Christoph Römer

Early drought stress detection in cereals: simplex volume maximisation for hyperspectral image analysis

Early water stress recognition is of great relevance in precision plant breeding and production. Hyperspectral imaging sensors can be a valuable tool for early stress detection with high spatio-temporal resolution. They gather large, high dimensional data cubes posing a significant challenge to data analysis. Classical supervised learning algorithms often fail in applied plant sciences due to their need of labelled datasets, which are difficult to obtain. Therefore, new approaches for unsupervised learning of relevant patterns are needed. We apply for the first time a recent matrix factorisation technique, simplex volume maximisation (SiVM), to hyperspectral data. It is an unsupervised classification approach, optimised for fast computation of massive datasets. It allows calculation of how similar each spectrum is to observed typical spectra. This provides the means to express how likely it is that one plant is suffering from stress. The method was tested for drought stress, applied to potted barley plants in a controlled rain-out shelter experiment and to agricultural corn plots subjected to a two factorial field setup altering water and nutrient availability. Both experiments were conducted on the canopy level. SiVM was significantly better than using a combination of established vegetation indices. In the corn plots, SiVM clearly separated the different treatments, even though the effects on leaf and canopy traits were subtle.

A review of advanced machine learning methods for the detection of biotic stress in precision crop protection

Effective crop protection requires early and accurate detection of biotic stress. In recent years, remarkable results have been achieved in the early detection of weeds, plant diseases and insect pests in crops. These achievements are related both to the development of non-invasive, high resolution optical sensors and data analysis methods that are able to cope with the resolution, size and complexity of the signals from these sensors. Several methods of machine learning have been utilized for precision agriculture such as support vector machines and neural networks for classification (supervised learning); k-means and self-organizing maps for clustering (unsupervised learning). These methods are able to calculate both linear and non-linear models, require few statistical assumptions and adapt flexibly to a wide range of data characteristics. Successful applications include the early detection of plant diseases based on spectral features and weed detection based on shape descriptors with supervised or unsupervised learning methods. This review gives a short introduction into machine learning, analyses its potential for precision crop protection and provides an overview of instructive examples from different fields of precision agriculture.

Automatic classification of building types in 3D city models

This article presents a classifier based on Support Vector Machines (SVMs), an advanced machine learning method for semantic enrichment of coarse 3D city models by deriving the building type. The information on the building type (detached building, terraced building, etc.) is essential for a variety of relevant applications of 3D city models like spatial marketing, real estate management and marketing, and for visualization. The derivation of the building type from coarse data (mainly 2D footprints, building heights and functions) seems impossible at first sight. However it succeeds by incorporating the spatial context of a building. Since the input data can be derived easily and at very low cost, this method is widely applicable. Nevertheless, as with all supervised learning algorithms, obtaining labelled training data is very time-consuming. Herewith, we provide a method which uses outlier detection and clustering methods to support users in efficiently and rapidly obtaining adequate training data.

“Deep Phenotyping” of Early Plant Response to Abiotic Stress Using Non-invasive Approaches in Barley

The basic mechanisms of yield maintenance under drought conditions are far from being understood. Pre-symptomatic water stress recognition would help to get insides into complex plant mechanistic basis of plant response when confronted to water shortage conditions and is of great relevance in precision plant breeding and production. The plant reactions to drought stress result in spatial, temporal and tissue-specific pattern changes which can be detected using non-invasive sensor techniques, such as hyperspectral imaging. The “response turning time-point” in the temporal curve of plant response to stress rather than the maxima is the most relevant time-point for guided sampling to get insights into mechanistic basis of plant response to drought stress. Comparative hyperspectral image analysis was performed on barley (Hordeum vulgare) plants grown under well-watered and water stress conditions in two consecutive years. The obtained massive, high-dimensional data cubes were analysed with a recent matrix factorization technique based on simplex volume maximization of hyperspectral data and compared to several drought-related traits. The results show that it was possible to detect and visualize the accelerated senescence signature in stressed plants earlier than symptoms become visible by the naked eye.

Landcover classification with self-taught learning on archetypal dictionaries

This paper introduces archetypal dictionaries for a self-taught learning framework for the application of landcover classification. Self-taught learning, an unsupervised representation learning method, is exploited to learn low-dimensional and discriminative higher-level features, which are used as input into a classification algorithm. Experiments are conducted using a multi-spectral Landsat 5 TM image of a study area in the north of Novo Progresso located in South America. Our results confirm that self-taught learning with archetypal dictionaries provide features, which can be used as input into a linear logistic regression classifier. The obtained classification accuracies are comparable to kernel-based classifier using the original features.

Feeding the World with Big Data: Uncovering Spectral Characteristics and Dynamics of Stressed Plants

Modern communication, sensing, and actuator technologies as well as methods from signal processing, pattern recognition, and data mining are increasingly applied in agriculture, ultimately helping to meet the challenge of “How to feed a hungry world?” Developments such as increased mobility, wireless networks, new environmental sensors, robots, and the computational cloud put the vision of a sustainable agriculture for anybody, anytime, and anywhere within reach. Unfortunately, data-driven agriculture also presents unique computational problems in scale and interpretability: (1) Data is gathered often at massive scale, and (2) researchers and experts of complementary skills have to cooperate in order to develop models and tools for data intensive discovery that yield easy-to-interpret insights for users that are not necessarily trained computer scientists. On the problem of mining hyperspectral images to uncover spectral characteristic and dynamics of drought stressed plants, we showcase that both challenges can be met and that big data mining can—and should—play a key role for feeding the world, while enriching and transforming data mining.

Optimalwavelengths for an early identification of Cercospora beticola with Support Vector Machines based on hyperspectral reflection data

Automatic classification of plant diseases at an early stage is vital for precision crop protection. Our aim was to identify sugar beet leaves inoculated with Cercospora beticola before symptoms are visible. Therefore hyperspectral reflection between 400 and 1050 nm was observed. Relevant wavelengths have to be found in order to implement practical sensor systems with reduced development costs. The main contribution of this study is to identify a minimal subset which is sufficient for separating healthy and inoculated leaves. The heuristic of Hall which analyses the relevance of a feature subset considering the intercorrelation among the features was applied. In order to select a good subset in a reasonable amount of time a genetic algorithm was used. This way enabled a subset of only seven out of 462 wavelengths, which nevertheless enabled us to identify low disease severity ≤ 5% with a classification accuracy of 84.3%. Disease severity above 5% was classified with 99.8%.