6.5 Polarization Techniques to Enhance Remote Sensed Imagery
Investigators:
Dr. Timothy Stombaugh, Biosystems and Agricultural Engineering, tstomb@bae.uky.edu
Dr. M. Pinar Mengüç, Mechanical Engineering, menguc@engr.uky.edu
Dr. Scott Shearer, Biosystems and Agricultural Engineering, shearer@bae.uky.edu
Introduction
Good Precision Agriculture (PA) management decisions cannot be accomplished without accurate spatial data. The primary tools that most producers use to gather spatial data are intensive soil sampling and yield monitoring. Remote sensing (RS) is a technology that has received much attention recently. The technical definition of RS is any sensor that can measure some quantity remotely or without coming in contact with it. In agriculture, most people understand RS to be crop imagery obtained from satellites or aerial vehicles. A less common but very important variation of RS is closer range sensing as in a land vehicle-mounted sensor.
RS is a very desirable sensing technique for several reasons. A sensor located at one point in time and space can instantaneously obtain data from a wide area, which eliminates the need for extensive human sampling and measurement. RS is a non-contact technique, which means that the crop is not disturbed or damaged in any way. Some RS equipment, such as near-infrared (NIR) cameras, can measure quantities that cannot be seen or observed by a human.
Many researchers have attempted to quantify different crop parameters using RS imagery, and have had some success. The frustration commonly encountered is that relationships that are identified often are only valid for a limited set of environmental, geographic, or crop parameters. Very few globally applicable relationships have been developed. RS studies to date almost exclusively used the reflectance intensity of different wavelengths of radiation (light). These images are affected by shadows cast within the plant canopy on bright sunny days, or by background noise from the soil surface prior to the point in the growing season when the full crop canopy is developed.
One concept that has not been explored for use in RS is the use of polarized (or, more generally speaking, elliptically polarized) light. Most people are familiar with the primary effect of polarized sunglasses or polarized filters for photographic equipment, which is to reduce glare from highly reflective surfaces. There are many advanced types of polarizing filters that can be used to create a variety of image enhancement effects. One possible application is to use polarization and cross polarization to enhance weed competition in RS images. The objective would be to use polarizing filters to eliminate image features that are not of interest while enhancing the more pertinent features. For example, in soybean crops it may be Johnson grass that is of interest to the producer because of the expense required to control this perennial weed, while foxtail is easily controlled with cost-effective herbicides. With polarizing filters it may be possible to enhance the Johnson grass infestations in RS imagery while suppressing crop and foxtail information from the image. However, a straightforward application of polarization with the off-shelf filters may never give the desired effects. To achieve the desired image enhancement effects, it is imperative that researchers gain an understanding of the variation of contrast in different images via the use of linear, circular or elliptical polarized filters.
Objectives
The primary goal of this project is to explore the use of polarized RS imagery for sensing critical field crop parameters. This goal will be achieved through the completion of the following specific objectives.
The investigators believe that many applications for use of polarized light may exist for field crop sensing. This exploratory project will focus specifically on detecting weed infestations in corn and soybeans, and detecting nitrogen stress in wheat, which the investigators believe are quantities that can be enhanced using polarizing filters and identified unambiguously using novel algorithms.
Background
RS has been implemented with some degree of success to measure a variety of field parameters. For example, Dr. Stombaugh has been involved in a study to identify areas of compaction in fields using visible and NIR imagery (Wells et al., 2001). Some researchers have tried to predict crop yield variation from RS data (Yang and Anderson, 2000; Gopalapillai and Tian, 1999). Bajwa and Tian (2001) used RS to map weed infestations in soybeans. Burks et al. (2000a and 2000 b) used texture of the plant canopy to discriminate between weed species. In these studies, models were developed relating RS spectral intensity data to ground truth data; however, the models only had limited ability to predict the same parameters from imagery taken in other fields.
The studies mentioned above utilized combinations of reflectance intensity at different colors or spectra of visible and invisible radiation to compute prediction models. It is important to note that the visible spectrum is limited to wavelengths from 400 nm (deep blue) to 700 nm (dark red). Many modern sensors can also measure NIR (700 to 1200 nm), mid-infrared (1200 to 4000 nm) or far-infrared (4,000 to 12,000 nm) radiation.
One common parameter used in many analyses of crop or plant performance is the Normalized Difference Vegetative Index (NDVI). The NDVI uses both red and NIR reflectance information and computed by:
The NDVI is based on the spectral intensity profile of the RS data. The ability to analyze the images based on the spectral and intensity profiles has helped agriculture extensively over the years. However, these approaches have had only limited success in visualizing more subtle variations in the vegetation physio-chemical properties. Therefore, it is desirable to use more advanced approaches to develop the next generation tools for agricultural use.
In addition to its spectra and intensity, light (or any electro-magnetic radiation) has phase information, which can be quantified in terms of the polarization of radiation. The phase comes about because light is essentially a wave. When a light wave interacts with a surface or a particle, its polarization may change. This change usually occurs as a result of variations in the index of refraction and shape/structure of the matter (surface or particle). The degree of change in polarization may also depend on the wavelength of radiation, which may be significant at certain wavelengths. It follows that polarized filters, which filter out certain phases of light, can be used to reduce or enhance reflectance from certain plants or other objects with unique surface or shape characteristics. To date, spectral-polarization sensing for precision agriculture has never been investigated.
Dr. Mengüç has been involved in polarization-based sensing research for more than a decade. He and his students have developed a novel elliptically-polarized light-scattering-based particle characterization methodology. This approach has already been applied to detection of soot agglomerates in flames, phytoplankton in water, ceramic particles, nanocrystals, as well as to cotton fibers (Govindan et al., 1996; Manickavasagam and Mengüç, 1997; Mengüç and Manickavasagam, 1998; Manickavasagam et al., 1998; Kozan et al, 2002). During the last three years, there have been commercialization efforts for this approach, supported by the National Science Foundation in terms of four SBIR (Small Business Initiative Research) Grants. There is a pending patent (Mengüç and Manickavasagam, 2001) that covers not only the detection technique (eyes) but also the computational-procedure to interpret the results (brain) for application to small particles.
Extension of the polarization-based characterization modalities from single-scattering applications to multiple-scattering media is already underway in Mengüçs research group. For example, recently detailed Monte Carlo models have been developed, and the idea is already being tested for characterization of foam size distributions (Vailon et al., 2000; Wong and Mengüç, 2002a, 2002b). Therefore, implementation of these concepts to precision agriculture will have a very strong chance for success.
Based on preliminary data and discussions, we have observed that the polarization-based sensing may be most promising in the near-infrared spectra, from 700 nm to 900 nm wavelength range. This spectrum is within the range of commercially available infrared black-white films from Kodak and common digital cameras that utilize CCD or CMOS sensing elements. Dr. Mengüç has extensive experience with the films, and Dr. Stombaugh has experience with the digital media (Fig. 1).
Figure 1. NIR photographs using film (left, by Mengüç, see the corresponding web page http://www.engr.uky.edu/~menguc/MPM-IR2/home.htm) and digital (right, by Stombaugh) media.
Procedures
Since there has been no prior research in agricultural applications with polarized NIR reflectance, the first task will be to evaluate the utility of this technique. We will first identify a combination of specific polarization orientations and wavelength spectra that can be used to identify weeds in soybean and corn crops. We will utilize the University of Kentucky weed research plots for these initial studies. Since we know that healthy plants are highly reflective of NIR radiation, we will use NIR film and filters to enhance the plant material. An example of one specific inquiry we will conduct is to identify grassy weeds in soybeans. Since soybean leaves are generally round compared to the grassy weeds, we will use a elliptic (including circular) polarizing filters to suppress the soybean information thus enhancing the more linear weed information.
The second quantity we will attempt to measure is nitrogen stress in wheat. This quantity has been intensively studied by several UK researchers and by other organizations in collaboration with UK. We will partner closely with these studies to increase the potential for success of all parties. Again, we will identify filter and spectra combinations that are affected by nitrogen content.
Initial imagery that we collect will be ground based. As we begin to develop the relationships, we will collect aerial imagery of weed-infested fields using a small aircraft to make sure the techniques are not adversely affected by sun angle or atmospheric conditions. We will attempt to collect data at six critical times throughout the growing season including pre-plant, early growth, full canopy, bloom, onset of senescence, and later in senescence.
Once the acceptable spectral and polarization filtering combinations are identified, the next task will be to develop a lens/filter combination to collect imagery that will identify a certain set of parameters. We expect that the best correlations may come from combinations of several images with different filter combinations. To collect these images simultaneously, we will develop a multi-faceted lens that creates duplicate images in the same frame (Fig. 2). Each facet of the lens will be coated with a different filter so that all required image data are collected simultaneously.
Figure 2. Multifaceted lens that can be used to create duplicate images in one frame (Mengüç et al, 2002 (Patent Disclosure made to the University of Kentucky).
One of the challenges faced when using RS imagery is the fact that a significant amount of image processing is often required to glean useful information. The inherent effect of filtering techniques is that the useful features of the image are already enhanced in the raw image. Nevertheless, there may be some image processing necessary, especially when the multifaceted lens is used. We will develop the algorithms and associated software that can be used to extract pertinent information from the images. This package will be as simple as possible so that general PA users can analyze images with little or no training.
A longer-term goal will be to build relatively inexpensive digital cameras, which can be mounted in small remote-controlled airplanes. This approach will give the farmer the required local on-demand measurement opportunities, and combination of these approaches with the data available from satellite images is likely to yield the optimum solution. During the course of this project, we will build the inexpensive cameras for these purposes.
Expected Benefits
The techniques developed in this project will provide a new tool for PA data collection. We believe there is tremendous potential for reliably identifying problem areas in a field such as weed pressure or plant stress due to nitrogen deficiency or other factors. Because the power of the tool is in the hardware development, it should be a relatively simple tool to use.
Since this is a new exploratory research area, there is the potential for many other applications of the technology in agriculture. We plan to look at other factors that could be identified using remote sensing imagery.
Deliverables
The primary deliverable from this project will be a procedure to filter and preprocess remote sensed images with the intent of identifying areas of crop stress and weed competition. The scientific merits of these procedures will be disseminated through appropriate conference presentations and journal articles. In addition, Extension training will be provided through the mobile computing laboratory that was funded through Phase III of this grant program. The researchers have also filed a patent disclosure for this technology.