to Chun-Chieh Yang's Study Room
and Read My Publications
PUBLICATIONS THESIS PROCEEDINGS PRESENTATIONS
PUBLICATIONS

Yang, C.-C., K. Chao, Y. R. Chen, M. S. Kim and H. L. Early. 2006. Simple region of interest analysis for systemically diseased chicken identification using multispectral imaging. Transactions of the ASAE, 49(1): 245-257.

            A simple multispectral differentiation method for the identification of systemically diseased chickens was developed and demonstrated. Color differences between wholesome and systemically diseased chickens were used to select interference filters at 488 nm, 540 nm, 580 nm, and 610 nm for the multispectral imaging system. Over a period of 6 months, 660 chicken images were collected in three batches. An image processing algorithm to locate the region of interest (ROI) was developed in order to define four classification areas on each image: whole carcass (WC), region of interest (ROI), upper region (UR), and lower region (LR). Three feature types, average intensity (AI), average normalization (AN), and average difference normalization (ADN), were defined using several wavebands for a total of 12 classification features. A decision tree algorithm was used to determine threshold values for each of the 12 classification features in each of the four classification areas. The AI feature type was found to identify wholesome and systemically diseased chickens better than the AN and ADN features types. Classification by AI in the ROI area, using the 540 nm and 580 nm wavebands, achieved the best accuracies. AI540 achieved 96.3% and 97.1% classification accuracies for wholesome and systemically diseased chickens, respectively. AI580 achieved 96.3% and 98.6% classification accuracies for wholesome and systemically diseased chickens, respectively. This simple differentiation method shows potential for automated on-line chicken inspection.

Yang, C.-C., K. Chao, Y. R. Chen and H. L. Early. 2005. Systemically diseased chicken identification using multispectral images and region of interest analysis. Computers and Electronics in Agriculture, 49(2): 255-271.

            A simple image differentiation method for the identification of systemically diseased chickens was developed and cross-system validated using two different multispectral imaging systems. The first system acquired images at three wavelengths, 460 nm, 540 nm, and 700 nm, for a batch of 164 wholesome and 176 systemically diseased chicken carcasses. The second system acquired images at four wavelengths, 488 nm, 540 nm, 580 nm, and 610 nm, for a second batch of 332 wholesome and 318 systemically diseased chicken carcasses. Image masking was performed using the wavelengths of 700 nm and 610 nm for the first and second imaging systems, respectively. The relative reflectance intensity at individual wavelengths, ratio of intensities between pairs of wavelengths, and intensity combinations based on Principal Component Analysis (PCA) were analyzed. It was found that the wavelengths of 540 nm and 580 nm are vital for successful chicken image differentiation. With proper wavelength selection, PCA can be useful for multispectral image analysis. The wavelength of 540 nm, selected as the key wavelength, was used in both imaging systems for image differentiation. An image processing algorithm was developed to define and locate the region of interest (ROI) as the differentiation area on the image. Based on ROI analysis, a single threshold was generated for image differentiation. The average relative reflectance intensity of the ROI was calculated for each chicken image. The Classification and Regression Trees (CART) decision tree algorithm was used to determine the threshold value to differentiate systemically diseased chickens from wholesome ones. The first differentiation threshold, based on the first image batch and generated by the decision tree method, was applied to the second image batch for cross-system validation, and vice versa. The accuracy from validation was 95.7% for wholesome and 97.7% of systemically diseased chickens for the first image batch, and 99.7% for wholesome and 93.5% for systemically diseased chickens for the second image batch. The threshold values, each generated using only one of the two image batches, were similar. The results showed that using a single key wavelength and a threshold, this simple image processing and differentiation method could be used in automated on-line applications for chicken inspection.

Yang, C.-C., K. Chao and Y. R. Chen. 2005. Development of multispectral imaging processing algorithms for identification of wholesome, septicemia, and inflammatory process chickens. Journal of Food Engineering, 69(2): 225-234.

            A multispectral imaging system and image processing algorithms for food safety inspection of poultry carcasses were demonstrated.  Three key wavelengths of 460, 540, and 700 nm, previously identified using a visible/near-infrared spectrophotometer, were implemented in a common-aperture multispectral imaging system, and images were collected for 174 wholesome, 75 inflammatory process, and 170 septicemic chickens.  Principal component analysis was used to develop an algorithm for separating septicemic chickens from wholesome and IP chickens based on average intensity of first component images.  A threshold value of 105 was able to correctly separate 95.6% of septicemic chickens.  To differentiate inflammatory process chickens, a region of interest was defined from which spectral features were determined.  The algorithm was able to correctly identify 100% of inflammatory process chickens by detecting pixels that satisfied the spectral feature conditions.  A decision tree model was created to classify the three chicken conditions using inputs from the two image processing algorithms.  The results showed that 89.6% of wholesome, 92.3% of inflammatory process, and 94.4% of septicemic chickens were correctly classified.

Yang, C.-C., S. O. Prasher, R. Lacroix, and S. H. Kim. 2004. Application of multivariate adaptive regression splines (MARS) to simulate soil temperature. Transactions of the ASAE, 47(3): 881-887.

            A new and flexible regression model, Multivariate Adaptive Regression Splines (MARS), is introduced and applied to simulate soil temperature at three depths. MARS uses a divide-and-conquer approach to automatically classify the training data into several groups. In each group, a regression line or hyperplane is generated. Compared to other intelligent computing technologies, MARS is fast, flexible, and capable of determining the important inputs to the model. The inputs to the model include the day of the year, the maximum and minimum air temperatures, rainfall, and potential evapotranspiration. The outputs contain the soil temperatures at depths of 100, 500, and 1500 mm. The performance of MARS was compared to that of artificial neural networks (ANNs). The correlation coefficients of linear regression from both MARS and ANNs were always higher than 0.950. MARS also indicated that the day of the year is the input that is most significant to the output, followed by the minimum air temperature. The results demonstrate the potential of MARS to be used as a regression technology in agricultural applications.

Yang, C.-C., S. O. Prasher and P. K. Goel. 2004. Differentiation of crop and weeds by decision-tree analysis of multi-spectral data. Transactions of the ASAE, 47(3): 873-879.

            The purpose of this study was to use a data mining technique, decision trees, to classify multi-spectral images of experimental plots having different crop and weed populations. Eleven types of plots were prepared for this study. Eight types were seeded with corn or soybeans and were either: 1) weed-free; 2) co-populated by velvet leaf only; 3) co-populated with a mixture of grass species; or 4) , co-populated with the predominant weed species of the regions. The other three types were as 2), 3), and 4), with neither corn nor soybeans. An aircraft-mounted pushbroom imaging spectrometer was used to obtain scans of the plots in one blue, five green, five red, and thirteen infrared bands. Eight classification problems involving different degrees of recognition complexity were set up. Each was tested using three different input types from the multi-spectral data. The three types of input were: a) absolute values of radiance from the 24 wave bands, b) vegetation index (VI) which consists of 12 inputs, and c) normalized difference vegetation index (NDVI) which consists of 65 inputs. Results showed that the most complex classification problem (distinguishing between 11 crop/weed combinations) was best resolved using the NDVI inputs (success classification of 0.85 as compared with 0.79 and 0.55 for the absolute radiance and VI respectively). Moreover, NDVI performed best as inputs in seven out of the eight problems.

Yang, C.-C., S. O. Prasher, R. Lacroix and S. H. Kim. 2003. A multivariate adaptive regression spines model for simulation of pesticide transport in soils. Biosystems Engineering, 86(1): 9-15.

            In this study, an innovative and intelligent computing regression algorithm, multivariate adaptive regression splines (MARS), was applied to simulate pesticide transport in soils. Using a divide-and-conquer method, the algorithm classifies the training data into several groups, in each of which a regression line or hyperplane is fitted. Compared to other intelligent computing technologies, MARS is fast, flexible, and capable of determining the important sequence of inputs to the output. This study evaluated MARS by applying it to simulate pesticide concentration levels at different soil depths and at various times. The model inputs included the number of days after pesticide application, accumulated rainfall, accumulated potential evapotranspiration, accumulated soil temperatures at depths of 100 mm in the morning as well as in the afternoon, and tillage practices. Several MARS models were developed to simulate the concentration levels of atrazine, deethylatrazine, and metolachlor at depths of 0 to 75 mm and 75 to 150 mm, respectively. The performance of MARS was compared to that of artificial neural networks (ANNs) using standard errors and correlation coefficients of linear regression. The results show the strong potential of MARS to be applied to agriculture as a regression technology.

Yang, C.-C., S. O. Prasher, P. Enright, C. Madramootoo, M. Burgess, P. K. Goel and I. Callum. 2003. Application of decision tree technology for image classification using remote sensing data. Agricultural Systems, 76(3): 1101-1117.

           Hyperspectral images of plots, cropped with silage or grain corn and cultivated with conventional tillage, reduced tillage, or no till, were classified using the classification and regression tree (C&RT) approach, an innovative intelligent computational algorithm in data mining. Each tillage/cropping combination was replicated three times, for a total of 18 plots. Five hyperspectral reflectance measurements per plot were taken randomly to obtain a total of 90 measurements. Images were taken on June 30, August 5, and August 25, 2000 to reflect three stages of crop development. Each measurement consisted of reflectances in 71 wave bands ranging from 400 to 950 nm. C&RT models were developed separately for the three observation dates, using the 71 reflectances as inputs to classify the image according to: a) tillage practice, b) residue level, c) cropping practices, d) tillage/cropping (residue) combination. C&RT models could generally distinguish tillage practices with a classification accuracy of 0.89 and residue levels with a classification accuracy of 0.98.

Yang, C.-C., S. O. Prasher, J.-A. Landry and H. S. Ramaswamy. 2003. Development of a herbicide application map using artificial neural networks and fuzzy logic. Agricultural Systems, 76(2): 561-574.

            The primary objective in this project was to develop a precision herbicide-spraying system in a corn field. Ultimately, such a system would involve real-time image collection and processing, weed identification, mapping of weed density, and sprayer control using a digital camera. A proposed image processing method involving artificial neural networks (ANNs) was evaluated for image recognition accuracy, computer time and memory requirements. The greenness method, based on a pixel-by-pixel comparison of red-green-blue intensity values, was successfully developed. The recognition of weeds in the field was then simplified by taking images between the corn rows. The images were processed by the greenness method to obtain percent greenness in each image. This information was used to create weed coverage and weed patchiness maps. Based on these maps, herbicide application rates were determined for each spot in the field. This was done by using the weed coverage and weed patchiness maps as inputs to a simulated fuzzy logic controller, and integrating the output of the controller over the field area corresponding to the input images. Simulations using different fuzzy rules and membership functions indicated that the precision spraying has potential for reducing water pollution from herbicides needed for weed control in a corn field.

Yang, C.-C., S. O. Prasher and J.-A. Landry. 2003. Development of an image processing system and a fuzzy controller for site-specific herbicide applications. Precision Agriculture, 4(1): 5-18.

            In precision farming, image analysis techniques can aid farmers in the site-specific application of herbicides and thus lower the risk of soil and water pollution by reducing the amount of chemicals applied. Using weed maps built with image analysis techniques, farmers can learn about the weed distribution in the crop. In this study, a digital camera was used to take a series of grid-based images covering the soil between rows of corn in a field in southwestern Quebec in May of 1999. Weed coverage was determined from each image using a “greenness method” in which the red, green and blue intensities of each pixel were compared. Weed coverage and weed patchiness were estimated based on the percent greenness area in the images. This information was used to create a weed map. Using weed coverage and weed patchiness as inputs, a fuzzy logic model was developed for use in determining site-specific herbicide application rates. A herbicide application map was then created for further evaluation of herbicide application strategy. Simulations indicated that significant amounts of herbicide could be saved using this approach.

Yang, C.-C., S. O. Prasher and J.-A. Landry. 2002. Weed recognition in corn fields using back-propagation neural network models. Canadian Biosystems Engineering, 44:715-722.

            The objective of this study was to develop back-propagation artificial neural network (ANN) models to distinguish young corn plants from weeds. Digital images were taken in the field under various natural lighting conditions. The images were cropped and resized to smaller sub-images containing either corn plants or weeds. The green objects in the images were extracted with the greenness method, thus counting the pixels with a green intensity larger than red and blue intensities, and replacing other background pixels with the intensity of zero. The extracted colour images were then converted to intensity images to save computational efforts during the ANN model development. The number of images available for training was quadrupled by rotating counter-clockwise each image by 90, 180 and 270 degrees. Several hundred images of corn plants and weeds were used for training the model. The ability of the ANN models to discriminate weeds from corn was tested. The highest success recognition rate was for corn at 100%, followed respectively by Abutilon theophrasti at 92%, Chenopodium album at about 62% and Cyperus esculentus at 80%. The ANN model required less than one minute to process two-hundred-and-fifty images containing 80x80 pixels. The inability of ANN models to properly classify weeds and corn plants was due to our inability to use a greater number of processing elements in the formulation of the ANN models. This was most likely caused by issues related to PC memory management.

Yang, C.-C., S. O. Prasher, J. Whalen and P. K. Goel. 2002. Use of hyperspectral imagery for identification of different fertilization methods with decision tree technology. Biosystems Engineering, 83(3): 291-298.

            This paper introduces data mining technology designed to classify agricultural fields under different manure/fertiliser application strategies. During the summer of 2000, airborne hyperspectral data was collected three times at two field sites in southwestern Quebec, Canada. One field site contained 24 plots (20 m by 24 m) that were amended with manure treatments and planted with maize and soya beans. The second field site contained 18 plots (18.5 m by 75 m) that received chemical fertilisers and were planted with maize. Reflectances of 71 wave bands of hyperspectral data (400 nm for violet to 940 nm for near infrared) were collected from five subplots within each of the 42 plots. The decision tree algorithm of data mining technology was used to distinguish between manure and chemical fertiliser treatments. The decision tree algorithm divides the data to reduce the deviance, and classifies them into the pre-defined categories as many tree branches. The success of the classification rate was as high as 91% for the early planting season, 99% for the mid planting season, and 95% for the late planting season. The accuracy of the results, demonstrates that data mining technology could be used for remote sensing imagery classification of fertiliser applications.

Yang, C.-C., S. O. Prasher, J.-A. Landry and H. S. Ramaswamy. 2002. Development of neural networks for weed recognition in corn fields. Transactions of the ASAE, 45(3): 859-864.

            The main objective of this project was to develop a weed recognition system based on artificial neural networks to assist in the precision application of herbicides in corn fields. Digital images were collected in May 1998 using a commercially-available digital camera. The intensities of the three primary colors, red, green, and blue, were compared for each pixel of the images. The three intensities of a pixel remained unchanged when, in this pixel, the green intensity was greater than each of the other two; otherwise, the three intensities of this pixel were set to zero. Background objects, except plants, were thus removed from the images. The resulting pixel intensities of the modified images were used as the inputs for Learning Vector Quantization (LVQ) artificial neural networks (ANNs). ANNs were trained to distinguish corn from weeds, as well as to differentiate between weed species. The success rate for a single ANN in distinguishing a given weed species from corn was as high as 90%, and as high as 80% in distinguishing any of four weed species from corn. Better success rates might be obtainable with more elaborate schemes for data input and/or structural improvements such as cascading. The image-processing time for the ANNs was as short as 0.48 s per image, thus making it useful for real-time data processing and application of herbicides. The development of such ANNs for weed recognition could be useful in precision farming to guide site-specific herbicide application and ultimately reduce the total amount of herbicide applied as well as lowering the risk of pollution.

Yang, C.-C., S. O. Prasher, J.-A. Landry and R. Kok. 2002. The development of image processing and weed localization algorithms for precision farming. Biosystems Engineering, 81(2): 137-146.

    The primary goal of this study was to develop the image processing component of a weed detection and mapping system that can be applied to precision farming.  The image processing algorithm first located the green objects in the digital image, then calculated weed density based on the greenness ratio.  The greenness ratio was the portion of an image in which the green intensity was larger than the red and blue intensities.  Some smoothing filters were also investigated to remove noise from the image.  Three approaches to determine the greenness ratio were evaluated and compared to the manual survey using a planimeter.  With the similar greenness ratios obtained from these approaches, the approach using MATLAB built-in function ‘vectorisation of loops’ gave comparable results to those using a planimeter in the shortest amount of time, 0.8 seconds per image, compared to 83 to 88 seconds per image by the other two approaches applying the greenness method to different image formats without the vectorisation function.  Using digital images taken of a corn field in 1999, the weed distribution on a portion of the field was mapped.  The results showed that only 59% of the corn field required full herbicide application based on having a weed density of 5% or more.  The proposed image processing method is a simple and fast way to detect weeds with a commercial digital camera.  It can become an important component of a weed detection and mapping system by assisting in site-specific application of herbicides in order to reduce herbicide costs and environmental pollution.

Yang, C.-., S. O. Prasher, J.-A. Landry, J. Perret and H. S. Ramaswamy. 2000. Recognition of weeds with image processing and their use with fuzzy logic for precision farming. Canadian Agricultural Engineering, 42(4): 195-200.

    Herbicide use can be reduced if the spatial distribution of weeds in the field is taken into account. This paper reports the initial stages of development of an image capture/processing system to detect weeds, as well as a fuzzy logic decision-making system to determine where and how much herbicide to apply in an agricultural field. The system used a commercially available digital camera and a personal computer. In the image processing stage, green objects in each image were identified using a greenness method that compared the red, green and blue (RGB) intensities. The RGB matrix was reduced to a binary form by applying the following criterion: if the green intensity of a pixel was greater than the red and the blue intensities, then the pixel was assigned a value of one; otherwise the pixel was given a value of zero. The resulting binary matrix was used to compute greenness area for weed coverage, and greenness distribution of weeds (weed patch). The values of weed coverage and weed patch were inputs to the fuzzy logic decision-making system, which used the membership functions to control the herbicide application rate at each location. Simulations showed that a graduated fuzzy strategy could potentially reduce herbicide application by 5 to 24%, and that an on/off strategy resulted in an even greater reduction of 15 to 64%.

Yang, C.-C., S. O. Prasher, J.-A. Landry, H. S. Ramaswamy and A. DiTommaso. 2000. Application of artificial neural networks in image recognition and classification of crop and weeds. Canadian Agricultural Engineering, 42(3): 147-152.

    The objective of this study was to develop a back-propagation artificial neural network (ANN) model that could distinguish young corn plants from weeds. Although only the colour indices associated with image pixels were used as inputs, it was assumed that the ANN model could develop the ability to use other information, such as shapes, implicit in these data. The 756x504 pixel images were taken in the field and were then cropped to 100x100-pixel images depicting only one plant, either a corn plant or weeds. There were 40 images of corn and 40 of weeds. The ability of the ANNs to discriminate weeds from corn was then tested on 20 other images. A total of 80 images of corn plants and weeds were used for training purposes. For some ANNs, the success rate for classifying corn plants was as high as 100%, whereas the highest success rate for weed recognition was 80%. This is considered satisfactory, given the limited amount of training data and the computer hardware limitations. Therefore, it is concluded that an ANN-based weed recognition system can potentially be used in the precision spraying of herbicides in agricultural fields.

Yang, C.-C., C. S. Tan and S. O. Prasher. 2000. Artificial neural networks for subsurface drainage and subirrigation systems in Ontario, Canada. Journal of the American Water Resources Association, 36(3): 609-618.

    Artificial neural network (ANN) models were developed to simulate fluctuations in midspan water table depths (WTD) given rainfall, potential evapotranspiration, and irrigation inputs on a Brookston clay loam in Woodslee, Ontario, having a dual-purpose subsurface drainage/subirrigation setup. Water table depths and meteorologic data collected at this site from 1992 to 1994 and from 1996 to 1997 were used totrain the ANNs. The ANNs were then used for real-time control and time series simulations. The lowest root mean squared errors (RMSE) for teh various ANNs were 60.6 mm for real-time control simulation, and 88.4 mm for time-series simulation of water table depths. It was possible to simulate WTD for the different modes of water table management in one network by incorporating an indicator for switching from one to the other. The ANN simulations were quite good even though the training data sets had irregular measurement intervals. With fewer input parameters and small network structures, ANNs still provided accurate results and required little time for training and execution. ANNs are therefore easier and faster to develop and run than conventional models and can contribute to the proper management of subsurface drainage and subirrigation systems.

Yang, C.-C., S. O. Prasher and C. S. Tan. 1999. An artificial neural network model for water table management systems. Canadian Water Resources Journal, 24(1): 25-33.

    This paper presents the development of an artificial neural network (ANN) model to simulate fluctuations in midspan water table depths, influenced by daily rainfall and potential evapotranspiration rates. Unlike conventional mathematical models, ANN models do not require, a priori, explicit knowledge of the relationship between inputs and outputs. This knowledge is obtained through training: field observations of inputs and outputs. Compared with conventional mathematical models, ANN models require fewer input parameters and can be run quickly on a PC. These benefits prove to be significant in the real-time control of subsurface drainage and subirrigation systems. In this study, ANN models were developed to simulate water table depths during subsurface drainage and subirrigation. They were trained and tested using field observations on water table depths, made at an agricultural field in Woodslee, southern Ontario, Canada, from 1992 to 1994. The results show the applicability of ANNs in drainage modeling. They can simulate water table depth processes effectively, without requiring a large number of input parameters. ANN models are simple to build and easy to run, allowing accurate results to be obtained quickly. Thus, they can be used in the real-time control of subsurface drainage and subirrigation systems.

Yang, C.-C., R. Lacroix, and S. O. Prasher. 1998. The use of back-propagation neural networks for the simulation and analyses of time-series data in subsurface drainage systems. Transactions of the ASAE, 41(4): 1181-1187.

            This study was undertaken to investigate the application of artificial neural networks (ANNs) in the simulation of subsurface drainage systems. Back-propagation ANNs were trained to imitate a conventional mathematical model, DRAINMOD, in the simulation of water table depths. For good representation of the dynamics of a soil system, the time lag procedure was developed to feed the input values of previous time steps. The results show that the use of time lag procedures produced significant impacts on the ANN performances. In this study, two methods are introduced to analyze and compare the impact of various strategies of data input into the ANNs and DRAINMOD. In the model-response analysis, one input was varied with different impulses while the other inputs were kept constant. The results showed that the optimal time-dependence period of the ANN inputs should be determined by the saturated hydraulic conductivity and actual distance from the soil surface to the impermeable layer. In the sensitivity analysis, each processing element in the input layer of an ANN with the lag procedure was disabled respectively. When the processing elements corresponding to the inputs of rainfall and previous water table were disabled, the r² values of linear regression could decrease to less than 0.1. The results showed that these inputs were more important to the ANNs. These methods can be used to evaluate the performances of simulation models for time-series data and real-time control, particularly when the real situation is not available.

Yang, C.-C., S. O. Prasher, S. Sreekanth, N. K. Patni and L. Masse. 1997. An artificial neural network model for simulating pesticide concentrations in soil. Transactions of the ASAE, 40(5): 1285-1294. American Society of Agricultural Engineers.

            The simulation of pesticide concentrations in soil requires knowledge of complex physico-chemical processes that pesticides undergo, in both unsaturated and saturated zones. Generally, conventional models are used for this purpose. This paper reports on the use of artificial neural networks (ANNs) to simulate pesticide concentrations in agricultural soils. The main advantages of ANN modeling are significantly fewer input parameters and a very short execution time. An ANN model can be executed in real-time, while the sprayer is working in the field, in order to adjust application rates to the real extent of the problem. In this study, an ANN model was built and trained with inputs of; accumulated daily rainfall, soil temperature, potential evapotranspiration, as well as tillage practices and the number of days elapsed after pesticide application. The outputs of the ANN model were the daily accumulated amounts of pesticide levels in the soil. The results were compared with the data collected in 1992 and 1993 from an agricultural field in Ottawa, Canada. The results show the benefits of ANNs in predicting pesticide concentrations in agricultural soils. In this study, only six input parameters are required with fast execution. The ANN-based model can be very helpful in making quick and appropriate decisions during real-time application of pesticides. In this study, the performance of ANNs was investigated when the amount of available training data was limited. The results indicated that the performance of ANNs was good, in spite of limited data, with the values of root-mean-square error and standard deviation being generally lower than 0.2 µg/g. However, the performance of ANNs could be improved with more training data obtained from field experiments.

Yang, C.-C., S. O. Prasher and G. R. Mehuys. 1997. An artificial neural network to estimate soil temperature. Canadian Journal of Soil Science, 77(3): 421-429. Agricultural Institute of Canada. Canadian Society of Soil Science. Canadian Society of Agrometeorology.

           This study was undertaken to develop an artificial neural network (ANN) model for transient simulation of soil temperature at different depths in the profile. The capability of ANN models to simulate the variation of temperature in soils was investigated by considering readily available meteorologic parameters. The ANN model was constructed by using five years of meteorologic data, measured at a weather station at the Central Experimental Farm in Ottawa, Ontario, Canada. The model inputs consisted of daily rainfall, potential evapotranspiration, and the day of the year. The model outputs were daily soil temperatures at the depths of 100, 500 and 1500 mm.
            The estimated values were found to be close to the measured values, as shown by a root-mean-square error ranging from 0.59 to 1.82oC, a standard deviation of errors from 0.61 to 1.81oC, and a coefficient of determination from 0.937 to 0.987. Therefore, it is concluded that ANN models can be used to estimate soil temperature by considering routinely measured meteorologic parameters. In addition, the ANN model executes faster than a comparable conceptual simulation model by several orders of magnitude.

Yang, C.-C., S. O. Prasher, R. Lacroix, S. Sreekanth, A. Madani and L. Masse. 1997. Artificial neural network model for subsurface-drained farmlands. Journal of Irrigation and Drainage Engineering, 123(4): 285-292. Water Resources Engineering Division, American Society of Civil Engineers.

            This paper describes the development of an artificial neural network (ANN) model, to simulate fluctuations in midspan water-table depths and drain outflows, as influenced by daily rainfall and potential evapotranspiration rates. Unlike conventional models, ANN models do not require explicit relationship between inputs and outputs. Instead, ANNs map the implicit relationship between inputs and outputs through training by field observations. Compared with conventional models, the ANN model requires fewer input parameters since the inputs that remain constant are not considered by ANNs. Therefore, ANNs can be executed quickly on a microcomputer. These benefits can be exploited in the real-time control of water-table management systems.
            The model was developed using field observations of water-table depths from 1991 to 1993 and drain outflows from 1991 to 1994, made at an agricultural field in Ottawa, Canada. The root mean squared errors and standard deviation of errors of simulated results were found to range from 46.5 to 161.1 mm and 46.6 to 139.2 mm, respectively, thus showing potential applications of ANNs in land drainage engineering.

Yang, C.-C., S. O. Prasher, G. R. Mehuys and N. K. Patni. 1997. Application of artificial neural networks for simulation of soil temperature. Transactions of the ASAE, 40(3): 649-656. American Society of Agricultural Engineers.

            In an agricultural ecosystem, soil temperature can affect the growth of plants and organisms, the fate and transport of chemicals, and many other natural phenomena. Simulation of soil temperature is essential to support many agricultural models. Modelling the fluctuations of soil temperature at different depths is complicated considering the great number of variables. In this study, a simple model, based on an artificial neural network (ANN), was developed to simulate daily soil temperatures at 100, 500 and 1500 mm depths, in a  soil from Ottawa, Ontario, Canada in an attempt to develop a simple, fast and more accurate ANN model than the conceptual models currently used to simulate soil temperature. The inputs for the ANN model included: daily rainfall, potential evapotranspiration, maximum and minimum air temperature, and the day of the year. These input factors are all easy to obtain and are measured at most weather stations world-wide. The parity between the measured and the simulated data, resulting from ANNs, shows the ability of simple ANN models to simulate soil temperature. The results obtained from ANN models varied within a root- mean-square error range from 0.63 to 1.39 C, standard deviations from 0.61 to 1.39 C and coefficients of determination (r²) from 0.937 to 0.985. The accuracy of the simulations shows the simplicity with which ANNs can be used to model complicated phenomena in agricultural systems. The short time of execution (a few seconds for a one-year simulation) is another benefit of ANN models. Many simulation models, such as for pesticide fate and transport, nutrient movement in soils, and soil bioremediation, require timely fluctuations of soil temperatures. For such uses, the fast execution of ANNs is very helpful. Therefore, this technology could prove very useful for decision support systems which require real-time control in agricultural applications.

Yang, C.-C., S. O. Prasher, R. Lacroix and A. Madani. 1997. Application of Artificial neural networks in subsurface drainage system design. Canadian Water Resources Journal, 22(1): 1-12. Canadian Water Resources Association.

            In this paper we describe the use of artificial neural networks (ANNs) to model the performance of a subsurface drainage system in Nova Scotia, Canada. The ANN model was built and trained by using measured data on midspan water-table depths and drain outflows from an alfalfa field. The results obtained by the ANN model were compared with the measured data, and with the simulated results from a conventional mathematical model, DRAINMOD. The results show that the ANN model can simulate midspan water-table fluctuations and drain outflows quite well. The ANN model runs significantly faster and requires significantly fewer inputs than DRAINMOD. The ANN simulations depend heavily on the quality of the input data for both average and extreme conditions. This study indicates that an ANN model may be used effectively for the design and evaluation of subsurface drainage systems. The benefits of ANNs are speed, accuracy, ease-of-use and flexibility.

Yang, C.-C., S. O. Prasher and R. Lacroix. 1996. Applications of artificial neural networks to land drainage engineering. Transactions of the ASAE, 39(2): 525-533. American Society of Agricultural Engineers.

            An ANN model was developed and trained by using the simulated midspan water-table depths from DRAINMOD: a conventional water-table management model. Compared to DRAINMOD, the model is very simple to run, and requires only a small amount of data, such as precipitation, evapotranspiration and initial midspan water-table depth.
            The results indicate that the ANN model can make predictions similar to DRAINMOD, with the least root mean square error of 0.1193, and doing this significantly faster and with fewer input data. The results also indicated that the successful prediction of midspan water-table depths depends upon the inclusion of data indicating average as well as extreme conditions, in order to train the ANNs. Given such data, ANNs perform well under general conditions.
            Generally, the ANN structure with 6 processing elements and 1 hidden layer was sufficient for this study. It was found that the networks should be trained with at least 145,000 cycles, but more than 200,000 cycles are unnecessary. A feedback procedure was implemented which fed the previous water-table depth output back into the current input. In addition, a lag procedure was suggested which improved the performance of ANNs under irregular situations, such as sudden and large rainstorms. A three-day lag of all input parameters was the best choice when the weather conditions were irregular.
            The benefits of ANNs are speed, accuracy, ease-of-use and flexibility, thus making ANN models suitable for water-table management systems that require a real-time control.

Yang, C.-C., S. O. Prasher and R. Lacroix. 1996. Applications of artificial neural networks to simulate water-table depths under subirrigation. Canadian Water Resources Journal, 21(1): 27-44. Canadian Water Resources Association.

            In this paper we report on the development of an artificial neural network (ANN) model for the design and evaluation of subirrigation systems. The model was formulated and trained by using simulation data from DRAINMOD, a well-known water-table management model. The DRAINMOD model was used to simulate subirrigation in a clay loam soil with 26 years of weather data. One of the main model outputs, the midspan water-table depth, was used to train the ANN model. In addition, the ANN model required data on rainfall, evapotranspiration, weir levels for water-table control, and midspan water-table depths for a learning situation. The results show that the ANN model was able to simulate as well as DRAINMOD. The root mean square (RMS) errors between the ANN and DRAINMOD simulated water-table depths were less than 0.1. Compared to DRAINMOD, the ANN model required very little data to run, and it also executed a lot faster. Therefore, there is a possibility of using such ANN models in real-time control of subirrigation systems, where important decisions need to be made on very short notice. In addition, given the fact that the ANN technology emulated the complex conventional model DRAINMOD, it should be possible to apply this technology to other problems in soil hydrology and contaminant movement.

Sreekanth, S., S. O. Prasher and C.-C. Yang. 1997. Importance of choice of input parameters in artificial neural network simulation of water-table depths. Canadian Water Resource Journal, 22(2): 111-124. Canadian Water Resources Association.


 
My Home My Laboratory My Lounge My Library

Any of your comments are WELCOME!!!chun-chieh.yang@mcgill.ca