Backward elimination model construction for regression and classification using leave-one-out criteria

A fundamental principle in practical nonlinear data modeling is the parsimonious principle of constructing the minimal model that explains the training data well. Leave-one-out (LOO) cross validation is often used to estimate generalization errors by choosing amongst different network architectures (M. Stone, "Cross validatory choice and assessment of statistical predictions", J. R. Stast. Soc., Ser. B, 36, pp. 117-147, 1974). Based upon the minimization of LOO criteria of either the mean squares of LOO errors or the LOO misclassification rate respectively, we present two backward elimination algorithms as model post-processing procedures for regression and classification problems. The proposed backward elimination procedures exploit an orthogonalization procedure to enable the orthogonality between the subspace as spanned by the pruned model and the deleted regressor. Subsequently, it is shown that the LOO criteria used in both algorithms can be calculated via some analytic recursive formula, as derived in this contribution, without actually splitting the estimation data set so as to reduce computational expense. Compared to most other model construction methods, the proposed algorithms are advantageous in several aspects; (i) There are no tuning parameters to be optimized through an extra validation data set; (ii) The procedure is fully automatic without an additional stopping criteria; and (iii) The model structure selection is directly based on model generalization performance. The illustrative examples on regression and classification are used to demonstrate that the proposed algorithms are viable post-processing methods to prune a model to gain extra sparsity and improved generalization.

Land-cover classification from airborne LIDAR data fused with aerial optical images

Airborne LIght Detection And Ranging (LIDAR) provides accurate height information for objects on the earth, which makes LIDAR become more and more popular in terrain and land surveying. In particular, LIDAR data offer vital and significant features for land-cover classification which is an important task in many application domains. In this paper, an unsupervised approach based on an improved fuzzy Markov random field (FMRF) model is developed, by which the LIDAR data, its co-registered images acquired by optical sensors, i.e. aerial color image and near infrared image, and other derived features are fused effectively to improve the ability of the LIDAR system for the accurate land-cover classification. In the proposed FMRF model-based approach, the spatial contextual information is applied by modeling the image as a Markov random field (MRF), with which the fuzzy logic is introduced simultaneously to reduce the errors caused by the hard classification. Moreover, a Lagrange-Multiplier (LM) algorithm is employed to calculate a maximum A posteriori (MAP) estimate for the classification. The experimental results have proved that fusing the height data and optical images is particularly suited for the land-cover classification. The proposed approach works very well for the classification from airborne LIDAR data fused with its coregistered optical images and the average accuracy is improved to 88.9%.

High-precision measurements of the co-polar correlation coefficient: non-Gaussian errors and retrieval of the dispersion parameter µ in rainfall

The co-polar correlation coefficient (ρhv) has many applications, including hydrometeor classification, ground clutter and melting layer identification, interpretation of ice microphysics and the retrieval of rain drop size distributions (DSDs). However, we currently lack the quantitative error estimates that are necessary if these applications are to be fully exploited. Previous error estimates of ρhv rely on knowledge of the unknown "true" ρhv and implicitly assume a Gaussian probability distribution function of ρhv samples. We show that frequency distributions of ρhv estimates are in fact highly negatively skewed. A new variable: L = -log10(1 - ρhv) is defined, which does have Gaussian error statistics, and a standard deviation depending only on the number of independent radar pulses. This is verified using observations of spherical drizzle drops, allowing, for the first time, the construction of rigorous confidence intervals in estimates of ρhv. In addition, we demonstrate how the imperfect co-location of the horizontal and vertical polarisation sample volumes may be accounted for. The possibility of using L to estimate the dispersion parameter (µ) in the gamma drop size distribution is investigated. We find that including drop oscillations is essential for this application, otherwise there could be biases in retrieved µ of up to ~8. Preliminary results in rainfall are presented. In a convective rain case study, our estimates show µ to be substantially larger than 0 (an exponential DSD). In this particular rain event, rain rate would be overestimated by up to 50% if a simple exponential DSD is assumed.

Model-based vehicle detection and classification using orthographic approximations

This paper reports the current state of work to simplify our previous model-based methods for visual tracking of vehicles for use in a real-time system intended to provide continuous monitoring and classification of traffic from a fixed camera on a busy multi-lane motorway. The main constraints of the system design were: (i) all low level processing to be carried out by low-cost auxiliary hardware, (ii) all 3-D reasoning to be carried out automatically off-line, at set-up time. The system developed uses three main stages: (i) pose and model hypothesis using 1-D templates, (ii) hypothesis tracking, and (iii) hypothesis verification, using 2-D templates. Stages (i) & (iii) have radically different computing performance and computational costs, and need to be carefully balanced for efficiency. Together, they provide an effective way to locate, track and classify vehicles.

Correlated observation errors in data assimilation

Data assimilation provides techniques for combining observations and prior model forecasts to create initial conditions for numerical weather prediction (NWP). The relative weighting assigned to each observation in the analysis is determined by its associated error. Remote sensing data usually has correlated errors, but the correlations are typically ignored in NWP. Here, we describe three approaches to the treatment of observation error correlations. For an idealized data set, the information content under each simplified assumption is compared with that under correct correlation specification. Treating the errors as uncorrelated results in a significant loss of information. However, retention of an approximated correlation gives clear benefits.

Improvements to the accuracy of measurements of NO2 by zenith-sky visible spectrometers II: errors in zero using a more complete chemical model

Using a flexible chemical box model with full heterogeneous chemistry, intercepts of chemically modified Langley plots have been computed for the 5 years of zenith-sky NO2 data from Faraday in Antarctica (65°S). By using these intercepts as the effective amount in the reference spectrum, drifts in zero of total vertical NO2 were much reduced. The error in zero of total NO2 is ±0.03×1015 moleccm−2 from one year to another. This error is small enough to determine trends in midsummer and any variability in denoxification between midwinters. The technique also suggests a more sensitive method for determining N2O5 from zenith-sky NO2 data.

Use of airborne remote sensing to detect riverside Brassica rapa to aid in risk assessment of transgenic crops

High resolution descriptions of plant distribution have utility for many ecological applications but are especially useful for predictive modelling of gene flow from transgenic crops. Difficulty lies in the extrapolation errors that occur when limited ground survey data are scaled up to the landscape or national level. This problem is epitomized by the wide confidence limits generated in a previous attempt to describe the national abundance of riverside Brassica rapa (a wild relative of cultivated rapeseed) across the United Kingdom. Here, we assess the value of airborne remote sensing to locate B. rapa over large areas and so reduce the need for extrapolation. We describe results from flights over the river Nene in England acquired using Airborne Thematic Mapper (ATM) and Compact Airborne Spectrographic Imager (CASI) imagery, together with ground truth data. It proved possible to detect 97% of flowering B. rapa on the basis of spectral profiles. This included all stands of plants that occupied >2m square (>5 plants), which were detected using single-pixel classification. It also included very small populations (<5 flowering plants, 1-2m square) that generated mixed pixels, which were detected using spectral unmixing. The high detection accuracy for flowering B. rapa was coupled with a rather large false positive rate (43%). The latter could be reduced by using the image detections to target fieldwork to confirm species identity, or by acquiring additional remote sensing data such as laser altimetry or multitemporal imagery.

Identifying management zones in agricultural fields using spatially constrained classification of soil and ancillary data

Site-specific management requires accurate knowledge of the spatial variation in a range of soil properties within fields. This involves considerable sampling effort, which is costly. Ancillary data, such as crop yield, elevation and apparent electrical conductivity (ECa) of the soil, can provide insight into the spatial variation of some soil properties. A multivariate classification with spatial constraint imposed by the variogram was used to classify data from two arable crop fields. The yield data comprised 5 years of crop yield, and the ancillary data 3 years of yield data, elevation and ECa. Information on soil chemical and physical properties was provided by intensive surveys of the soil. Multivariate variograms computed from these data were used to constrain sites spatially within classes to increase their contiguity. The constrained classifications resulted in coherent classes, and those based on the ancillary data were similar to those from the soil properties. The ancillary data seemed to identify areas in the field where the soil is reasonably homogeneous. The results of targeted sampling showed that these classes could be used as a basis for management and to guide future sampling of the soil.