Dimitrios Moshou

Early Disease Detection in Wheat Fields using Spectral Reflectance

The difference in spectral reflectance between healthy and diseased wheat plants infected with Puccinia striiformis (yellow rust) was investigated. In-field spectral images were taken with a spectrograph mounted at spray boom height. A normalisation method based on reflectance and illumination adjustments was applied. To consider the entire canopy reflection, a spatially moving average was introduced. A classification model based on quadratic discrimination was built on a selected group of wavebands obtained by stepwise variable selection. Through this method, confusion rates dropped from 12 to 4% error classification, based on four different wavebands. These results are very encouraging for the development of a cost-effective optical device for recognising diseases, such as yellow rust, in the field in early spring.

A neural network based plant classifier

The Self-Organizing Map (SOM) neural network is used in a supervised way for a classification task. The neurons of the SOM are extended with local linear mappings. Error information obtained during training is used in a novel learning algorithm to train the classifier. The proposed method achieves fast convergence and good generalization. The classification method is then applied in a precision farming application, the classification of crops and weeds using spectral properties. The proposed method compares favorably with an optimal Bayesian classifier that is presented in the form of a probabilistic neural network. The classification performance of the proposed method is proven superior compared with other statistical and neural classifiers.

Simultaneous identification of plant stresses and diseases in arable crops using proximal optical sensing and self-organising maps

The objective of this research was to detect plant stress caused by disease infestation and to discriminate this type of stress from nutrient deficiency stress in field conditions using spectral reflectance information. Yellow Rust infected winter wheat plants were compared to nutrient stressed and healthy plants. In-field hyperspectral reflectance images were taken with an imaging spectrograph. A normalisation method based on reflectance and light intensity adjustments was applied. For achieving high performance stress identification, Self-Organising Maps (SOMs) and Quadratic Discriminant Analysis (QDA) were introduced. Winter wheat infected with Yellow Rust was successfully recognised from nutrient stressed and healthy plants. Overall performance using five wavebands was more than 99%.

Dynamic muscle fatigue detection using self-organizing maps

Wavelets are used for the processing of signals that are non-stationary and time varying. The electromyogram (EMG) contains transient signals related to muscle activity. Wavelet coefficients are proposed as features for identifying muscle fatigue. By observing the approximation coefficients it is shown that their amplitude follows closely the muscle fatigue development. The proposed method for detecting fatigue is automated by using neural networks. The self-organizing map (SOM) has been used to visualize the variation of the approximation wavelet coefficients and aid the detection of muscle fatigue. The results show that a 2D SOM separates EMG signatures from fresh and fatigued muscles, thus providing a visualization of the onset of fatigue over time. The map is able to detect if muscles have recovered temporarily. The system is adaptable to different subjects and conditions since the techniques used are not subject or workload regime specific.

Apple mealiness detection using fluorescence and self-organising maps

The chlorophyll fluorescence kinetics of ‘Jonagold’ and ‘Cox’ apples, stored under different conditions to induce mealiness, were measured. Three different storage conditions were considered causing three mealiness levels: not mealy, moderately and strongly mealy. Also destructive measurements of the texture (firmness, hardness, juice content and soluble solids content) were done. Classification into different mealiness levels based on the fluorescence measurements was more performant than a classification based on the destructive measurements. To estimate the mealiness level in a non-destructive way from the fluorescence features, a number of different classifiers were constructed. Quadratic discriminants and supervised and unsupervised neural networks were tested and compared. The self-organising map gives promising results when compared with the multi-layer perceptrons and quadratic discriminant analysis. The different advantages of the constructed classifiers suggest that fluorescence can be used in an automatic sorting line to assess certain types of mealiness.

A Weed Species Spectral Detector Based on Neural Networks

A new neural network architecture for classification purposes is proposed. The Self-Organizing Map (SOM) neural network is used in a supervised way for a classification task. The neurons of the SOM become associated with local linear mappings (LLM). Error information obtained during training is used in a novel learning algorithm to train the classifier. The proposed method achieves fast convergence and good generalization. The classification method is then applied in a precision farming application, the classification of crops and different kinds of weeds by using spectral reflectance measurements. The classification performance of the proposed method is proven superior compared to other neural classifiers. Also, the proposed method compares favorably with the results obtained by using an optimal Bayesian classifier.

Wheat yield prediction using machine learning and advanced sensing techniques

A yield potential prediction model was developed and evaluated for winter wheat.The soil parameters were estimated through online soil spectroscopy with a prototype sensor.Main inputs for yield potential prediction were estimated soil parameters and remote sensing vegetation indices.The proposed architecture provided visual information about the factors affecting the yield potential. Understanding yield limiting factors requires high resolution multi-layer information about factors affecting crop growth and yield. Therefore, on-line proximal soil sensing for estimation of soil properties is required, due to the ability of these sensors to collect high resolution data (>1500 sample per ha), and subsequently reducing labor and time cost of soil sampling and analysis. The aim of this paper is to predict within field variation in wheat yield, based on on-line multi-layer soil data, and satellite imagery crop growth characteristics. Supervised self-organizing maps capable of handling existent information from different soil and crop sensors by utilizing an unsupervised learning algorithm were used. The performance of counter-propagation artificial neural networks (CP-ANNs), XY-fused Networks (XY-Fs) and Supervised Kohonen Networks (SKNs) for predicting wheat yield in a 22ha field in Bedfordshire, UK were compared for a single cropping season. The self organizing models consisted of input nodes corresponded to feature vectors formed from normalized values of on-line predicted soil parameters and the satellite normalized difference vegetation index (NDVI). The output nodes consisted of yield isofrequency classes, which were predicted from the three trained networks. Results showed that cross validation based yield prediction of the SKN model for the low yield class exceeded 91% which can be considered as highly accurate given the complex relationship between limiting factors and the yield. The medium and high yield class reached 70% and 83% respectively. The average overall accuracy for SKN was 81.65%, for CP-ANN 78.3% and for XY-F 80.92%, showing that the SKN model had the best overall performance.

Neural recognition systems for swine cough

Coughing is one of the most frequent presenting symptoms of many diseases affecting the airways and the lungs of humans and animals. The aim of this paper is to build up an intelligent alarm system that can be used for the early detection of cough sounds in piggeries. Registration of coughs from different pigs in a metallic chamber was done in order to analyse the acoustical signal. A new approach is presented to distinguish cough sounds from other sounds like grunts, metal clanging and noise using neural networks (NN) as classification method. Other signals (grunts, metal clanging, etc.) could also be detected. Self-organising maps are used for visualisation of data relationships. Several types of NN are compared with statistical methods for the classification of the cough sounds. The early detection of coughs can be used for the construction of an intelligent alarm that can inform about the presence of a possible viral infection.

Evaluation of UAV imagery for mapping Silybum marianum weed patches

ABSTRACT Silybum marianum (L.) Gaertn weed has the tendency to grow in patches. In order to perform site-specific weed management, determining the spatial distribution of weeds is important for its eradication. Remote sensing has been used to perform species discrimination and it offers promising techniques for operational weed mapping. In the present study, the feasibility of high-resolution imaging for S. marianum detection and mapping is reported. A multispectral camera (green–red–near-infrared) mounted on a fixed wing unmanned aerial vehicle (UAV) was used for the acquisition of high-resolution images with pixel size of 0.1 m. The maximum likelihood (ML) classifier was used to classify the S. marianum among other weed species present in a field, with Avena sterilisL. being predominant. As input to the classifier, the three spectral bands and the texture were used. The scale of the mapping was varied by degrading the image resolution to evaluate classification performance, with 1 m resolution providing the highest classification accuracy. The classification rates obtained using ML reached an overall accuracy of 87.04% with a K-hat statistic of 74%. The results prove the feasibility of operational S. marianum mapping using UAV and multispectral imaging.

An intelligent alarm for early detection of swine epidemics based on neural networks

Coughing is one of the most frequent presenting symptoms of many diseases affecting the airways and the lungs of humans and animals. The aim of this research is to build an intelligent alarm system that can be used for the early detection of cough sounds in pig houses. Registration of coughs from different pigs in a metallic chamber was done in order to analyze the acoustical signal. A new approach is presented to distinguish cough sounds from other sounds like grunts, metal clanging, and noise using neural network classification methods. Other signals (grunts, metal clanging, etc.) could also be detected. A hybrid classifier is proposed that achieves the highest classification accuracy in both the off-line and the on-line detection of coughs and other sounds. The best correct classification performance was obtained with a hybrid classifier that classified coughs and metal clanging separately from other sounds, giving better results compared to a multi-layer perceptron alone. The hybrid classifier, which consisted of a 2-class probabilistic neural network and a 4-class multi-layer perceptron, gave high discrimination performance in the case of grunts and noise (91.3% and 91.3% respectively) and a performance of 94.8% for correct classification in the case of coughs. The early detection of coughs can be used for the construction of an intelligent alarm that can signal the presence of a possible viral infection so that early treatment can be implemented.