the open and under forest canopies. In fact, at the present time it may not be possible to do this with any high degree of accuracy. However, there is potential that remote sensing and information systems offer considerable hope for the future. The amount of this forage harvest over time (utilization) is, in many ways, the sine qua non of range management. This paper is a review of the literature and possibly some new ideas on how we might classify landscapes as to the level of forage production and the amount removed or harvested over some period of time using remotely sensed data.

Monitoreo remoto ha sido usado en cierta mancra como una herramienta principal para trazar varios ecosistemas de praderas y comunidades de plantas. Losesfuerzos se han enfocado en el uso de la interpretacion de fotos, asi como elementos. La determinacion del rango de produccion y la utilizacion de esta es basicamente un problema de clasificacion. La manera en que se distribuyen o se alincan los surco de tierra se basa en la composicion cientifica de la tierra, esto por lo tanto, crea el problema de distribucion de information a traves de los paisajes. La necesidad consiste en la identifiacion y en el trazo de las diferencias que existen en las areas, las cuales permiten una produccion de forraje a traves del tiempo y que a su vez sirve para describir los diferentes niveles de uso. La produccion de forraje se define como el incremento. Esto es un problema muy dificil que implica la neccsidad de examinar numerosas especies de forraje herbaceo y pastura, las cuales se hayan en campo abierto y bajo la copa forestal. Es importante tomar en cuenta que esto en la actualidad no puede ser posible cualquiera que sea el grando de exactitud. Sin embargo, es posible que el monitoreo remoto y los sistemas de informacion ofrezcan unaesperanza alentadora en el futuro. La cantidad de cosecha del forraje a traves del tiempo, es en muchas maneras el sine qua non del manejo y distribucion del campo. Este trabajo es una revision de articulos escritos en este campo asi como podriamos clasificar los paisajes, el nivel de produccion de forrajc y la cantidad de cosecha buena y la que se desecha: todo esto a traves de un cierto periodo de tiempo tomando en cuenta el uso de informacion de monitoreo remoto.

Range scientists have always been concerned about forage production. Numerous methods have been developed over the years to determine forage production or standing crop. The primary concern has been to determine carrying capacity and the determination of proper stocking for a given unit of rangeland for a given grazing season. Traditional methods have often been time consuming and field sampling plots have not always been successful in terms of accurately measuring the amount of forage produced over large areas of rangeland. This suggests the potential use of remote sensing as a means of determining forage production and/or standing crop. (Tueller 1989).

The removal of vegetation or extraction of plant material from the forage base, is an indication of utilization. If the amount of standing crop is reduced from time A to time B and this is correlated with the happenstance that a given number of livestock or wild animals (including invertebrates) were on the range unit for the time interval between A and B, then the difference can be attributed to utilization.

This paper considers the application of remote sensing techniques that may be useful in measuring forage production, standing crop and forage removal by grazing and browsing animals (utilization). Even though we often think of remote sensing as a tool to classify and map vegetation and other features on a landscape, the remote determination of production and utilization is essentially a classification process. The question then is how can one identify spectral signatures that hold true under specified conditions to indicate the amount of standing crop and the amount of forage utilized? And what are the spatial requirements to do this if the spectral data is useful. What part of the EMR spectrum might provide useful data in this regard? What analytical procedures can be used to extract such information from the remotely sensed data including upscaling from high-resolution to low-resolution imagery?

Remote sensing technology has been steadily improving since its beginning somewhere in the late 1960s. Initially black-and-white aerial photography was the standard, although the medium changed to color as the technology advanced. Since that time, there have been numerous improvements in remote sensing technology. The landmark remote sensing experiment began in 1972 with the first Landsat satellite and the acquisition of multispectral data from space with low spectral and spatial resolution. Thereafter, data sets increased in both spectral and spatial resolution. Review of the Range Resources chapter in the 1975 Manual of Remote Sensing (Poulton et al. 1975) shows very little information about site productivity and utilization as measured by remote sensing. However, the subject of change detection is discussed and emphasis was given to image analysis by the photo interpreter. Image processing techniques coupled with GIS and GPS technology did not exist at this time. Only now are we seeing their rapidly expanding role. Below I attempt to describe some of the newer kinds of remote sensing technology that might lead to reasonable measurements of forage productivity/utilization.

Can we measure net primary productivity using present day remote sensing technology? Net primary productivity (NPP) is the difference between the accumulation of photosynthate and the accumulative autotrophic respiration by green plants per unit time and space. The NPP is measured by evaluating meteorological or evapotranspiration models, from incident solar radiation and its absorption coefficiency by plant canopies, or by the simulation of biological processes affecting NPP such as photosynthesis, respiration, and transpiration. Liu et al. (1997) used a model utilizing daily meteorological and AVHRR NDVI data to predict NPP for boreal ecosystem productivity. Empirical models for determining biomass from remotely sensed data have limitations (Begue 1993). Deterministic models are now used to determine biomass from remotely sensed data. These models simulate the functioning of the canopy on the basis of the absorption of photosynthetically active radiation (PAR: 400-700nm) and its conversion into dry matter by the plant. The absorption efficiency of the canopy (i.e., the fraction of incident PAR absorbed by the canopy) depends on several factors: vegetation clumping, leaf area index, a leaf inclination distribution function, optical properties of the leaves, solar angles, and fraction of the incoming diffuse radiation (Asner 1998a, Asner 1998b and Asner and Wessman 1997). Both reflectance and absorption are strongly related; therefore it may be possible to assess the efficiency from remotely sensed reflectance data. Vegetation indices are often substituted for reflectance because they are less dependent on external factors such as soil reflectance, irradiance geometry, etc. Actual leaf area indices are difficult to measure and thus they may not provide a useful relationship to biomass and biomass changes.

We are on the edge of an unprecedented period in remote sensing history. We are at the point of having high spatial resolution imagery from space. This is in the form of the IKONOS satellite data that is just becoming available from SpaceImaging, Inc. Also, Interferometric Synthetic Aperture Radar (IFSAR) system will now provide 1-m elevation data, which in the future may assist in the vegetation and production classification process for range landscapes. The IFSAR is a dual antennae active sensor that pulses the topographic surface. The resulting processing of signals provides a relatively high-resolution digital elevation model (DEM).

Fig. 1 an IFSAR image for a portion of the Camp Williams National Guard facility south of Salt Lake City, Utah. The IFSAR image is on the left and a Digital Elevation Model (DEM) on the right. The DEM surrogate shows promise as a predictor of forage production along with other remotely sensed data.

In the last few years there has been considerable discussion of the use of hyperspectral data. These systems have very high spectral resolution. Over 45 different systems are in use worldwide with spectral resolutions varying from 16 to 511 bands. The commonly used Airborne Visible Infrared Imaging Spectrometer (AVRIS) has 225 spectral bands. It is a unique optical sensor that delivers calibrated images of the upwelling spectral radiance in 224 contiguous spectral channels (also called bands) with wavelengths from 400 to 2500 nanometers (nm). Most of the bands are 10 nm wide or less. The only problem with these system is that they are so new that they have not been researched thoroughly. Most of these systems still provide pixels of 5 to 20 meters, although it is possible to acquire 1-m pixels. To this date a hyperspectral system has yet to be flown from space. Other systems with both increased spatial and increased spectral resolution are also becoming available. Many assume that this increased spectral and spatial resolution will be the final solution of many remote sensing sampling problems. However, they fail to realize that we, on most rangelands situations, are still working with a mixed pixel with considerable spectral ambiguity. Hence it is still too early to champion these systems because even though they have great promise, they have not yet been proven for determining ecosystem productivity and change.

A relatively new technique, spectral demixing, is being used to analyze the hyperspectral data. This procedure defines endmembers (groups of relatively pure spectra) in the spectral space (distribution of spectra when placed in a two dimensional graph) whereby it is possible to define what these endmembers are in terms of ground components such as bare ground, shrub vegetation, herbaceous vegetation, rock, gravel and litter. The relative proportion of these endmembers may serve as surrogates of site productivity on certain rangelands (Wessman et al 1997, Roberts et al 1998).

We have been working with large scale, high resolution systems including digital aerial cameras and multispectral videography. These systems give submeter pixels. While we can often recognize certain species and in the infrared obtain a relative indication of vigor, there is yet no way to accurately quantify the NPP and determine changes in the standing crop.

The classification paradigm assumes that the remote sensing scientists and practitioners can place like pixels in groups to represent features of interest on a landscape. The features of interest could be sites with different levels of productivity or different levels of utilization. This depends upon the success of defining a spectral signature that can describe the polygon and its attributes. So far, the success for describing and quantifying such features has been problematical for various reasons. Some of these reasons are: 1) different phenologies, leaf shapes, and stem shape characteristics among the numerous species making up the biomass on a particular range site, 2) the variation in substrates, 3) the time of day and time of the year the remotely sensed data is obtained and other factors. These all act to provide noise in the data sets used to define the signature with respect to the definition of the feature. (Level of biomass production on a range site level of utilization on a range site). A whole host of algorithms are available to bunch these like pixels together. Some are unsupervised where the brightness values and their mean and standard deviation provide a means of clumping each of the pixels into like classes within the specified ranges of these parameters. Often the unsupervised classification can be accomplished quickly and the distance between means, the number of iterations and the number of classes allowed can be specified.

A supervised classification is often beneficial. In this case, known pixels within the scene (raster file) are selected as "training sets," that is, pixels are selected based on ground knowledge that they represent a class of interest to the range manager, e.g., a specified level of production/utilization. The DEMs provide slope and aspect data which may be useful along with vegetation indices to increase classification accuracy on sites with differing levels of NPP.

Even when the classification algorithm allows identification of discrete pixels that identify a given level of forage production and utilization, there is often considerable noise in the system based on the spectral variability found on these sites. The spectral brightness values representing these situations are seldom homogeneous. Therefore, the resulting classes have considerable noise, i.e., pixels that will not correspond to the definition, this noise or influx of unlike pixels can often be reduced by using some sort of a noise reduction filter. A common filter is a smoothing algorithm where each pixel value is related to the surrounding pixels by taking the mean of several pixels surrounding each pixel in the image raster file. Such a low pass filter may not be ideal because the mean value of the filter window (the filter output for each pixel) can be skewed by the extreme values of noise cells. If such a filter is used to reduce noise, it is often possible to drop the highest and lowest pixel values to reduce this problem. Other filters such as a modal filter can be used. The idea is to create homogeneous classes in a classification that are true to a situation (high forage production/high utilization or low forage production/utilization).

As each class is defined and determined closely correlated to a production/utilization situation, it is of value to determine the success or accuracy of the classification results. To interpret the results of a classification, it is useful to compare the classes to any available information about the types of materials and ground cover found in the scene. What is the level of correlation between the ground information and the level of production/utilization?

A ground data error matrix can be developed. In such a matrix the accuracy values of each column indicate the percentage of cells in the ground data class that were correctly classified. Values less than 100 % percent indicate errors of omission (ground data cells omitted from the output class). These values are sometimes referred to as producers accuracy. The accuracy values for each row shows the percentage of sample pixels in each output class that were correctly classified. In a like manner for the rows with values less than 100% indicate errors of commission (pixels incorrectly included in the output class). These values are sometimes referred to as user’s accuracy. Many recent studies and proposals that this writer has reviewed have failed to provide good methodology for determining accuracy. This, of course, provides remote sensing data that is often meaningless.

A new paradigm, that of landscape ecology, is now providing insight into the distribution of rangeland ecosystems at a level that has been little studied. Patch dynamics, a powerful way of dealing explicitly with spatial heterogeneity, and has emerged as a unifying concept across different fields of earth sciences (Wu 1999). Upscaling from large to small-scale data will often be required as we expand remotely sensed data to large land areas. This is discussed later in this paper. I like to think about or talk about the concept of the ‘total sample.’ Field sampling methods or point sources often are only ill defined as we extrapolate the results to the larger land area. The ‘total sample’ is what we may be able to obtain using remote sensing technology. The point is this - if we can accurately measure rangeland vegetation characteristics that are indicative of production, utilization or other attributes from the remotely sensed images then we have that data across the total landscape, the questions of scale and spectral ambiguity not withstanding.

Factors influencing the measurement of plant growth and utilization are Net Primary Productivity (NPP), Photosynthetically Active Radiation (PAR), Leaf Area Index (LAI), and other factors influencing relative photosynthesis. The literature on this subject focuses on estimating factors rather than measuring them. This is indicative of the difficulties for using remote sensing for measuring many of these attributes useful to range managers. I suppose it is possible to think of PAR and NPP, e.g., as surrogates to weight per unit area and the accumulation of biomass on a rangeland site. However, the level of accuracy in doing this is not clear.

There are few examples in the literature of attempts to measure production and utilization on rangelands. There are a few which report; e.g. NPP estimates (again) based on AVHRR over large areas almost on a continental or sub continental scale (Liu et al 1997, Mougin et al 1995). These may be useful for describing the general production of ecosystems over these large land areas but unfortunately these studies do very little to help the local range manager make correct decisions on stocking rates, maintaining surface soil, assuring biodiversity, etc. Failures are usually based on a misunderstanding of just what remote sensing can do for us and with what level of accuracy.

Van Leeuwen and Huete (1996) examined the potential of using vegetation indices to generate canopy reflectance "mixtures" and to estimate fractions of absorbed photosynthetically active radiation (fAPAR) with varying LAIs, soil background, combinations of vegetation component spectral properties, and one or two horizontal vegetation layers. They found that the vegetation index response of a standing leaf, litter canopy was shown to decrease or increase, depending on the vegetation index response of bare soil. Dark leaf litter canopies were associated with a higher vegetation index response than the bright leaf litter canopies. Litter on top of green vegetation generally decreased the vegetation index response, depending on how much of the green vegetation was obscured by litter. This and similar studies suggest the difficulty in understanding the significance of the mixed pixel.

Bork et al (1999) compared ground based narrow-band sensing with typical satellite broad band data. They found that multiple regression models had little advantage over simple regression models in predicting plant cover. They were able to separate grazing treatments with relatively minor ecological differences using the narrow-band data

When it becomes possible to measure standing crop and biomass of forage using remote sensing, it will be necessary to describe the production across range landscapes. How do we go about doing this? The approach will likely be a multiscale approach. The multispectral aspect will be used to define and quantitatively describe biomass on large scale data and then carry such data upward to larger land areas using some sort of an upscaling procedure. The example here is one of describing the use of high resolution data derived from a Kodak digital camera, measuring various attributes and then merging the high resolution with lower scale resolution so that the data can be extrapolated to the larger land areas. In this example the upscaling is from the Kodak 0.2 meter pixels to either Landsat Thematic Mapper images (30 m pixels) or the Indian Remote Sensing 6m data.

We measured various plant communities with different levels of biomass and species composition for a short grass prairie site with some shrub encroachment on the Otero Mesa on the Ft. Bliss U.S. Army reservation in southern New Mexico and west Texas near El Paso. Fig. 2 shows a site with a Tobosa grass (Hilaria mutica) swale, an adjacent site with a stand of the small shrub, Pectis lemmonii and then above the swale various communities dominated by blue (Bouteloua gracilis) and black gramma grass (Bouteloua eriopoda). Scattered across the area are the shrubs mesquite (Prosopsis glandulosa), yucca (Yucca elata), Crucifixion Thorn (Koberlinia spinosa) and little leaf sumac (Rhus microphyllum). These shrubs cannot always be identified at this scale but the total density of shrubs in these grasslands can be determined. The Kodak digital data with 0.36 m pixels is placed side by side with the same area extracted from the 30 m pixel Thematic Mapper data.

Fig. 3 is a IHS (Intensity, Hue, Saturation) merged image which is used to combine the pixels from each scale together. The merged image shows that the data can be combined successfully and then Fig. 4 is a stepwise linear classification of the merge. A successful merging of these data sets provides a way to extrapolate the data from the large scale data to large acreage typical of those found on rangelands. The idea here is to examine characteristics of a range site with high resolution remote sensing data. The characteristics ideally would include above ground biomass, as well as species composition. Then an upscaling procedure as outlined here could be used to define these characteristics across a larger landscape area.

Fig. 2 A gamma grass range site on the Otero Mesa, Ft. Bliss, Texas. The Kodak digital image mosaic shows a tobosa grass (Hilaria mutica) swale in the upper left and some roads. The TM extract is from the same areas as the mosaicked Kodak images.

Fig. 3 An IHS (Intensity, Hue, Saturation) merge of the Kodak digital images with the 30 m pixels of the Thematic Mapper data. This merge shows that the high resolution data can be carried across to the merge thus allowing expansion to the larger area.

Fig. 4 A classification of the Kodak/Tm merge. This classification is describes the area of the merge occupied by specific species and potentially their biomass characteristics which can then be extrapolated to the larger rangeland areas.

The word contemporary is used since remote sensing technology is evolving so quickly that it is difficult to predict what we will be able to measure even in the next few months as new systems continually become available. Some of these new systems have been briefly reviewed above.

I would like to acknowledge the data from our SERDP Project (CS 1098), Emerging Remote Sensing technology for the data and graphics on upsclaling. Thanks go to Adrienne Breeland, Graduate Research Assistant for preparation of the graphics in figures 2 - 4 and Kelly Ellsworth with assistance with the preparation of the manuscript.