Artificial Neural Networks in Agricultural Machinery Applications

Marvin L. Stone Glenn A. Kranzler

Associate Professor Professor

Biosystems and Agricultural Engineering Department

Oklahoma State University

Summary:

Control and management of agricultural machinery offers many opportunities for application of general purpose empirical models. The nature of agricultural machines creates the need for modeling systems that are robust, noise tolerant, adaptable for multiple uses, and are extensible. Artificial Neural Networks (ANNs) have these characteristics and are attractive for use in control and modeling in agricultural machinery.

This paper reviews five cases of the application of ANNs in agricultural machinery and discusses the development and performance of the models in each case. The capabilities of ANNs with respect to configuration, adaptation, noise tolerance, and training are addressed. In addition, the use of ANN models in embedded systems is discussed.

Introduction

Artificial neural network (ANN) based models have been explored for use in various agricultural machinery applications. The typical application has been based on Multiple Input / Single Output ANNs which can be used to model linear and non-linear surfaces (Govind and Ramamoorthy, 1991). These types of models may be effective where response surface modeling has been used in the past. Potential applications of this type of ANN in machinery applications include predicting:

· Spatial yield response in fields in precision farming applications (Drummond, 1995).

· Machine performance, e.g. combine harvesters (Hall, 1992).

· Plant characteristics from sensor signals (Stone, 1994a; Zheng et al., 1994).

· Temporal dynamics in control systems (Altendorf, 1993).

A major impetus for examining application of ANNs in agricultural machinery has been the rapid development of cost-effective control systems that are available for use in agricultural machinery. Electronic systems have increased in reliability and capability and are available at lower costs than ever before. At the same time, sensor technologies have been experiencing similar improvements. Examples in the sensor area are integrated photodiode sensors and amplifiers, micro-machined accelerometers and pressure transducers, and Hall Effect sensors. Costs for components in sensor technology have been reduced by a factor of ten in some cases, allowing consideration of sensors where it would have been impractical in the past. In addition to the improvement in the cost-to-performance ratio for sensors and computers, the environment for application of electronics technology in agricultural machinery is improving. Standardization of communications between electronic components for agricultural machinery is underway (Stone, 1994b). The proposed standardized communications technology allows sharing of sensor information among tractor and implement systems, as well as within other types of agricultural machinery and with off-board computers. The possibility exists for addition of sensor-based controls in a modular and cost-effective fashion.

The availability of cost-effective sensors, computers, and a conducive environment for addition of the components, poses the natural problem of the need for effective models to allow control to be exercised. The current strong interest in precision farming is an example of this problem. A premise in map-based precision farming is that development of maps of soil fertility characteristics combined with maps of previous yields may be used to forecast the necessary fertilization for the current year. Adequate models are needed to allow map-based precision farming to succeed. ANNs are a natural solution to this type of empirical model. Theoretical or analytical models will require more time and people than are available, while empirical models may be adapted to a particular situation, for example, a particular grower's fields in the case of precision farming.

The popularity and interest in the use of ANNs is also driven somewhat by the availability of effective tools for developing the networks. Many convenient tools are available which allow neural network models to be fitted to empirical datasets. One can contrast the availability of ANN software development packages with the available non-linear regression packages. The neural network option looks attractive from the perspective of the variety of development packages available.

Modeling of equipment performance to allow control is another area where ANNs have strong potential for agricultural equipment applications. Combine harvesters are a good example. These machines are used to process many different crops (e.g., mung beans to canola) which have large variations in characteristics (e.g., dry and brittle to wet and weed-contaminated) within a particular crop as well as large variations among crops. Models would be useful that could predict performance and allow the harvester to be "tuned" for particular crops and crop variations. ANNs have been used to create general performance models that could be "trained" during operation to deal with particular variations. This approach would mitigate the lack of comprehensive analytical models and the cost of people to create them.

ANNs have been credited with more noise tolerance than conventional regression-based empirical modeling. A case is presented later with testing to support this claim. In addition to noise tolerance, a related characteristic is the ability to function adequately in the presence of missing data. Both characteristics in ANNs stem from redundancy that can be built into the model as well as the nature of an ANN to seek a minimum error for each case.

ANNs have undesirable characteristics that limit their application in agricultural machinery applications. Those characteristics include computationally intensive training, network weighting based on training cases, non-closed form implementations, and lack of the traditional understanding that exists for regression approaches. Computational intensive training generally limits training to a powerful computer. One of the models reported below required between 6 and 15 hours of training time on a Cray Y-MAP (Zheng et al., 1994). In the author's experience, this time is large but not unusual. The ability to write self-adaptive ANNs is severely inhibited by this constraint. A field computer would require cost-prohibitive speed and memory capacity for self-adaption. The issue of more heavily weighting network output to follow the cases that occur more frequently in training data can be seen as a disadvantage. The creator of an ANN should deal with this problem during the training of the ANN to assure the frequency of cases in the training data have the intended frequency distribution. Non-closed form implementation of ANNs results in an inefficiency when implementing the feed-forward ANN model. The typical implementation requires a summation to be executed as shown in equation 1. This is equivalent to summation of the terms in a regression equation, except that here the function normally has more terms than the typical regression equation. An advantage over the typical regression equation is that the function f , which is typically non-linear, may be implemented in an efficient way and is used repeatedly in calculation of the sum (Stone, 1994a). A non-linear regression equation may not have uniform non-linear elements and will require special coding to achieve computational efficiency. The computational efficiency becomes an issue in embedded systems where processor speed and memory space come at premium prices.

(1)

Where:

is the current output state of the j-th processing element in layer s.

is the weight on the connection joining the i-th element in layer s-1 to the j-th element in layer s.

f is traditionally a hyperbolic tangent, sigmoid, or sine function.

The traditional understanding that exists with regression approaches has not yet developed for ANN approaches. ANN techniques are relatively new, with the earliest reported ANNs in the mid 1950's and significant recognition of the potential for ANNs in 1982 (NeuralWare, 1991). General purpose tools have become available for ANN development (e.g., NeuralWare, 1991; Baffes et al.,1989) and many citations now exist in the literature in many fields. Confidence in ANN technology will develop appropriately as experience is gained in applications.

Five neural network applications are reviewed below. The cases were selected from the literature to allow evaluation of ANN technique in agricultural machinery-related applications. The current applications of ANNs in agricultural machinery are diverse and not numerous and the review here must focus with some detail on each of the available cases. Each of the studies selected examines different aspects of the application of ANNs. The cases that are reviewed are listed in Table 1, along with a description of the application and the major developments in each of the works. Table 2 summarizes the network types and development methods used in each of the studies.

Table 1. Studies selected for review

Table 2. Software and method used in reviewed studies

1. NeuralWare, Inc., Pittsburg, PA. 3. The MathWorks, Natic MA.

2. California Scientific Software, Nevada City, CA. 4. NASA, Baffes et al., 1989

Combine Harvester Control

J. W. Hall (1992) examined the applications of ANNs for the purpose of controlling a combine harvester, as a part of his Ph.D. dissertation. Hall thoroughly reviewed the characteristics of neural networks and performed testing to demonstrate the desirable characteristics of neural networks. Summarized below are the results of the general purpose testing that he performed on neural network models.

Efficacy in modeling non-linear functions

Govind and Ramamoorthy (1991) compared Volterra-Weiner non-linear modeling to ANNs. They found that in most cases, ANNs performed better than Volterra-Weiner with highly non-linear functions. Hall modeled z = sin(xy) with a 2-10-1 network, with minimum RMS errors of 0.148 and a 4-6-3-1 with minimum RMS error of 0.070. and z = exp-(x2+y2) with a 4-6-2-1, with a minimum RMS error of 0.019. He noted that though additional hidden layers increase an ANNs ability to model non-linearity, it also reduces the likelihood of convergence.

Noise immunity

Hall conducted an experiment on y = x2 with x between 0 and 1 to evaluate noise immunity. Outliers were added at levels of between 0.5 and 10% and simple 1-2-1 networks were tested against quadratic and cubic regression analysis. Regression analysis were significantly poorer (statistically) than the neural network models. With 30-point data sets, average RMS error for neural network based models was 0.0535, compared to 0.1218 and 0.1342 for quadratic and cubic regression respectively. In addition, a 3-9-1 configuration of an ANN was compared to a quadratic model with interaction terms for grain loss in a combine harvester related to fan speed, chaffer opening, and sieve opening. The computed response from the model was generated and errors of 25% and 50% were added to 5 to 10% of the data points. A quadratic regression analysis was performed and a model fit and compared with the neural network based model. The regression model was found to be more noise-tolerant with low levels of noise. The ANN based model was more noise tolerant at high noise levels.

Extending a model constructed originally with sparse data

ANNs are trained in a manner that allows new data to be added to a training dataset and for the ANN to be retrained starting with the previous model. Generally, training can be done much more quickly than the original training, and the effect is incremental learning or improvement of the model. A related characteristic is that ANNs tend to reproduce the cases provided in training and connect the available surface response points with a "minimum energy" surface. Hall demonstrated the later characteristic and compared the performance to a regression approach. Hall contrasted the nature of regression to place the function embedded in the regression between the available points with the ANN tendency to place a minimum energy surface between the points.

Use of latent parameters to allow future extension of a network

Latent input variables may be included in an ANN model to allow additional variables to be included in a training dataset in the future. The latent inputs are initially set to zero. This approach allows the network architecture to be set initially without requiring changes as new variables are added. Hall demonstrated this technique with a 12-8-1 network where the initial training data had only two inputs modeling z = x² + y². He compared this network with a 2-8-1 without latent inputs and found no difference in convergence and accuracy of the networks.

Use of sub-models inheritance

Binary inputs to an ANN can be used to allow the network to switch between one independent model and another. Hall trained a network to model z = 1 + xy, z = 1 - sin(2x) + y, and z = x² + y² with a 5-20-1 ANN. Two inputs were used for x and y, three inputs to select the particular model and one output for y. No improvements in training time were found when training each sub-model separately compared to combining all of the training data and training all at once. Two important observations were made. First, the total number of hidden nodes should be equal to the sum of the hidden nodes required for the individual sub-models. When fewer hidden nodes are used, dependencies or interactions between the models occur. Second when sparse training data are used for any particular sub-model, the network tends to inherit features from the other models where more data sets are provided. This later characteristic has been interpreted as "inheritance" and also as the "ability to generalize". Hall explored the conditions required for a sub-model, z = (x-1)² + y² to inherit from a parent z = x² + y². The parent was trained with 250 cases and the sub-model with 5 and 25 cases. The sub-model did reflect the training data in the locations where the data were provided and reflected the parent model elsewhere. The submodel improved as more training cases were provided. Hall also examined the relationship between the number of hidden nodes (with a single hidden layer) and the parent dependency of a sub-model. The testing was done where the sub-model had sparse data (5 cases with 250 cases for the parent). An optimum number of hidden nodes was found, based on a judgement of the degree of independence required.

Combine performance prediction

Hall took advantage of the ANN characteristics above to form a model to predict grain loss and quality produced by a combine harvester operating in wheat. He used a 15-6-4-1 configuration, as shown in Figure 1, to model each of five variables; two describing quality of the harvested grain and three describing efficiency of the harvester. The data used to develop the models were gathered from various sources including the literature, from expert combine harvester operators, and field experiments designed for that purpose. Of the model inputs, two parameters were determined by hand. The first was whether the crop was irrigated. The second was a threshing parameter which was based on a judgement of the difficulty in removing the grain from the crop. Two "specific condition" inputs were provided and used to allow the model to be calibrated to specific local conditions. Field experiments were run to identify the effects of variations in sieve opening, chaffer opening, and fan speed under three different growing locations.

Hall used one third of the data in the training cases for testing and two thirds of the data for training. The errors between training and testing were used to assure the training was not overfitting the data. Missing data were dealt with by setting the particular missing value to the mean value for that parameter. The experimental data consisted of 16 cases from the 3 locations. Testing was done to determine the number of cases required to retrain the model for a new location. In general, five to ten cases had to be added to the training data set to allow the model to perform as well as it did with all of the new cases added. Normalized RMS error was reported for each of the parameters and was within 0.1 to 0.2 ( parameter ranges were 0 to 1) for models where the specific condition inputs were used.

Figure 1. Combine Performance Model - Hall 1992

Modeling Variability in Fields

Drummond et al. compared the performance of three different types of multivariate modeling techniques for use in predicting crop yield. Soil fertility was sampled on a 30-m grid and top soil depth was measured with a finer resolution using a soil conductivity meter. Yield was measured for two crops; corn, grown in the field in 1993, and soybeans grown in the field in 1994. Yield was measured at one-second intervals during harvest on the combine harvester with a yield monitor and position measured with a GPS unit. The spatial resolution of the combine data were not reported, but is estimated here at approximately 2 m in the direction of travel and 6 m across the direction of travel. The instantaneous yield was corrected in position for delay through the combine harvester. Both the yield data and the soils data were kriged to a 10 m grid. A 25-ha field was sampled, resulting in a 2576-point data set. The data were randomly divided into training and testing data sets for the neural network development, but the r2 results are reported for the combined dataset. The network geometry used is shown in Figure 2. Experimentation with other geometries was not reported.

Figure 2. Yield Prediction Model - Drummond et al. 1995

The authors reported that the training did not result in overfitting, based on comparisons of the results on the training and testing data sets. Further, additional training in the 1993 data set may have resulted in higher r2 but was not done to retain consistency with the 1994 data set. Table 3, taken from Drummond et al. (1995) compares the performance of the different modeling techniques.

The authors concluded that weather variations were potentially the major cause of un-explained variability in the modeling. In addition, they also concluded that some further un-explained variability may have been due to non-weather related factors not included in the models. Measurements of soil nitrogen levels or nitrogen application was not reported in the study.

Table 3. Goodness-of-fit for yield prediction for various modeling techniques (Drummond et al., 1995)

Weed Detection in Sprayers

An ANN was developed to allow color patterns to be recognized in an agricultural weed sprayer application by Stone, (1994a). The structure of the most successful network tested is shown in Figure 3.

Figure 3. Plant Color Recognition Model - Stone 1994a

Figure 4 presents a schematic of the sensor and spray nozzle element component of the sprayer. The complete sprayer consisted of many of the sensor-nozzle elements placed in parallel on a single spray boom.

Figure 4. Sprayer Sensor and Nozzle Element.

A sensor was fabricated to detect color on the surface of the ground in a 7.5 by 50-cm wide image. Three color bands; green, red, and near infra-red were sensed. The signals from the sensor were digitized with a 68HC11 based controller using the on-chip 8-bit A/D converter. The 68HC11 based computer was also used to activate a solid-state switch that energized a solenoid valve in the spray nozzle. The intent of control in the system was to sense the presence of a weed by color and to activate the nozzle to spray the plant at the point in time that the plant was under the nozzle. A time budget is shown in the figure. If computing time plus the time required for the fluid to reach the ground once it emerges from the nozzle was insignificant, the sensor and nozzle could be located together. The 0.25 second time period between when the fluid emerges from the nozzle and when it reaches the ground cannot be changed. The configuration of the system places a practical limit on computing time. For a sprayer traveling in the field at 3 m/s, a typical ground speed, separation between sensor and spray nozzle must be 1.1 meters. This physical separation is near the maximum limit practical for the machine.

Agricultural sprayers based on optical sensing and control of spray nozzle activation currently exist on the market. The current designs rely on look-up table based models. This approach limits the number of inputs that can be practically used in the controller. A look-up table with three or more variables and with 8 bit precision will not fit conveniently in a low-cost micro-controller memory. An alternative to a look-up table is to encode the necessary response into an equation. Determination of a simple equation to model the problem is not a simple task.

Many potential interferences exist in detecting the plant, including: amount of target plant in the image, light level, dead plant matter, many soil colors, and variation in the color of the target plant. The inteferences result in an unusual map of sensor response based on color inputs. A non-linear model of some type is necessary. An ANN appeared to be a suitable model for the problem.

Training data were created by exposing the sensor to many different conditions intended to span the possible conditions that would be seen in actual application of the system. Soils of different colors were collected from six locations in Oklahoma. The soils were exposed to the sensor, dry and wet. In addition, various percentages of plant cover including 0%, 10%, and 100% were placed on the soils. In addition, the system was tested under various natural lighting conditions from heavy overcast to bright. Early testing revealed that in-door conditions could not easily be made to model out-door lighting. All combinations of the input conditions were tested resulting in nearly 80 sets of training conditions. The tests were repeated with similar conditions resulting in nearly 80 sets of test conditions that could be used to determine the performance of the system. Lighting, and plant placement could not be repeated exactly, resulting in significant variations between the test and training data.

Neural networks with one and two hidden layers were tested with different numbers of nodes in each layer. Table 1 presents results of training different configurations. The table presents the performance of the model after training on the training data and evaluation of errors that were found comparing model predictions of the test data. Two types of errors were computed to evaluate performance of the model. they were percentage of tests where the plant was present and detected (Plant % correct), and the percentage of tests where the plant was not present and not detected (No Plant % correct). Many more training iterations were performed, but are not shown. The table presents only tests where the "% correct when the plant was present" was maximized. For the current application of the sprayer, it is much more important to assure a plant has been sprayed than to avoid spraying when unnecessary. Current sprayer designs operate continuously and would have values of 0 and 100% for the "No Plant % Correct" and the "Plant % Correct" performance measures.

Table 1. Performance of network configurations

Nodes in Hidden Layer 1	Nodes in Hidden Layer 2	No Plant % Correct	Plant % Correct	Epochs
3	2	75	90	85000
3	3	80	92.5	60000
4	4	70	92.5	55000
5	5	70	92.5	55000
6	6	70	92.5	50000
7	7	65	92.5	70000
8	8	75	90	55000
3	0	45	100	20000
4	0	45	97.5	20000
5	0	50	100	20000
6	0	50	100	30000
7	0	40	100	25000
10	0	92.5	70	20000

Models with a single hidden layer in general were able to detect plants but were not as effective at rejecting situations where no plant was present. The training "epochs," the number of cases for which the network was optimized for ranged between 20,000 and 85,000. As expected, the number of epochs required to optimize the single hidden layer model was less than the two layer models. Training was done on a SUN IPX workstation using NeuralWare's NeuralWorks Professional II/Plus neural network development package. Training required 2 to 5 minutes for each configuration.

The model described in the second entry in Table 1 with three nodes in each of two hidden layers was selected for use in the prototype sprayer. More complex models did not improve the accuracy, and the single layer models were judged to have too large an error when no plant was present. The less complex model also allowed faster executing and smaller code to be used in the embedded application.

The resulting model was coded in C, compiled, and placed in the micro-controller. Two approaches were used to develop the code for the application. NeuralWare's developers package, DPACK, was used to automatically convert the network description into C. Additionally, the network was hand coded in C using equation 1. Table 2 presents code size and execution times for different optimizations of the embedded code.

Several techniques were tested to reduce the execution time of the code. The hand-coded version of the model was converted to a look-up table rather than the C function htan(x). The look-up table increased code size but reduced execution time by more than 50%. An alternative activation function, f(x) = 1/(1-x) was also tested and compiled in a floating-point form. The model had to be retrained using the alternative activation function with the same results as the originally selected activation function, f(x) = htan(x). The look-up table based model performed better than using f(x) = 1/(1-x) coded in floating point. Finally, the whole implementation of the model in C was coded in integer arithmetic. Some components of the calculation required double precision integers to retain accuracy. The resulting code produced an output in 0.07 seconds, after supplying inputs. This computational speed permits the time budget presented in figure 4 to be met and allows a feasible geometry for the physical components of the system.

Table 2. Execution time and code size for the production model

Model Description	Arithmetic	Activation Function	Code Length (Bytes)	Execution Time (s)
DPACK1 generated	Floating Point	Floating Point, f(x) = htan(x)	36K	-
Hand Coded	Floating Point	Floating Point, f(x) = htan(x)	3.5K	0.3
Hand Coded	Floating Point	Floating Point, f(x) = x/(1-x)	3.5K	0.15
Hand Coded	Floating Point	Look-Up Table, f(x) = htan(x)	3.8K	0.13
Hand Coded	Integer	Look-Up Table, f(x) = htan(x)	3.5K	0.07

1. Proprietary NeuralWare neural network code generator.

Some degradation in the accuracy in detection of plants was expected when the model was converted to look-up tables and integerized. Testing of the resulting models on the original test data revealed no significant difference between the integerized model and the original floating point model.

A prototype sprayer using nozzle elements based on the design was tested in the field. Though performance of the prototype was consistent with initial testing of the model, field tests revealed several unexpected operating limitations. The sensitivity of the prototype detectors was greatly diminished under low light conditions. Training of the model was not done under light conditions as low as those experienced in the field. In addition, during dawn and dusk periods, the color of the natural light is shifted toward red. Both conditions resulted in reduced sensitivity.

Blueberry Bush Pruning

Zheng and Rohrbach (1994) reported the development and testing of an ANN which would process ultra-sonic distance measurements to determine plant position. They designed an array of 6 ultrasonic range finder devices which were focused towards the center of a target bluberry plant as depicted in Figure 4.

Figure 4. Bush Ultra-Sonic Sensor Arrangement - Zheng and Rohrbach - 1994

The purpose of the resulting system was to position a trimming apparatus to trim branches from the blueberry plant. The ultra-sonic transducers were set to sense the closest branch intercepted by the sound beam for a particular sampling. A 6-20-1 configuration of an ANN was trained to use the distance measurements from the ultrasonic transducers and to predict the center position of the target plant, as shown in Figure 5.

Figure 5. Bush Centroid Prediction - Zheng and Rohrbach - 1994

A total of 343 cases were collected and divided, with 75% of the cases used for training and 25% used to test and compute error for a combined model. Of that total, 269 of the cases were taken with a single plant stem target and 74 cases with multiple stem targets. The ANN was trained until the test dataset MSE was minimized, which required slightly more than 300,000 epochs (number of times cases were individually presented to the network.). A second network configuration consisting of three 6-20-1 networks was also tested. One of the networks was used only to determine distance for cases with a single stem, another to determine distance for cases with multiple stems, and a third network was used to classify cases as single or multiple stems. The strategy is depicted in Figure 6.

Figure 6. Modular network design - Zheng and Rohrbach - 1994

This configuration of the network required less than half the training time, while producing slightly less error in the predicted position. Training time was an issue with the software and problem combination presented here, taking 4.3 hours on a DEC 5000/25 computer. The authors estimated the training could have been shortened to slightly over an hour on a Cray Y-MP super-computer. Standard deviations of the position errors produced by both configurations of networks were approximately 0.008 m and were distributed with a central tendency.

Weed Identification

Zhang et al. (1994) reported the use of ANNs to process color images of weeds in a winter-wheat environment with the objective of being able to distinguish between weeds and other components of the image. They were particularly interested in detecting weeds with reddish stems. They initially attempted to apply a 6-10-5 network to process 6 spectral indices and classify the inputs into five classes; weed stem, leaf, soil, cracks and shadows, and stones and bright spots. They were unsuccessful in training the network to acceptably classify the image. A second configuration, a 48-24-5 network configuration as shown in Figure 7 was tested.

Figure 7. Image Classification - Zhang et al. - 1994

The input spectral indices were coded as 8-bit values with one input for each bit of the index. This resulted in 6 indices x 8 bits or 48 inputs which had 1 or 0 values. Six hundred pixels taken from three images were pre-classified and used to form training and testing data sets. The trained network was then tested on six images, three originally containing the training data and three additional images.

The ANN was compared with a discriminant analysis approach to classifying pixels in the images. The ANN approach was not successful in classifying images of weed varieties different from the wild buckwheat images of the training data. The ANN did perform well on images similar to the training data. Overall error rates where test data sets were different from training data sets were slightly under 40%, while overall error rates for discriminant analysis were slightly less than 30%. In this study, ANNs did not perform better than a conventional approach. The authors did not report efforts to optimize the ANN to improve performance.

Summary and Conclusion

Five cases of the application of ANNs in agricultural machinery were reviewed. The application and performance of the models in each case were discussed. Some general conclusions can be drawn from the applications.

1. Artificial Neural Networks are an effective alternative to non-linear regression analysis in fitting surfaces. This conclusion is demonstrated in the work by Drummond et al., 1995.

2. Artificial Neural Networks can be used in embedded applications, but special techniques are needed to allow effective speed and memory utilization as shown in Stone, 1994a.

3. Training time may be a limiting factor in applying ANNs, as demonstrated by Zheng et al., 1994. Training may require hours on fast workstations even for relatively simple networks. Modularization of the model or use of an existing model may reduce the training time as shown by Hall, 1992 and Zheng et al., 1994.

4. An ANN model may be adapted to fit local conditions with the addition of relatively few cases of training data as shown by Hall, 1992.

5. In cases where outliers may exist in data, ANNs appear to be less sensitive than conventional regression analysis, as shown by Hall, 1992.

6. Classification using an ANN is not always successful and may not be better than conventional approaches, as in Zhang et al. 1994.

Control and management of agricultural machinery offers many opportunities for application of general purpose empirical models. The nature of agricultural machines creates the need for modeling systems that are robust, noise tolerant, adaptable for multiple uses, and are extensible. Artificial Neural Networks (ANNs) have these characteristics and are being examined for use in control and modeling in agricultural machinery.

References

Altendorf, C. A. 1993. Estimating Soil Water Content using Soil Temperatures and a Neural Network. Unpublished Ph.D. Dissertation. Department of Biosystems and Agricultural Engineering. Oklahoma State University, Stillwater Oklahoma.

Baffes, P. T., R. O. Shelton, and T. A. Phillips. 1989. NETS User's Guide, Version 3.0. Software Technology Branch, Lyndon B. Johnson Space Center, Houston, TX.

Drummond, S. T., K. A. Suddeth, and S. J. Birrell. 1995. Analysis and Correlation Methods for Spatial Data. ASAE Paper 95-1335. ASAE, 2950 Niles Rd., St. Joseph, MI 49085-9659.

Govind, G. and P. A. Ramamoorthy. 1991. Multi-Layer Neural Networks for Arbitrary Approximation: an Explaination and Simulations, Intelligent Engineering Systems through Artificial Neural Networks. pp 589-594, ASME Press, New York.

Hall, J. W. 1992. Emulating Human Process Control Functions with Neural Networks. Unpublished Ph.D. Dissertation. Department of Mechanical Engineering. University of Illinois, Urbana Illinois.

NeuralWare. 1991. Neural Computing. NeuralWare, Inc. Pittsburg, PA.

Stone, M. L. 1994a. Embedded Neural Networks in Real Time Controls. SAE Paper 941067. 45th Annual Earthmoving Industry Conference. SAE, Warrendale PA.

Stone, M. L. 1994b. High Speed Networking in Construction and Agricultural Equipment. SAE Paper 941662. 1994 Symposium on Future Transportation Electronics: Multiplexing and In-Vehicle Networking. SAE, Warrendale PA.

Zhang, N., Y. Yang, and M. El-Faki. 1994. Neural-Network Application in Weed Identification using Color Digital Images. ASAE Paper No. 94-3511. ASAE, 2950 Niles Rd., St. Joseph, MI 49085-9659.

Zheng, D. and R. P. Rohrbach. 1994. Neural Networks for Ultrasonic Position Control during Blueberry Pruning. ASAE Paper No. 94-1058. ASAE, 2950 Niles Rd., St. Joseph, MI 49085-9659.