Crop type classification using Earth Observation (EO) data is challenging, particularly for crop types with similar phenological growth stages. In this regard, the synergy of optical and Synthetic-Aperture Radar (SAR) data enables a broad representation of biophysical and structural information on target objects, enhancing crop type mapping. However, the fusion of multi-sensor dense time-series data often comes with the challenge of high dimensional feature space. In this study, we (1) evaluate how the usage of only optical, only SAR, and their fusion affect the classification accuracy; (2) identify the combination of which time-steps and feature-sets lead to peak accuracy; (3) analyze misclassifications based on the parcel size, optical data availability, and crops’ temporal profiles. Two fusion approaches were considered and compared in this study: feature stacking and decision fusion. To distinguish the most relevant feature subsets time- and variable-wise, grouped forward feature selection (gFFS) was used. gFFS allows focusing analysis and interpretation on feature sets of interest like spectral bands, vegetation indices (VIs), or data sensing time rather than on single features. This feature selection strategy leads to better interpretability of results while substantially reducing computational expenses. The results showed that, in contrast to most other studies, SAR datasets outperform optical datasets. Similar to most other studies, the optical-SAR combination outperformed single sensor predictions. No significant difference was recorded between feature stacking and decision fusion. Random Forest (RF) appears to be robust to high feature space dimensionality. The feature selection did not improve the accuracies even for the optical-SAR feature stack with 320 features. Nevertheless, the combination of RF feature importance and time- and variable-wise gFFS rankings in one visualization enhances interpretability and understanding of the features’ relevance for specific classification tasks. For example, by enabling the identification of features that have high RF feature importance values but are, in their information content, correlated with other features. This study contributes to the growing domain of interpretable machine learning.
Information on land cover and land use at high spatial resolutions is essential for advancing earth system science as well as for environmental monitoring to support decision-making and reporting processes. In view of this, we present the first version of the DFD Land Use and Land Cover Product for Germany, DFD-LULC_DE, for the year 2014, generated from 702 Landsat-7 and Landat-8 scenes at 30m resolution. The results were derived based on a fully automated preprocessing chain that integrates data acquisition, radiometric, atmospheric and topographic correction, as well as spectral–temporal feature extraction for all Landsat surface reflectance bands, brightness temperature and various spectral indices. The classification followed a two-step approach: first, an initial classification is performed using a Random Forest classifier trained on ground truth data obtained from the LUCAS survey of EUROSTAT, followed by a semi-automated sampling of additional training data to further improve the initial classification results. Automatic selection of appropriate training samples is based on the vote entropy derived from the initial classification, thereby keeping manual user interaction low. The approach demonstrated is promising, also with respect to a European wide application, and contributes towards the advancement and enhancement of the DLR-DFD’s processing chains, which are directed towards the generation of land cover products at regular intervals being of central importance to related land monitoring and reporting services.
4 April 2013 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
A prerequisite for object-based image analysis is the generation of adequate segments. However, the parameters for the image segmentation algorithms are often manually defined. Therefore, the generation of an ideal segmentation level is usually costly and user-depended. In this paper a strategy for a semi-automatic optimization of object-based classification of multitemporal data is introduced by using Random Forest (RF) and a novel segmentation algorithm. The Superpixel Contour (SPc) algorithm is used to generate a set of different levels of segmentation, using various combinations of parameters in a user-defined range. Finally, the best parameter combination is selected based on the cross-validation-like out-of-bag (OOB) error that is provided by RF. Therefore, the quality of the parameters and the corresponding segmentation level can be assessed in terms of the classification accuracy, without providing additional independent test data. To evaluate the potential of the proposed concept, we focus on land cover classification of two study areas, using multitemporal RapidEye and SPOT 5 images. A classification that is based on eCognition's widely used multiresolution segmentation algorithm (MRS) is used for comparison. Experimental results underline that the two segmentation algorithms SPc and MRS perform similar in terms of accuracy and visual interpretation. The proposed strategy that uses the OOB error for the selection of the ideal segmentation level provides similar classification accuracies, when compared to the results achieved by manual-based image segmentation. Overall, the proposed strategy is operational and easy to handle and thus economizes the findings of optimal segmentation parameters for the Superpixel Contour algorithm.