Andreas Rabe

Sensitivity of Support Vector Machines to Random Feature Selection in Classification of Hyperspectral Data

The accuracy of supervised land cover classifications depends on factors such as the chosen classification algorithm, adequate training data, the input data characteristics, and the selection of features. Hyperspectral imaging provides more detailed spectral and spatial information on the land cover than other remote sensing resources. Over the past ten years, traditional and formerly widely accepted statistical classification methods have been superseded by more recent machine learning algorithms, e.g., support vector machines (SVMs), or by multiple classifier systems (MCS). This can be explained by limitations of statistical approaches with regard to high-dimensional data, multimodal classes, and often limited availability of training data. In the presented study, MCSs based on SVM and random feature selection (RFS) are applied to explore the potential of a synergetic use of the two concepts. We investigated how the number of selected features and the size of the MCS influence classification accuracy using two hyperspectral data sets, from different environmental settings. In addition, experiments were conducted with a varying number of training samples. Accuracies are compared with regular SVM and random forests. Experimental results clearly demonstrate that the generation of an SVM-based classifier system with RFS significantly improves overall classification accuracy as well as producers and users accuracies. In addition, the ensemble strategy results in smoother, i.e., more realistic, classification maps than those from stand-alone SVM. Findings from the experiments were successfully transferred onto an additional hyperspectral data set.

The EnMAP Spaceborne Imaging Spectroscopy Mission for Earth Observation

Imaging spectroscopy, also known as hyperspectral remote sensing, is based on the characterization of Earth surface materials and processes through spectrally-resolved measurements of the light interacting with matter. The potential of imaging spectroscopy for Earth remote sensing has been demonstrated since the 1980s. However, most of the developments and applications in imaging spectroscopy have largely relied on airborne spectrometers, as the amount and quality of space-based imaging spectroscopy data remain relatively low to date. The upcoming Environmental Mapping and Analysis Program (EnMAP) German imaging spectroscopy mission is intended to fill this gap. An overview of the main characteristics and current status of the mission is provided in this contribution. The core payload of EnMAP consists of a dual-spectrometer instrument measuring in the optical spectral range between 420 and 2450 nm with a spectral sampling distance varying between 5 and 12 nm and a reference signal-to-noise ratio of 400:1 in the visible and near-infrared and 180:1 in the shortwave-infrared parts of the spectrum. EnMAP images will cover a 30 km-wide area in the across-track direction with a ground sampling distance of 30 m. An across-track tilted observation capability will enable a target revisit time of up to four days at the Equator and better at high latitudes. EnMAP will contribute to the development and exploitation of spaceborne imaging spectroscopy applications by making high-quality data freely available to scientific users worldwide.

imageRF - A user-oriented implementation for remote sensing image analysis with Random Forests

An IDL implementation for the classification and regression analysis of remote sensing images with Random Forests is introduced. The tool, called imageRF, is platform and license independent and uses generic image file formats. It works well with default parameterization, yet all relevant parameters can be defined in intuitive GUIs. This makes it a user-friendly image processing tool, which is implemented as an add-on in the free EnMAP-Box and may be used in the commercial IDL/ENVI software.

The EnMAP-Box—A Toolbox and Application Programming Interface for EnMAP Data Processing

The EnMAP-Box is a toolbox that is developed for the processing and analysis of data acquired by the German spaceborne imaging spectrometer EnMAP (Environmental Mapping and Analysis Program). It is developed with two aims in mind in order to guarantee full usage of future EnMAP data, i.e., (1) extending the EnMAP user community and (2) providing access to recent approaches for imaging spectroscopy data processing. The software is freely available and offers a range of tools and applications for the processing of spectral imagery, including classical processing tools for imaging spectroscopy data as well as powerful machine learning approaches or interfaces for the integration of methods available in scripting languages. A special developer version includes the full open source code, an application programming interface and an application wizard for easy integration and documentation of new developments. This paper gives an overview of the EnMAP-Box for users and developers, explains typical workflows along an application example and exemplifies the concept for making it a frequently used and constantly extended platform for imaging spectroscopy applications.

A Comparison of Advanced Regression Algorithms for Quantifying Urban Land Cover

Quantitative methods for mapping sub-pixel land cover fractions are gaining increasing attention, particularly with regard to upcoming hyperspectral satellite missions. We evaluated five advanced regression algorithms combined with synthetically mixed training data for quantifying urban land cover from HyMap data at 3.6 and 9 m spatial resolution. Methods included support vector regression (SVR), kernel ridge regression (KRR), artificial neural networks (NN), random forest regression (RFR) and partial least squares regression (PLSR). Our experiments demonstrate that both kernel methods SVR and KRR yield high accuracies for mapping complex urban surface types, i.e., rooftops, pavements, grass- and tree-covered areas. SVR and KRR models proved to be stable with regard to the spatial and spectral differences between both images and effectively utilized the higher complexity of the synthetic training mixtures for improving estimates for coarser resolution data. Observed deficiencies mainly relate to known problems arising from spectral similarities or shadowing. The remaining regressors either revealed erratic (NN) or limited (RFR and PLSR) performances when comprehensively mapping urban land cover. Our findings suggest that the combination of kernel-based regression methods, such as SVR and KRR, with synthetically mixed training data is well suited for quantifying urban land cover from imaging spectrometer data at multiple scales.

Simplifying Support Vector Machines for classification of hyperspectral imagery and selection of relevant features

Support Vector Machines (SVM) for image classification proved to perform well in many applications. However, they are often not preferred in hyperspectral image analysis due to long processing times caused by a high number of support vectors and large data sets. We present two approaches that speed-up the classification process with SVM by a) simplifying the original SVM, i.e. reducing the number of support vectors, and b) reducing the number of features by selecting relevant, non-redundant features. Results for three classification problems are shown. By applying the two approaches, we observe reduction rates a) between 9.1% and 27.2% for the number of support vectors and b) from 86.8% to 93.0% of features, both without significant decreases in classification accuracy. This enables a fast mapping of complete hyperspectral scenes.

Impact of different morphological profiles on the classification accuracy of urban hyperspectral data

We present a detailed study on the classification of urban hyperspectral data with morphological profiles (MP). Although such a spectral-spatial classification approach may significantly increase achieved accuracy, the computational complexity as well as the increased dimensionality and redundancy of such data sets are potential drawbacks. This can be overcome by feature selection. Moreover it is useful to derive detailed information on the contribution of different components from MP to the classification accuracy by evaluating these subsets. We apply a wrapper approach for feature selection based on support vector machines (SVM) with sequential feature forward selection (FFS) search strategy to two hyperspectral data sets that contain the first principal components (PC) and various corresponding MP from an urban area. In doing so, we identify feature subsets of increasing size that perform best in terms of kappa for the given setup. Results clearly demonstrate that maximum classification accuracies are achieved already on small feature subsets with few morphological profiles.

Simplifying Support Vector Machines for Regression analysis of hyperspectral imagery

Support Vector Machines for Regression (SVR) proved to perform well. However, they are not preferred in image analysis due to a high number of needed support vectors (SV) and consequently long processing times. We present a method for simplifying the original SVR regression function up to a user-specified degree of accepted performance decrease. We show results for two regression problems: modelling leaf area index and dry vegetation mixing fraction using simulated hyperspectral EnMAP data. In both cases, SVR demonstrate high potential for modelling complex dependencies between hyperspectral reflectance and quantitative targets. By simplifying the original SVR, we observed reduction rates in number of SV in the 86% to 95% range for acceptable degrees of approximation quality. This enables a fast mapping of complete EnMAP scenes.

Optimisation of Savannah Land Cover Characterisation with Optical and SAR Data

Accurately mapping savannah land cover at the regional scale can provide useful input to policy decision making efforts regarding, for example, bush control or overgrazing, as well as to global carbon emissions models. Recent attempts have employed Earth observation data, either from optical or radar sensors, and most commonly from the dry season when the spectral difference between woody vegetation, crops and grasses is maximised. By far the most common practice has been the use of Landsat optical bands, but some studies have also used vegetation indices or SAR data. However, conflicting reports with regards to the effectiveness of the different approaches have emerged, leaving the respective land cover mapping community with unclear methodological pathways to follow. We address this issue by employing Landsat and Advanced Land Observing Satellite Phased Array type L-band Synthetic Aperture Radar (ALOS PALSAR) data to assess the accuracy of mapping the main savannah land cover types of woody vegetation, grassland, cropland and non-vegetated land. The study area is in southern Africa, covering approximately 44,000 km2. We test the performance of 15 different models comprised of combinations of optical and radar data from the dry and wet seasons. Our results show that a number of models perform well and very similarly. The highest overall accuracy is achieved by the model that incorporates both optical and synthetic-aperture radar (SAR) data from both dry and wet seasons with an overall accuracy of 91.1% (±1.7%): this is almost a 10% improvement from using only the dry season Landsat data (81.7 ± 2.3%). The SAR-only models were capable of mapping woody cover effectively, achieving similar or lower omission and commission errors than the optical models, but other classes were detected with lower accuracies. Our main conclusion is that the combination of metrics from different sensors and seasons improves results and should be the preferred methodological pathway for accurate savannah land cover mapping, especially now with the availability of Sentinel-1 and Sentinel-2 data. Our findings can provide much needed assistance to land cover monitoring efforts to savannahs in general, and in particular to southern African savannahs, where a number of land cover change processes have been related with the observed land degradation in the region.