Matt Aitkenhead

Weed and crop discrimination using image analysis and artificial intelligence methods

Abstract Development of a visual method of discriminating between crop seedlings and weeds is an important and necessary step towards the automation of non-chemical weed control systems in agriculture, and towards the reduction in chemical use through spot spraying. Two methods were applied to recognise carrot ( Daucus carota L.) seedlings from those of ryegrass ( Lolium perenne ) and Fat Hen ( Chenopodium album ) using digital imaging. The first method involved the use of a simple morphological characteristic measurement of leaf shape (perimeter 2 /area), which had varying effectiveness (between 52 and 74%) in discriminating between the two types of plant, with the variation dependent on plant size. The second involved a self-organising neural network more biologically plausible than many commonly used NN methods. While the latter did not give results as good as those required for commercial purposes, it showed that a neural network-based methodology exists which allows the system to learn and discriminate between species to an accuracy exceeding 75% without predefined plant descriptions being necessary.

A co-evolving decision tree classification method

Decision tree classification provides a rapid and effective method of categorising datasets. Many algorithmic methods exist for optimising decision tree structure, although these can be vulnerable to changes in the training dataset. An evolutionary method is presented which allows decision tree flexibility through the use of co-evolving competition between the decision tree and the training data set. This method is tested using two different datasets and gives results comparable with or superior to other classification methods. A final discussion argues for the utility of decision trees over algorithmic or other alternative methods such as neural networks, particularly in situations where a large number of variables are being considered.

A neural network face recognition system

Abstract A neural network based facial recognition program (FADER—FAce DEtection and Recognition) was developed and tested. The hardware and software components were all standard commercial design, allowing the system to be built for minimal cost. Using a set of 1000 face and 1000 ‘no-face’ images, we achieved 94.7% detection rate, and a 0.6% false positive rate. Three different neural network models were applied to face recognition, using single images of each subject to train the system. A novel adaptation of the Hebbian connection strength adjustment model gave the best results, with 74.1% accuracy achieved. Each of the systems components, including an intermediate substructure detection network, was subject to evolutionary computation in order to optimise the system performance.

A bioinformatic and transcriptomic approach to identifying positional candidate genes without fine mapping: an example using rice root-growth QTLs

Fine mapping can accurately identify positional candidate genes for quantitative trait loci (QTLs) but can be time consuming, costly, and, for small-effect QTLs with low heritability, difficult in practice. We propose an alternative approach, which uses meta-analysis of original mapping data to produce a relatively small confidence interval for target QTLs, lists the underlying positional candidates, and then eliminates them using whole-genome transcriptomics. Finally, sequencing is conducted on the remaining candidate genes allowing identification of allelic variation in either expression or protein sequence. We demonstrate the approach using root-growth QTLs on chromosomes 2, 5, and 9 of the Bala x Azucena rice mapping population. Confidence intervals of 10.5, 9.6, and 5.4 cM containing 189, 322, and 81 genes, respectively, were produced. Transcriptomics eliminated 40% of candidate genes and identified nine expression polymorphisms. Sequencing of 30 genes revealed that 57% of the predicted proteins were polymorphic. The limitations of this approach are discussed.

A novel method for training neural networks for time-series prediction in environmental systems

Abstract Soil, streamwater and climatic variables were measured hourly over several month periods in two situations in North-East (NE) Scotland, using data loggers and other measuring instruments. One of the locations was on agricultural land near Inverness while the other was at an acidic peat moorland site in the River Dee catchment. The data sets were used to train neural networks using three different methods, including a novel, biologically plausible system. Temporal pattern recognition capabilities using each method were investigated. The novel method proved equally capable in predicting future variable values using large data sets as the other two methods. An argument is made for this method, termed the ‘Local Interaction’ method, providing valid competition to other neural network and statistical methods in the detection of patterns and prediction of events in complex biological systems.

Modelling DOC export from watersheds in Scotland using neural networks

A wide variety of watershed-scale attributes can be used as predictors of the export of dissolved organic carbon (DOC) from a watershed. However, the complexity and number of relationships makes the development of generally applicable mechanistic models for prediction of DOC export based on measurement of factors difficult. Here we have applied neural network modelling methods to the prediction of stream flux and daily DOC export from several watersheds of varying size within the Dee valley, in north-east Scotland. A two-stage process was carried out in which first a model was developed which used a large number of variables thought to be relevant to DOC export, and then the possibility of using a restricted set of variables was investigated in order to reduce the amount of analysis required in order to produce accurate DOC export predictions. The results showed that it is possible to predict DOC export using input variables corresponding broadly to the factors responsible for soil formation, and that a single sample site may provide enough information to allow prediction for an entire watershed. However, in order to achieve a model with statistically significant results, it is necessary to use multiple sample sites per watershed, and to use measured rather than modelled flow values. Discussion is made of the effectiveness of the neural network method in developing models of DOC export, and of problems with the method (particularly in the inability to use NN models for process-based models).

Classification of Landsat Thematic Mapper imagery for land cover using neural networks

Landsat Thematic Mapper (TM) imagery can be used to classify different land cover types based on reflectance and emittance characteristics in seven wavelength bands. Various methods, including NDVI and other simple mathematical transformations, can be used to show strong variations in band intensity ratios from different surfaces. However, the number of land cover classes used is commonly low, preventing a detailed mapping of the region of interest. A neural network trained with the backpropagation method should be able to improve on these simple mathematical calculations by developing complex functions which allow recognition of different land cover or land use types. Landsat imagery of Aberdeen and the surrounding area was used to develop a land cover map highlighting areas of residential, commercial and industrial land use, along with various natural and semi‐natural land cover classes. Confusion between specific classes is highlighted by the use of a Kohonen self‐organizing map to categorize the Landsat multispectral imagery, resulting in a description of the land cover categories that can actually be distinguished from one another using Landsat TM imagery.

Gypsophile chemistry unveiled: Fourier transform infrared (FTIR) spectroscopy provides new insight into plant adaptations to gypsum soils

Gypsum soils are among the most restrictive and widespread substrates for plant life. Plants living on gypsum are classified as gypsophiles (exclusive to gypsum) and gypsovags (non-exclusive to gypsum). The former have been separated into wide and narrow gypsophiles, each with a putative different ecological strategy. Mechanisms displayed by gypsum plants to compete and survive on gypsum are still not fully understood. The aim of this study was to compare the main chemical groups in the leaves of plants with different specificity to gypsum soils and to explore the ability of Fourier transform infrared (FTIR) spectra analyzed with neural network (NN) modelling to discriminate groups of gypsum plants. Leaf samples of 14 species with different specificity to gypsum soils were analysed with FTIR spectroscopy coupled to neural network (NN) modelling. Spectral data were further related to the N, C, S, P, K, Na, Ca, Mg and ash concentrations of samples. The FTIR spectra of the three groups analyzed showed distinct features that enabled their discrimination through NN models. Wide gypsophiles stood out for the strong presence of inorganic compounds in their leaves, particularly gypsum and, in some species, also calcium oxalate crystals. The spectra of gypsovags had less inorganic chemical species, while those of narrow gypsum endemisms had low inorganics but shared with wide gypsophiles the presence of oxalate. Gypsum and calcium oxalate crystals seem to be widespread amongst gypsum specialist plants, possibly as a way to tolerate excess Ca and sulphate. However, other mechanisms such as the accumulation of sulphates in organic molecules are also compatible with plant specialization to gypsum. While gypsovags seem to be stress tolerant plants that tightly regulate the uptake of S and Ca, the ability of narrow gypsum endemisms to accumulate excess Ca as oxalate may indicate their incipient specialization to gypsum.

Mapping land cover from detailed aerial photography data using textural and neural network analysis

Automated mapping of land cover using black and white aerial photographs, as an alternative method to traditional photo‐interpretation, requires using methods other than spectral analysis classification. To this end, textural measurements have been shown to be useful indicators of land cover. In this work, a neural network model is proposed and tested to map historical land use/land cover (LUC) from very detailed panchromatic aerial photographs (5 m resolution) using textural measurements. The method is used to identify different land use and management types (e.g. traditional versus mechanized vineyard systems). These have been tested with known ground reference data. The results show the potential of the methodology to obtain automatic, historic, and very detailed cartography information from a complex landscape such as the mountainous and Mediterranean region to which it is applied here, and the advantages that this method has over traditional methods.

Improving land-cover classification using recognition threshold neural networks

The use of neural networks to classify land-cover from remote sensing imagery relies on the ability to determine a winner from the candidate land-cover types based on the imagery information available. In the case of a “winnertakes-all” scenario, this does not allow us a measure of how much the prediction of each pixel’s land-cover can be trusted. We present a three-stage method where only winning candidates which are given a clear lead over the other land-cover types are accepted, with a neighborhood relationship and the application of mixed pixels being used to provide full classification. This method allows us to place more faith in the resulting map than simply taking the winner, and results in a higher accuracy of classification. The method is applied to Landsat imagery of an area of the Philippines where natural, urban, and cultivated land-cover types exist.