Scott T. Drummond

Accuracy issues in electromagnetic induction sensing of soil electrical conductivity for precision agriculture

Soil apparent electrical conductivity (ECa) has been used as a surrogate measure for such soil properties as salinity, moisture content, topsoil depth (TD), and clay content. Measurements of ECa can be accomplished with commercially available sensors and can be used to efficiently and inexpensively develop the dense datasets desirable for describing within-field spatial variability in precision agriculture. The objective of this research was to investigate accuracy issues in the collection of soil ECa data. A mobile data acquisition system for ECa was developed using the Geonics EM38 1 sensor. The sensor was mounted on a wooden cart pulled behind an all-terrain vehicle, which also carried a GPS receiver and data collection computer. Tests showed that drift of the EM38 could be a significant fraction of within-field ECa variation. Use of a calibration transect to document and adjust for this drift was recommended. A procedure was described and tested to evaluate positional offset of the mobile EM38 data. Positional offset was due to both the distance from the sensor to the GPS antenna and the data acquisition system time lags. Sensitivity of ECa to variations in sensor operating speed and height was relatively minor. Procedures were developed to estimate TD on claypan soils from ECa measurements. Linear equations of an inverse or power function transformation of ECa provided the best estimates of TD. Collection of individual calibration datasets within each surveyed field was necessary for best results. Multiple measurements of ECa on a field were similar if they were obtained at the same time of the year. Whole-field maps of ECa-determined TD from multiple surveys were similar but not identical. There was a significant effect of soil moisture and temperature differences across www.elsevier.com:locate:compag

Statistical and Neural Methods for Site-Specific Yield Prediction

Understanding the relationships between yield and soil properties and topographic characteristics is of critical importance in precision agriculture. A necessary first step is to identify techniques to reliably quantify the relationships between soil and topographic characteristics and crop yield. Stepwise multiple linear regression (SMLR), projection pursuit regression (PPR), and several types of supervised feed-forward neural networks were investigated in an attempt to identify methods able to relate soil properties and grain yields on a point-by-point basis within ten individual site-years. To avoid overfitting, evaluations were based on predictive ability using a 5-fold cross-validation technique. The neural techniques consistently outperformed both SMLR and PPR and provided minimal prediction errors in every site-year. However, in site-years with relatively fewer observations and in site-years where a single, overriding factor was not apparent, the improvements achieved by neural networks over both SMLR and PPR were small. A second phase of the experiment involved estimation of crop yield across multiple site-years by including climatological data. The ten site-years of data were appended with climatological variables, and prediction errors were computed. The results showed that significant overfitting had occurred and indicated that a much larger number of climatologically unique site-years would be required in this type of analysis.

SIMULTANEOUS SOIL MOISTURE AND CONE INDEX MEASUREMENT

Soil compaction can restrict root growth and water infiltration, resulting in yield reduction. Maps of yield monitor data aid in visualization of variations in yield, without identifying underlying factors for these variations. Soil penetration resistance can help identify areas where soil physical characteristics are negatively impacting yield. However, penetration resistance is a function of soil moisture content and soil type as well as compaction. A standard penetrometer cone was modified to collect near-infrared reflectance and estimate moisture content. The instrument was tested in the laboratory on a selection of soil types with varying moisture tension levels using stepwise and continuous probe insertions. Soil moisture, dry bulk density, and clay content were significant variables in predicting soil cone index at the lower moisture tension level.

WAVELENGTH IDENTIFICATION AND DIFFUSE REFLECTANCE ESTIMATION FOR SURFACE AND PROFILE SOIL PROPERTIES

Optical diffuse reflectance spectroscopy (DRS) has been used to estimate soil physical and chemical properties, but much of the previous work has been limited to surface soils or to samples obtained from a restricted geographic area. Our objectives in this research were: (1) to assess the accuracy of DRS for estimating variation in several important surface and profile soil properties across a wide range of soils from the U.S. Corn Belt, and (2) to determine the wavelength ranges and/or specific wavelengths that should be included in a DRS soil property sensor. Soil cores were obtained to a 120 cm depth from ten fields, two each in Missouri, Illinois, Michigan, South Dakota, and Iowa. Cores were segmented by pedogenic horizon and samples (n = 165 for the surface soil horizon, n = 697 for all soil horizons in the profile) were analyzed for texture fractions, cations (calcium, magnesium, and potassium) and cation exchange capacity (CEC), pH, total and organic carbon, and total nitrogen using standard laboratory procedures. Spectra were obtained on sieved, air-dried soils from 350 to 2500 nm using a commercial three-detector spectrometer. Reflectance data were related to soil properties using partial least squares (PLS) regression and stepwise multiple linear regression (SMLR). Calibration accuracies varied among the different soil properties, but for a given soil property, similar accuracies were generally obtained with PLS and SMLR. The most accurate estimates, with R2 values above 0.8, were obtained for organic carbon, clay, CEC, and calcium. When data from each of the three spectrometer detector ranges were analyzed separately with PLS, the third detector range (1770 to 2500 nm) provided results similar to those obtained using the complete spectral range. Discrete wavelength models that described 90% or more of the variance described by a full model were obtained using eight or fewer wavelengths for the profile dataset and six or fewer wavelengths for the surface dataset. Several wavelengths and wavelength ranges common to models for multiple soil properties were identified: 2070 nm, 1870 to 1915 nm, and 2220 to 2410 nm. Because additional wavelengths important for individual soil properties were dispersed across the 1770 to 2500 nm spectral range, a full-spectrum sensing approach is recommended for simultaneous estimation of multiple soil properties. A discrete-waveband sensor could be practical for estimating one or two individual soil properties.

Application of neural networks: precision farming

Precision farming is a relatively new field of study whose goal is to improve cropping efficiency by variable application of crop treatments such as fertilizers, pesticides, etc. A deeper understanding of the functional relationship between yield, soil and site properties is of critical importance to precision farming. A number of feedforward neural network methods were investigated in an attempt to identify techniques able to functionally relate soil properties and crop yields on a point by point basis. Both training accuracy and generalization ability were evaluated for these previously reported neural techniques. Several techniques were able to provide a relatively high degree of accuracy, while retaining good generalization characteristics.

CALIBRATION METHODS FOR SOIL PROPERTY ESTIMATION USING REFLECTANCE SPECTROSCOPY

Optical diffuse reflectance sensing is a potential approach for rapid and reliable on-site estimation of soil properties. One issue with this sensing approach is whether additional calibration is necessary when the sensor is applied under conditions (e.g., soil types or soil moisture conditions) different from those used to generate an initial calibration, and if so, how many sample points are required in this additional calibration. In this study, these issues were addressed using data from ten fields in five states in the U.S. Corn Belt. Partial least squares (PLS) regression was used to develop calibrations between soil properties and reflectance spectra. Model evaluation was based on the ratio of standard deviation to RMS error (RPD), a statistic commonly used in spectral analysis. When sample data from the field where calibrations were to be applied (i.e., test field) were included in the calibration stage (full information calibration), RPD values of prediction models were increased by an average of 0.55 (from 1.08 to 1.63) compared with results from models not including data from the test field (calibration without field-specific information). Including some samples from the test field (hybrid calibration) generally increased RPD to 90% of that from full information calibration (average increase = 0.49) by using data from 8 to 20 soil cores, with little further improvement given additional data. Using test field points as a bias adjustment (two-stage calibration) increased RPD by an average of 0.29 with two to six sample points, a finding that was confirmed by Monte Carlo simulation. These results show the importance of including in a calibration set samples similar (i.e., obtained from the same or similar fields) to those in the test set. These similar samples could be included directly in the calibration or could be used to implement a post-calibration bias adjustment. Although results were more accurate with the recalibration approach, the bias adjustment approach was more efficient computationally and required less data. Thus, either might be preferred depending on specific circumstances.

DETERMINING YIELD MONITORING SYSTEM DELAY TIME WITH GEOSTATISTICAL AND DATA SEGMENTATION APPROACHES

In combine harvesting, knowledge of the delay time from cutting the crop to sensing the grain flow is required for accurate spatial location of grain yield data. Currently, either an assumed, fixed delay time is used or the delay time is determined by visual inspection of yield maps. Geostatistical and data segmentation methods were developed to estimate yield monitoring system delay time using objective criteria. The methods were validated with an ideal dataset and with elevation and soil electrical conductivity datasets having known delay times. When applied to yield and moisture content measurements collected with a commercial yield monitoring system, the methods successfully estimated delay time. In most cases, the results agreed (µ1 s) with results achieved using a visual method. Grain yield and grain moisture content exhibited different delay times at different locations within test fields. Thus, it may be appropriate to apply delay time corrections to homogeneous sub–field areas, instead of on a whole–field basis. Use of these new estimation methods could allow for more accurate and efficient processing of yield monitor data.

Yield Editor 2.0: Software for Automated Removal of Yield Map Errors

Yield maps provide important information for developing and evaluating precision management strategies. The high-quality yield maps needed for decision-making require screening raw yield monitor datasets for errors and removing them before maps are made. To facilitate this process, we developed the Yield Editor interactive software which has been widely used by producers, consultants and researchers. Some of the most difficult and time consuming issues involved in cleaning yield maps include determination of combine delay times, and the removal of “overlapped” data, especially near end rows. Our new Yield Editor 2.0 automates these and other tasks, significantly increasing the reliability and reducing the difficulty of creating accurate yield maps. This paper describes this new software, with emphasis on the Automated Yield Cleaning Expert (AYCE) module. Application of Yield Editor 2.0 is illustrated through comparison of automated AYCE cleaning to the interactive approach available in Yield Editor 1.x. On a test set of fifty grain yield maps, AYCE cleaning was not significantly different than interactive cleaning by an expert user when examining field mean yield, yield standard deviation, and number of yield observations remaining after cleaning. Yield Editor 2.0 provides greatly improved efficiency and equivalent accuracy compared to the interactive methods available in Yield Editor 1.x.

VALIDATION TESTING OF A SOIL MACRONUTRIENT SENSING SYSTEM

Rapid on-site measurements of soil macronutrients, i.e., nitrogen (N), phosphorus (P), and potassium (K), are needed for site-specific crop management, where fertilizer nutrient application rates are adjusted spatially based on local requirements. This study reports on validation testing of a previously developed ion-selective electrode (ISE) based soil macronutrient sensing system using 36 soil samples from a single site, the Northern Illinois Agronomy Research Center (NIARC), and previously developed calibration models. Objectives were to (1) validate calibration models with a new array of membranes and electrodes, and (2) evaluate the ability of the system to estimate variations in soil NO3-N, P, and K within a single test site. Soil extract samples were obtained using the Kelowna extractant. Electrode responses were measured with five ISEs for each of NO3-N, P, and K and were normalized using the baseline correction and two-point normalization methods developed in our previous work. The array of ISEs fabricated with new membranes and cobalt rod, in conjunction with the previously developed normalization methods and calibration models, accurately estimated NO3-N, P, and K in solution without need to recalibrate the ISE system through standard laboratory analysis of soil samples from the new test site. ISE-measured NO3-N, P, and K concentrations in Kelowna-based soil extracts were similar to those determined by standard instruments, validating the ability of the system to identify within-field macronutrient differences. The use of a calibration factor to adjust ISE measurements for the difference in extraction efficiency between Kelowna and standard extractants resulted in a slope near unity between soil NO3-N, P, and K concentrations determined by ISEs and standard methods. However, a relatively large offset in soil P concentration between calibrated ISEs and standard methods will require further investigation to identify the cause. This study showed that it was possible to transfer existing calibration equations to new membranes and electrodes through application of the baseline correction and two-point normalization methods and an adjustment for differences in extraction efficiency. This finding enhances the applicability of the ISE-based soil macronutrient sensing system and methodology for rapid soil analysis.

Automated Yield Map Delay Identification Using Phase Correlation Methodology

Crop yield data are a key component of precision agriculture and are critical for both development and evaluation of precision management strategies. Ideally, software that generates grain yield maps from raw yield monitor data should automatically correct errors associated with machine and operating characteristics. Perhaps the most basic correction required is to properly compensate for the time lag (or position lag) between the cutting of the crop from the field and the grain flow measurement by the flow sensor in the combine. Past research has suggested several approaches to automatically determine delay time, but for various reasons these have not been implemented in mapping software. In this article, we present a new, computationally efficient method that can accurately determine delay time for individual fields using the image processing method of phase correlation. The phase correlation delay identification (PCDI) method was evaluated using a number of yield maps with varying degrees of harvest complexity, and results were compared to a geostatistical method. The PCDI method produced accurate estimates of delay time in approximately 90% of test datasets and provided a way to evaluate the reliability of the estimate. Additionally, the PCDI method was more computationally efficient than previous methods. Results of this study will increase the feasibility of including automatic delay time compensation in yield mapping software.