Dragoljub Pokrajac

Modeling spatial-temporal data with a short observation history

Abstract.A novel method is proposed for forecasting spatial-temporal data with a short observation history sampled on a uniform grid. The method is based on spatial-temporal autoregressive modeling where the predictions of the response at the subsequent temporal layer are obtained using the response values from a recent history in a spatial neighborhood of each sampling point. Several modeling aspects such as covariance structure and sampling, as well as identification, model estimation and forecasting issues, are discussed. Extensive experimental evaluation is performed on synthetic and real-life data. The proposed forecasting models were shown capable of providing a near optimal prediction accuracy on simulated stationary spatial-temporal data in the presence of additive noise and a correlated model error. Results on a spatial-temporal agricultural dataset indicate that the proposed methods can provide useful prediction on complex real-life data with a short observation history.

Distribution comparison for site-specific regression modeling in agriculture

A novel method for problem decomposition and for local model selection in a multimodel prediction system is proposed. The proposed method partitions the data into disjoint subsets obtained by the local regression modeling and then it learns the distributions on these sets in order to identify the most appropriate regression model for each test point. The system is applied to a site specific agriculture domain and is shown to provide a substantial improvement in the prediction quality as compared to a global model. Also, some aspects of local learner choice and setting of their parameters are discussed and an overall ability of the proposed model to accurately perform regression is assessed.

A Data Generator for Evaluating Spatial Issues in Precision Agriculture

With the rapid rise in site-specific data collection, many research efforts have been directed towards finding optimal sampling and analysis procedures. However, the absence of widely available high quality precision agriculture data sets makes it difficult to compare results from separate experiments and to assess the optimality and applicability of procedures. To provide a tool for spatial data experimentation, we have developed a spatial data generator that allows users to produce data layers with given spatial properties and a response variable (e.g. crop yield) dependent upon user specified functions. Differences in response functions within fields can be simulated by assigning different models to regions in coordinate-(x and y) or feature space (multidimensional space of attributes that may have an influence on response). Noise, either unexplained variance or sensor error, can be added to all spatial layers. Sampling and interpolation error is modeled by sampling a continuous data layer and interpolating values at unsampled locations. The program has been successfully tested for up to 15000 grid points, 10 features and 5 models. As an illustration of the potential uses of generated data, the effect of sampling density and kriging interpolation on neural network prediction of crop yield was assessed. Yield prediction accuracy was highly related (correlation coefficient 0.98) to the accuracy of the interpolated layers indicating that unless data are sampled at very high densities relative to their geostatistical properties, one should not attempt to build highly accurate regression functions using interpolated data. By allowing users to generate large amounts of data with controlled complexity and features, the spatial data generator should facilitate the development of improved sampling and analysis procedures for spatial data.

Neural network-based software for fertilizer optimization in precision farming

A novel technique for providing fertilizer recommendation in precision agriculture is proposed. The method is based on the maximization of the profit function approximated using a decision support system based on artificial neural networks. The software implementation of the proposed approach is described and its use is illustrated on simulated realistic data. Experimental results suggest that the proposed technique is applicable for site-specific crop management.

Noise-resilient detection of moving objects based on spatial-temporal blocks

In this paper we discuss the resilience of moving objects detection algorithm based on spatiotemporal blocks on VQriOtLY types of ditive and multiplicative noise. After a given video is decomposed into the spariotemporal blockr, the algorithm uses dimensiondity reduction technique to obtain a compact vector representation of each block and to suppress the injluence of noise. We evaluate the algorithm performance by comparing pound truth (hand-labeled moving objects) to properly de

A Neural Network-Based Method for Site-Specific Fertilization Recommendation

ned spatial-windows based evaluation statistics. Our results on a PETS repository video show that detection and tracking of moving objects is substantially improved in presence of Gawsim, speckle, multiplicative and Poisson noise.

Effective classification of 3D image data using partitioning methods

The objective of this study is to optimize financial gain in agricultural business by nproviding site-specific fertilizer recommendations using an advanced neural network-based ndecision support system. Crop yield is considered as a function of both controllable attributes n(e.g. concentrations of various nutrients, irrigation intensity) and non-controllable attributes (e.g. nterrain attributes such as slope and profile curvature). Both direct and inverse modeling was nperformed using linear models and multi-layer neural networks with sigmoidal and radial-basis nactivation functions. Using the estimated model, profit as a function of controllable attributes nwas maximized by independent as well as by simultaneous optimization of a site-specific nfertilization rate. In independent optimization, the optimal fertilizer concentration was obtained nfor each nutrient separately. Simultaneous optimization aimed towards a global maximization of nfinancial gain as a function of all administered nutrients. Applicability of the proposed method nwas evaluated on simulated agricultural data where we were able to compare the obtained nfertilization recommendations with known optimums. Our preliminary results suggest that the nproposed direct modeling technique has a high potential for significantly increased financial gain nassuming training examples are available from all regions of the attribute space.

Learning heterogeneous functions from sparse and non-uniform samples

We propose partitioning-based methods to facilitate the classification of 3-D binary image data sets of regions of interest (ROIs) with highly non-uniform distributions. The first method is based on recursive dynamic partitioning of a 3-D volume into a number of 3-D hyper-rectangles. For each hyper-rectangle, we consider, as a potential attribute, the number of voxels (volume elements) that belong to ROIs. A hyper-rectangle is partitioned only if the corresponding attribute does not have high discriminative power, determined by statistical tests, but it is still sufficiently large for further splitting. The final discriminative hyper-rectangles form new attributes that are further employed in neural network classification models. The second method is based on maximum likelihood employing non-spatial (k-means) and spatial DBSCAN clustering algorithms to estimate the parameters of the underlying distributions. The proposed methods were experimentally evaluated on mixtures of Gaussian distributions, on realistic lesion-deficit data generated by a simulator conforming to a clinical study, and on synthetic fractal data. Both proposed methods have provided good classification on Gaussian mixtures and on realistic data. However, the experimental results on fractal data indicated that the clustering-based methods were only slightly better than random guess, while the recursive partitioning provided significantly better classification accuracy.

Image processing in precision agriculture

A boosting-based method for centers placement in radial basis function networks (RBFNs) is proposed. Also, the influence of several methods for drawing random samples on the accuracy of RBFNs is examined. The new method is compared to trivial, linear and non-linear regressors including the multilayer perceptron and alternative RBFN learning algorithms and its advantages are demonstrated for learning heterogeneous functions from sparse and non-uniform samples.

Distinguishing Among 3-D Distributions for Brain Image Data Classification

A brief review of our signal and image processing application in precision agriculture is presented. A method for determining sampling frequency for agriculture data is proposed, and some initial results based on data simulation and image processing are reported.