Nonlinear Backpropagation Neural Network Research Articles

Unconstrained Estimation of Multitype Car Rental Demand

The unconstrained demand forecast for car rentals has become a difficult problem for revenue management due to the need to cope with a variety of rental vehicles, the strong subjective desires and requests of customers, and the high probability of upgrading and downgrading circumstances. The unconstrained demand forecast mainly includes repairing of constrained historical demand and forecasting of future demand. In this work, a new methodology is developed based on multiple discrete choice models to obtain customer choice preference probabilities and improve a known spill model, including a repair process of the unconstrained demand. In addition, the linear Holt–Winters model and the nonlinear backpropagation neural network are combined to predict future demand and avoid excessive errors caused by a single method. In a case study, we take advantage of a stated preference and a revealed preference survey and use the variable precision rough set to obtain factors and weights that affect customer choices. In this case study and based on a numerical example, three forecasting methods are compared to determine the car rental demand of the next time cycle. The comparison with real demand verifies the feasibility and effectiveness of the hybrid forecasting model with a resulting average error of only 3.06%.

Assessment of calibration methods for nitrogen estimation in wet and dry soil samples with different wavelength ranges using near-infrared spectroscopy

Choosing an appropriate wavelength range, extracting optimal wavelength variables, and selecting suitable statistical analysis methods are of great importance for improving the prediction accuracy of soil nitrogen (N) with near-infrared (NIR) spectroscopy. In this study, the prediction performances of two different wavelength ranges, a short wavelength range (SWR) of 900–1,700 nm and a full wavelength range (FWR) of 900–2,500 nm, are evaluated for the measurement of soil N content. Spectral scanning is performed on wet and dry-sieve soil samples to assess the effect of moisture on the prediction performance of soil N. Two calibration methods, a commonly used linear partial least squares regression (PLSR) and a nonlinear back propagation neural network (BPNN), are used. To understand if it is possible to reduce the number of wavelength variables without decreasing prediction accuracy, we introduce a successive projection algorithm (SPA) to extract wavelength variables that are minimally redundant. The results show that models developed within FWR outperform those developed within SWR, regardless of wet or dry soil conditions, which can be attributed to the presence of more spectral information related to soil N in FWR. Compared with PLSR, BPNN is a better choice for predicting soil N, because BPNN models provide higher accuracy. The best prediction performance is achieved by BPNN method in FWR using a SPA with Rp2 = 0.93, RMSEP = 0.0297% and RPD = 4.00 of wet soil samples, and Rp2 = 0.99, RMSEP = 0.0132% and RPD = 8.76 of dry soil samples. Additionally, we demonstrate that using the SPA algorithm significantly reduces the number of wavelength variables while maintaining high prediction accuracy. The characteristic wavelengths selected by the SPA algorithm follow the principle of material spectral absorption. It is worth noting that dry soil conditions lead to superior performance over wet soil conditions for the measurement of soil N, which can be attributed to the removal effect of moisture content from the wavelength region and the utilization of important absorption features. However, even under wet soil conditions, the simplified calibration models based on selected SPA variables obtain excellent quantitative prediction using the BPNN method in the SWR range, with Rp2 = 0.91, RMSEP = 0.0305%, and RPD = 3.47. It is important to expand large-scale detection applications for the measurement of soil N.

Soil total nitrogen estimation of alpine grassland using visible/near-infrared spectra: a comparison of multivariate techniques with different spectral transformations

Rapid, precise, and quantitative assessment of soil total nitrogen (TN) of alpine grassland is crucial for monitoring the environmental conditions of grasslands. The right choice of preprocessing and calibration methods may improve the model performance of soil property estimation in visible/near infrared reflectance spectroscopy (VNIRS). VNIRS (350 to 2500 nm) measurements were made of 80 soil samples from the northeast part of the Tibetan Plateau. To reduce the signal noise, the Savitzky–Golay technique was implemented and six spectral transformations were evaluated with the aim of identifying the best transformation for predicting soil TN. The performance of three linear calibration models, namely, stepwise multiple linear regression, principal component regression, partial least squares regression (PLSR), and a combination of these models with a nonlinear back propagation neural network (BPNN) was tested for accuracy in the measurement of soil TN. Our results indicate the following: (1) models using the derivative transformations of spectra as inputs perform better than those using nonderivative transformations, and the first derivative of the reciprocal of reflectance [(1 / R) ′ ] is the best spectral transformation for soil TN estimation; (2) models based on the PLSR and companion BPNN-latent variables (LVs) yield better predictions and smallest errors of soil TN. We demonstrate that (1 / R) ′ is the best spectral transformation and consistently improves the estimation models of soil TN. Among multivariate techniques, the BPNN-LVs is recommended for estimation of soil TN with great accuracy.

Monitoring urban impervious surface area change using China-Brazil Earth Resources Satellites and HJ-1 remote sensing images

Impervious surface area (ISA) plays an important role in monitoring urbanization and related environmental changes, and has become a hotspot in urban and environmental studies. Xuzhou City, located in northwest Jiangsu Province, China, is chosen as the study area, and two scenes of China-Brazil Earth Resources Satellites images and one scene of HJ-1 image are employed to estimate ISA percentage and analyze the change trend from 2001 to 2009. Using a linear spectral mixture model (LSMM) and nonlinear backpropagation neural network (BPNN) method, all pixels are decomposed to derive four fraction images representing the abundance of four endmembers: vegetation, high-albedo objects, low-albedo objects, and soil. The ISA percentage is then derived by the combination of high- and low-albedo fraction images after removing the influence of water. Some high spatial resolution images are selected to validate the ISA estimation results, and the experimental results indicate that the accuracy of BPNN is higher than LSMM. By comparing the urban ISA abundances derived by BPNN from three dates, it is found that the ISA of Xuzhou City has increased rapidly from 2001 to 2009, especially in the northeast and southeast regions, corresponding to the urban planning scheme and fast urbanization. Compared to other medium remote sensing images, the revisit cycle of HJ-1 multispectral image is only two days, demonstrating the potential of such data for ISA extraction in urbanization, disaster, and other related applications.

Nondestructive determination of nutritional information in oilseed rape leaves using visible/near infrared spectroscopy and multivariate calibrations

Nitrogen, phosphorus and potassium content are the three most important nutritional parameters for growing oilseed rape. We investigated visible and near infrared (Vis/NIR) spectroscopy combined with chemometrics for the fast and nondestructive determination of nutritional information in oilseed rape leaves. A total of 154 leaf samples were collected, with 104 randomly selected as the calibration set, and the remaining 50 samples used as the validation set. The performance of eight different preprocessing methods was compared in partial least squares (PLS) models. Some effective wavelengths selected by a successive projections algorithm (SPA) were also used to develop linear SPA-PLS, nonlinear back propagation neural network (BPNN), and nonlinear least squares-support vector machine (LS-SVM) models to determine nutritional information. The best prediction models were DOSC-PLS for nitrogen with r=0.9743 and RMSEP=0.1459, DOSC-SPA-BPNN for phosphorus with r=0.7054 and RMSEP=0.0594, and DOSC-SPA-BPNN for potassium with r=0.9380 and RMSEP=0.1788. The prediction precision for nitrogen and potassium determinations was acceptable for further practical applications, whereas further studies are needed to improve the prediction precision for phosphorus. The results indicated that Vis/NIR spectroscopy is feasible for nondestructive determination of nutritional information in oilseed rape leaves. It also provided an alternative technique for detecting other growth information about oilseed rape leaves.

Performance Evaluation of Computer Aided Diagnostic Tool (CAD) for Detection of Ultrasonic Based Liver Disease

Recent advances in digital imaging technology have greatly enhanced the interpretation of critical/pathology conditions from the 2-dimensional medical images. This has become realistic due to the existence of the computer aided diagnostic tool. A computer aided diagnostic (CAD) tool generally possesses components like preprocessing, identification/selection of region of interest, extraction of typical features and finally an efficient classification system. This paper enumerates on development of CAD tool for classification of chronic liver disease through the 2-D image acquired from ultrasonic device. Characterization of tissue through qualitative treatment leads to the detection of abnormality which is not viable through qualitative visual inspection by the radiologist. Common liver diseases are the indicators of changes in tissue elasticity. One can show the detection of normal, fatty or malignant condition based on the application of CAD tool thereby, further investigation required by radiologist can be avoided. The proposed work involves an optimal block analysis (64 x 64) of the liver image of actual size 256 x 256 by incorporating Gabor wavelet transform which does the texture classification through automated mode. Statistical features such as gray level mean as well as variance values are estimated after this preprocessing mode. A non-linear back propagation neural network (BPNN) is applied for classifying the normal (vs) fatty and normal (vs) malignant liver which yields a classification accuracy of 96.8%. Further multi classification is also performed and a classification accuracy of 94% is obtained. It can be concluded that the proposed CAD can be used as an expert system to aid the automated diagnosis of liver diseases.

The application of a non-linear back-propagation neural network to study the mass balance of Grosse Aletschgletscher, Switzerland

AbstractGlacier mass changes are considered to represent key variables related to climate variability. We have reconstructed a proxy for annual mass-balance changes in Grosse Aletschgletscher, Swiss Alps, back to AD 1500 using a non-linear back-propagation neural network (BPN). The model skill of the BPN performs better than reconstructions using conventional stepwise multiple linear regression. The BPN, driven by monthly instrumental series of local temperature and precipitation, provides a proxy for 20th-century mass balance. The long-term mass-balance reconstruction back to 1500 is based on a multi-proxy approach of seasonally resolved temperature and precipitation reconstructions (mean over a specific area) as input variables. The relation between the driving factors (temperature, precipitation) used and the reconstructed mass-balance series is discussed. Mass changes in Grosse Aletschgletscher are shown to be mainly influenced by summer (June–August) temperatures, but winter (December–February) precipitation also seems to contribute. Furthermore, we found a significant non-linear part within the climate–mass-balance relation of Grosse Aletschgletscher.

Attribution and detection of anthropogenic climate change using a backpropagation neural network

The climate system can be regarded as a dynamic nonlinear system. Thus traditional linear statistical methods are not suited to describe the nonlinearities of this system which renders it necessary to find alternative statistical techniques to model those nonlinear properties. In addition to an earlier paper on this subject (WALTER et al., 1998), the problem of attribution and detection of the observed climate change is addressed here using a nonlinear Backpropagation Neural Network (BPN). In addition to potential anthropogenic influences on climate (CO 2 -equivalent concentrations, called greenhouse gases, GHG and SO 2 emissions) natural influences on surface air temperature (variations of solar activity, volcanism and the El Nino/Southern Oscillation phenomenon) are integrated into the simulations as well. It is shown that the adaptive BPN algorithm captures the dynamics of the climate system, i.e. global and area weighted mean temperature anomalies, to a great extent. However, free parameters of this network architecture have to be optimized in a time consuming trial-and-error process. The simulation quality obtained by the BPN exceeds the results of those from a linear model by far; the simulation quality on the global scale amounts to 84% explained variance. Additionally the results of the nonlinear algorithm are plausible in a physical sense, i.e. amplitude and time structure. Nevertheless they cover a broad range, e.g. the GHG-signal on the global scale ranges from 0.37 K to 1.65 K warming for the time period 1856-1998. However the simulated amplitudes are situated within the discussed range (HOUGHTON et al., 2001). Additionally the combined anthropogenic effect corresponds to the observed increase in temperature for the examined time period. In addition to that, the BPN succeeds with the detection of anthropogenic induced climate change on a high significance level. Therefore the concept of neural networks can be regarded as a suitable nonlinear statistical tool for modeling and diagnosing the climate system.

Estimation of Bench-Scale Chlorine Decay in Drinking Water Using nth-Order Kinetic and Back Propagation Neural Network Models

Abstract This paper presents the development of two models for simulating residual chlorine decay in raw and treated waters collected from six different utilities of the Quebec City area. The data for modelling was generated by means of several bench-scale chlorination assays undertaken with chlorine doses that varied according to the content of natural organic matter in the water. The first model is a classical kinetic model in which chlorine decay is represented by a first or a second order function, according to the contact time of chlorine. A decay coefficient is estimated based on water quality and operational parameters, using linear regression. The second is a non-linear back propagation neural network model in which all the parameters responsible for chlorine decay are processed within a single model. The performances of both models were evaluated and compared using the database for model development and an independent database for validation. Both models demonstrated acceptable abilities for simulating residual chlorine decay. However, the back propagation neural network model gave significantly better results for conditions involving high chlorine dosage, high organic matter content and long reaction times during chlorination experiments. In the other cases, the performance of the kinetic model was slightly superior. Back propagation neural networks are also advantageous because they do not require assumptions about model order or the estimation of a chlorine decay coefficient. Also included in the paper is a discussion of possible strategies for use in future research work in order to generalize the results obtained.

Neural Network Model: BID Pricing Fraud

This paper focuses on judgments made by Defense Contract Audit Agency (DCAA) auditors. They must assess the likelihood of the existence of defective pricing in defense contracts. Defective pricing is a legal term referring to abuses in the federal contracting process.The research employed a triangulation approach to examine the processing strategy used by auditors in an environment of intermediate complexity. The first methodology utilized case studies to determine if there were differences in decision outcomes between groups where the DCAA auditors were forced to use step-by-step processing and groups where the DCAA auditors were exposed to all cues of the cases at the same time and could use end-of-sequence processing. The second methodology involved the development of two competing data-driven quantitative models using evaluation sheets from actual field visits. The predictive validity of a linear discriminate analysis model was compared with a nonlinear back propagation neural network model.This study demonstrated that processing strategy can impact decision outcomes, in accordance with existing theory. It supports the use of nonlinear pattern matching techniques, such as a back propagation neural network model, to model decision-making behavior where end-of-sequence processing is used.