Download PDF version here
Project Title:  Development Of Maps For Site-Specific Application Of Cotton And Soybean Pesticides
Principal Investigators:  Daniel B. Reynolds and David R. Shaw

Technical Summary: 

Although major advances have occurred in both application technology and herbicides available, weed control is still one of the most expensive portions of row crop production.  Technologies that can minimize herbicide applications and maximize weed control will be beneficial both environmentally and economically.  Objectives of this project are to utilize data generated by the prior and related research projects to develop methods for accurately creating site‑specific weed maps and to compare site‑specific management programs to conventional production practices.  Results of this project will demonstrate the advantages and limitations of site‑specific weed management in cotton and soybean. 


1.      Develop a classification system to delineate cotton and soybean from weedy species.

2.      Develop a classification system to delineate crop species from monocot weedy species and dicot weedy species.

3.      Develop and evaluate the effectiveness of directed sampling maps based upon late season imagery from the previous cropping year.

4.      Compare the efficacy and economics of site‑specific weed control programs with conventional broadcast herbicide applications. 


1.      Develop a classification system to delineate cotton and soybean from weedy species.  Research from related research aimed at characterizing spectral properties of distinct species has shown that crop species can be classified with a high degree of certainty utilizing multi‑spectral data (Figure 1).  In a transgenic cropping system, where the crop has been genetically modified to tolerate topical applications of glyphosate (Roundup), there is generally less of a need for application of multiple active ingredients.  In most instances, glyphosate is the only herbicide applied topically.  In this scenario, identification of specific weed species is of less importance than where multiple active ingredients may be utilized.  If one could accurately assess an image for the presence of non‑crop vegetation then herbicide treatment maps could be developed based solely upon the presence or absence of weedy species.  This would facilitate a site‑specific application based upon this criteria thus applying glyphosate only where needed.  This would facilitate the control of weedy species while minimizing herbicide costs, and pesticide loading in the environment.  This can be achieved by interpolating maps from data taken on a grid; however, the time requirements are prohibitive thus nullifying any economic benefits to the technology. 

Data from previous and on‑going research efforts will be utilized to develop classification parameters to assess the ability to accurately delineate areas of weed infestations in the crop.  Spectral information from multiple years and growth stages will be utilized as training sets for the current cropping year.  Weed distribution maps generated from the image classification procedure will be compared to the ground truthed grid data.  Additionally, polygons identified as being weed infested will be ground truthed to validate their presence.  In order to achieve this, geo‑corrected classification maps will be downloaded to portable computing devices interfaced with DGPS receivers for directed sampling and verification. 

2.      Develop a classification system to delineate crop species from monocot weedy species and dicot weedy species.  Objective two is very similar to the first.  The primary difference is that in this system we will attempt to separate the weedy vegetation into classes based upon them being monocots or dicots.  This approach would be important if the producer is utilizing conventional varieties/hybrids where they wished to apply a single product for broadleaf weeds and another for grassy weeds.  Additionally, this would be important in a cotton system where bromoxynil (Buctril) transgenic varieties are being utilized.  Unlike the Roundup Ready system, where the herbicide used (Roundup) controls both monocots and dicots, Buctril controls only dicot species; thus an additional herbicide would be needed to control the monocot species present.  Data from related projects have also shown some promise in delineating differences between monocot and dicot weeds (Figure 2).  These data were generated on a limited number of species and further field testing is needed for validation.  Although the classification efficiency was not as good as that obtained for weed vs crop, it appears that it may be effective enough to produce treatment maps. 

The approach will be the same as that utilized in Objective 1 where data from previous and on‑going research efforts will be utilized to develop classification parameters to assess the ability to accurately delineate areas of monocot and dicot weed infestations in the crop.  Spectral information from multiple years and growth stages will be utilized as training sets for the current cropping year.  Weed distribution maps generated from the image classification procedure will be compared to the ground truthed grid data.  Additionally, polygons identified as being infested with monocots or dicots will be ground truthed to validate their presence.  In order to achieve this, geo‑corrected classification maps will be downloaded to portable computing devices interfaced with DGPS receivers for directed sampling and verification. 

3.      Develop and evaluate the effectiveness of directed sampling maps based upon late season imagery from the previous cropping year.  Previous research at MSU as well as by other investigators has shown that as weed age increases, classification accuracy increases.  In the case of weedy species, one generally needs to apply herbicides prior to them reaching 6 inches in height; otherwise control may not be achieved.  Additionally, weed competition can occur early‑season and may impact yields even when removed mid‑ to late‑season.  Thus weeds need to be identified and treated when they are very small and more difficult to classify.  Unlike insects, weeds are not mobile and their propagules may not be highly mobile either, thus the expansion of affected areas may occur slowly.  Since agronomically important weedy species propagate primarily via seed production, rhizomes and root‑stocks, they may become relatively localized.  These factors may contribute to the findings of other researchers who have clearly shown that weeds occur in highly aggregated distributions.  One would assume that areas infested in a given year would likely have subsequent infestations the following growth cycle due to "seed rain" where weed seeds are spread in an area of fairly close proximity of its parent, when left undisturbed.  Thus it is possible that a weed distribution map produced in the proceeding cropping year may accurately depict the current years weed composition and distribution. 

This objective will utilize past ground truthed data as well as past and current aerial imagery.  Late season imagery will be utilized to develop distribution maps for the current year and to develop directed scouting for the following year.  Maps generated from the previous year will be compared to weed distributions as assessed on a grid sampling pattern, to determine their accuracy. 

4.      Compare the efficacy and economics of site‑specific weed control programs with conventional broadcast herbicide applications.  In this objective, ground‑truthed weed density data will be used to create site‑specific treatment maps in each of two 20 acre production fields.  The fields will be evaluated on at least a 0.5‑A grid.  The 0.5‑A grids will be divided into large plots, approximately 2‑acres.  The 2 acre blocks will be randomized across each field, in a randomized complete block design, with two treatments and at least 4 replications. Treatment one represents a site specific herbicide treatment and treatment two is a total over‑spray treatment of the 2 acre block area.  For each treatment, weed species and density data will be utilized to develop treatment criteria by analysis with MSU‑HADDS.  The fields will be sprayed according to the prescribed treatment maps with a site‑specific sprayer and economics, accuracy, weed control, and cotton yield of site‑specific treatments will be compared to conventional oversprays.  Economic analysis will be conducted in conjunction with Dr. Steve Martin at the Delta Research & Extension Center, Stoneville. 

Justification & Literature Review: 

Because weeds are a major impediment in profitable crop production in the southern U.S., solutions to weed problems have historically been major emphases of federal and state research programs.  Herbicides account for over 60% of the pesticides used in row crops in the U.S.  As a result, a wealth of background information exists on the ecology and control of major weed problems.  Weed scientists have developed detailed control recommendations and knowledge‑based systems to aid growers in selecting the best treatment for their particular situation.  However, these recommendations have usually been made with the assumption that weeds are uniformly distributed across the unit to be treated, typically either a field or farm.  Numerous studies have demonstrated the fallacy of this assumption.  Weeds tend to be highly aggregated within fields (Marshall, 1988; Mortensen et al., 1990; Johnson et al., 1996).  In many instances, substantial portions of fields contain no weeds, while other areas have high populations (Wiles et al., 1992; Johnson et al., 1995).  Data from Thornton et al. (1990) indicate that this scenario can lead to poor decisions in weed management when based upon the threshold concept.  Highly clumped populations will often lead to overestimation of yield losses, since intraspecific competition becomes substantial within the weed population.  A second issue of concern when aggregated populations occur is accuracy of population estimates.  Wiles et al. (1992) found that at least 18 samples per ha were required to accurately estimate populations in a study of 17 fields.  Few producers or consultants are willing to devote time and resources for this intensity of sampling effort; overestimation usually occurs because weed infestations are typically higher at field edges where sampling normally occurs.  Therefore, aggregate populations present two significant barriers to effective implementation of integrated weed management strategies. 

Currently, an integrated approach to weed management fundamentally relies on a threshold‑based concept. Efforts in recent years have focused on development of computerized decision aids that recommend the most effective and economical treatments only when threshold weed populations are present.  This focuses on postemergence herbicides that are applied only when necessary, rather than prophylactic blanket applications of preemergence or postemergence herbicides.  The model MSU‑HADSS has been developed and refined for recommending the most appropriate postemergence herbicide (including none if weed populations are below threshold) based on species, populations, weed size, herbicide costs and efficacy, and environmental conditions.  The model again assumes that an accurate assessment has been made of the weed population, and that this population is uniformly dispersed throughout the field.  False assumptions on either part will lead to inaccuracies in yield loss estimates and in determining threshold populations necessitating herbicide applications.  Thus, research is needed to determine how to overcome these barriers to sound, threshold‑based decisions on weed control efforts. 

The only viable mechanism for addressing the problem of over‑application of herbicides, even using the computer‑generated recommendations, is through effective field scouting and variable‑rate applications.  However, as already acknowledged, field scouting is time‑consuming and has simply not proven cost‑effective.  Thus, an alternative must be found.  Remotely sensed images provide the opportunity to develop these population maps.  However, past experience with satellite images has indicated that weeds must be large and populations must be fairly homogeneous, which makes image utility in row‑crop production minimal.  Producers must be able to accurately detect populations when weeds are small for effective herbicide control and before significant crop yield loss.  With greater spatial and spectral resolution from a new suite of sensors, the opportunity may exist to detect weeds at a time suitable for control in a cost‑effective manner.  The related research projects funded by the Remote Sensing Technologies Center (RSTC) are focused on identifying unique spectral characteristics to enable species differentiation.  The RSTC project requires high resolution hyper‑spectral data.  This project will focus on the ability to user lower resolution multi‑spectral data to identify groups of like species (monocots vs dicots) and crops to enable a less refined "on‑off" site‑specific herbicide application system. 

Current Research: 

Summary of previous ASTA project

Continue validation and refinement of a cotton database on weed competitiveness and herbicide efficacy to integrate into the MSU‑HADSS program for an expert herbicide recommendation system for cotton.  Studies were conducted at four locations across Mississippi.  Treatments were arranged as a split‑split‑plot in a randomized complete block design with four replications.  Main plots were comprised of  Roundup Ready, BXN and a conventional cotton variety, sub‑plots were no‑preemergence or Cotoran 1.25 lbs ai/A PRE and sub‑sub‑plots consisted of an early postemergence (early‑POST) HADSS recommendation, a mid‑postemergence (mid‑post) HADSS recommendation, a weedy check, and a weed free check.  These treatments were evaluated for efficacy, crop safety, and yield.  Weed populations were quantified, recommendations were generated by HADSS and treatments were applied at the 2‑4 and 6‑8 leaf stages.  HADSS recommendations containing preemergence treatments provided better than 90% control.  Treatments void of a preemergence application gave at least 80% control except for a single recommendation in the BXN system. The yield from the weed free plots did not yield significantly more than the two herbicide recommendations.  Thus these data indicate that yields were optimized when utilizing the HADSS recommendations.  The use of this decision‑aid should facilitate the use of the most efficacious and economical treatment in Roundup Ready, BXN, and conventional cotton varieties. 

Develop scouting methods to accurately map weed locations and densities, and integrate this information into the MSU‑HADSS system for recommendations of variable postemergence herbicide applications. 

Three soybean fields of 7, 16, and 15 ha and were selected at the Black Belt Branch Experiment Station, Brooksville, MS. Preemergence applications of 85 g ai/ ha flumetsulam and 3.1 kg ai/ ha metolachlor in 1998 and 46 g ai/ ha flumetsulam and 698 g ai/  ha pendimethalin in 1999 were applied to reduce the overall grass infestation.  Sampling of each field occurred 8 weeks after planting (WAP) and 6 WAP, in 1998 and 1999, respectively. All weed species were counted on a 10‑m grid, using a 0.58 m2 quadrate.  Data were eliminated from the original dataset of weeds for each field to develop 40‑, 60‑, and 80‑m independent data sets.  Distribution and population maps were interpolated using an inverse distance weighted method.  Data were extracted from the interpolated maps at known coordinates so that the observed population and the predicted population could be compared.  The 10‑m grid served as a standard to which all others were compared.  Accuracy of the larger scales in predicting weed populations decreased with increasing weed populations in the fields, with Pearson correlation coefficients of  ‑0.81, ‑0.82, and 0.83, for the 40‑, 60‑, and 80‑m scales, respectively.  Results were subjected to analysis of variance and no significant differences were detected when results were compared on a per weed basis, except when weed populations were extremely high, generally exceeding 1000 plants ha‑1.  

Field experiments were established to evaluate multispectral imagery for discriminating weed‑infested and weed‑free soybean late in the growing season.  Canopy composition estimates for soybean and weeds were collected four times after canopy closure while soybean was in the vegetative to late‑senescence stages of development.  Weed canopy estimates ranged from 30 to 36% for plots infested with browntop millet, barnyardgrass, and large crabgrass.  Within two days of visual canopy estimates, aerial multispectral imagery data were collected in the green, red, and near‑infrared spectrums.  The red and near‑infrared spectral data collected for each plot were used to develop a normalized difference vegetation index (NDVI).  All spectral bands and NDVI were used as classification variables in discriminant classification procedures used to discriminate grass‑infested and weed‑free soybean.  In most cases, reflectance in the green, red, and NIR bands, and NDVI was higher in weed‑infested plots compared to weed‑free plots.  Weed presence was detected with at least 90% accuracy, regardless of soybean growth stage.  Discriminant analysis functions developed from one image were 81 to 90% accurate in detecting weed presence for other images when soybean growth stage did not differ substantially.  However, overall classification accuracy decreased to 47 to 75% when using discriminant functions developed when soybean was in vegetative or late‑senescent growth stages to test classification accuracy when soybean was in the opposing growth stage.  Multispectral imagery has potential for late‑season weed detection across a range of soybean growth stages as long as soybean growth stage does not differ substantially between images. 

Field experiments were also established to evaluate the potential for hyperspectral data to discriminate reflectance properties of pitted morningglory intermixed with soybean and weed‑free soybean at various growth stages in conventional till and no‑till plots containing no cover, rye, and hairy vetch cover crop residue.  With the use of a hand‑held spectroradiometer, hyperspectral reflectance data were collected between 350 and 2,500 nm.  Pitted morningglory plant size had more influence on discriminant capabilities than tillage and residue systems.  Across all tillage and residue systems, classification accuracy of soybean plus pitted morningglory and weed‑free soybean was 71% at the cotyledon to 3‑leaf pitted morningglory growth stage, 76 to 79% at the 2‑ to 5‑leaf stage, 87 to 93% at the 4‑ to 8‑leaf stage, and 95% at the 6‑ to 9‑leaf stage.  Classification accuracy within till and no‑till systems (pooled across residue systems) was 67 to 81% at the cotyledon to 3‑leaf pitted morningglory growth stage, and 83 to 97% at the 4‑ to 9‑leaf growth stage.  Within each tillage and residue system, overall classification accuracy was 83 to 100% when pitted morningglory growth stage was beyond the cotyledon to 3‑leaf growth stage and 60 to 86% at the cotyledon to 3‑leaf stage.  The versatility of the eight 50‑nm hyperspectral bands for discriminating soybean intermixed with pitted morningglory and weed‑free soybean at each pitted morningglory growth stage was tested by using a discriminant function developed for one location to test discriminant capabilities for the opposite location.  Overall classification accuracy across all tillage and residue systems was 48 to 51% at the cotyledon to 3‑leaf growth stage, and 53 to 72% when pitted morningglory was beyond the 3‑leaf growth stage.  Overall classification accuracy when testing the discriminant capabilities within each tillage and residue system was 36 to 69% in the different tillage and residue systems at the cotyledon to 3‑leaf pitted morningglory growth stage and 76 to 93% by the 2‑ to 5‑ leaf growth stage.  Ability to discriminate weed‑free soybean from soybean intermixed with pitted morningglory was influenced by amplitude differences in reflectance between soybean and pitted morningglory, with soybean having a higher degree of reflectance than pitted morningglory at the 2‑ to 4‑leaf pitted morningglory growth stage in the visible spectrum. Conversely, pitted morningglory had a higher degree of reflectance in the infrared portion of the spectrum. 

Integrate remote sensing techniques (satellite, aircraft) for development of weed mapping and resulting postemergence herbicide recommendations through MSU‑HADSS.  Weed populations of three soybean fields (B‑East, B‑South, B‑North), located at the Black Belt Branch Experiment Station, Brooksville, MS, were estimated in 1998 and 1999.  Sampling occurred July 8 9, 1998 (8 weeks after planting), and June 30 July 1, 1999 (6 WAP). An established 10‑m x 10‑m UTM grid coordinate system was used to divide the fields into 100‑m2 cells, with the sample point located in the center of each cell.  Optimal herbicide recommendations were obtained for each sample location within each field by subjecting the weed information to the Herbicide Application Decision Support System (HADSS). An average of the weed populations for the entire field was also subjected to HADSS to obtain an optimal recommendation for a broadcast application for comparison purposes.  Data from 1998 resulted in 25 and 15% of the field not requiring a herbicide treatment for the B‑North and B‑South when compared to the whole‑field recommendations to receive broadcast treatments.  However, B‑East received a "no treatment" recommendation for the whole‑field analysis.  This was attributed to the sicklepod population exceeding a level deemed economically controllable by HADSS.  However, when SSWM recommendations were generated, 49% of the field received this recommendation, while 51% resulted in a herbicide application as an economical choice.  In 1999, glyphosate‑resistant transgenic soybean was used, thereby increasing the POST herbicide treatment options available in HADSS.  Herbicide treatment recommendations resulted in 100, 56, and 91% of the total area requiring herbicide treatments for B‑East, B‑North, and B‑South, respectively.  Comparing the projected net returns for each field can develop a better estimate of the value of SSWM.  In 1998, data from the B‑East resulted in a projected net return increase of $21.63 ha‑1 over that of the broadcast application.  Estimated net return increased $5.42 ha‑1 at B‑North with simulated SSWM applications over broadcast applications and  $14.67 ha‑1 increase at B‑South.  Net returns for 1999 resulted in only a $0.32 ha‑1 increase by using SSWM for B‑East, but a $21.00 and $13.56 ha‑ 1 increase for B‑North and B‑South, respectively. The extra expenses of SSWM, such as sampling and technology costs, are not included in the net returns calculations and, when included, would reduce the difference between SSWM and conventional methods.  This research has demonstrated the potential value of SSWM from an economic standpoint; environmental benefits through reductions in herbicide applications are also apparent. 

Experiments were conducted to evaluate the use of spectral data in classifying different crop and weed species.  The experiment was designed as a randomized complete block with 4 replications.   Species were maintained free of other species and hyperspectral data were taken on two week intervals.  Although spectral data were taken from 350‑2500 nanometers (nm) discrete bandwiths were selected to correspond with available airborne sensors.  Bands selected were: Green (545‑555 nm); Red (670‑680 nm); and Near Infrared (835‑845 nm).  The following vegetation indices were computed: Normalized difference vegetation index (NDVI); Green normalized difference vegetation index (NDVIg); Global Environmental Monitoring Index; Near Infrared (NIR); Red vegetation Index (RVI); and Difference vegetation index (DVI).  The computed indices and band widths were evaluated with a linear discriminant analysis technique, to categorize the dependent variable.  Both the cross‑validation (leave one out testing) and resubstitution options were used.  As one would expect the resubstitution option resulted in greater overall classification than cross‑validation.  The cross‑validation technique resulted in 91, 63, 63, 100, 46, and 33% correct classification of velvetleaf, redroot pigweed, broadleaf signalgrass, cotton, johnsongrass, and corn, respectively.  Generally, broadleaf species are more easily differentiated from each other than grass species.  These preliminary data indicate that spectral data holds promising potential in discriminating among species. Pure populations of pitted morningglory, entireleaf morningglory, sicklepod, common cocklebur, and soybean were established.  Mixed weed vegetation, both with and without soybean and bare soil were also included in the study for comparison.  Data were collected using two separate portable field spectroradiometers with spectral ranges of 400 to 1050 nm (3@ FOV) and 300 to 2500 nm (8@ FOV), with sampling intervals of 1.4 and 1.5 nm, respectively.  Data sets were of considerable size and therefore subjected to four different data reduction methods.  Stepwise discriminant analysis, principal components, selected vegetative indices, and selection of specific wavelengths, 12 in total, which had been found in previous research to have particular associations with agricultural crops.  Sequential evaluation of bands resulted in a reduction to a range of 7 to 119, depending on data set, that were most useful in discriminant analysis. Bands ranging in the blue to green portions of the spectrum were most often selected as having the highest correlation.  This was also the case when the 12 specifically selected bands were subjected to stepwise discriminant analysis.  When selecting for vegetation indices with discriminatory power, the SAVI had the highest correlation with the 3@  FOV data.  When selecting indices using the 8@ FOV data the most power was given to the GNDVI and the RVI.  Each of the data sets were subjected to classificatory discriminant analysis using the variables selected from the stepwise discriminant analysis.  To determine the best method of data reduction, while maintaining classification ability, results were subjected to analysis of variance.  The best method for differentiating between observations was using wavelengths selected from stepwise discriminant analysis, with classification accuracy of >98 % for sicklepod, 100% for bare soil, >92% for entireleaf morningglory, >86% for pitted morningglory, >81% for soybean, and >92%, for three of the data sets.  Data collected in late August were limited due to heat stress of the instrument and results were not as successful as other data sets. 

Evaluate the use of site‑specific point‑injection sprayers for applying herbicides in a site‑specific manner according to herbicide treatment maps generated by MSU‑HADSS.  Weed distribution maps were generated on a 0.5‑A grid on two 20‑A fields at Brooksville, MS.  Data collected were subjected to analysis by MSU‑HADSS.  MSU‑HADSS recommended specific treatment regimes for each grid cell based on species composition and density.  Both the weed distribution and herbicide treatment maps were interpolated using inverse weighted distance to develop treat maps for use with the point‑injection sprayer.  Each of the 20‑A fields were divided into 6 sections.  Each of the sections were to either be treated with a broadcast treatment based on the field average or site‑specifically based on individual weed populations.  Numerous technical problems precluded the use of the site‑specific sprayer.  Although individual components of the "whole" system are available, they currently do not work easily with each other.  It appears that all software and hardware requirements have not been adequately interfaced.  Although we can show the effects of theoretical applications and theoretical savings we have not been able to actually achieve site‑specific application of multiple active ingredients.  We have made several software and hardware changes that hopefully will enable site‑specific applications of multiple active ingredients in 2002.


Back to Top

Back to Menu

Visit: MSU ||  || MAFES Intranet || Remote Sensing Technologies Center (RSTC)
For information about this web site, contact Wade Givens
Last Modified: 01/06/2004