Many environmental data are inherently imbalanced, with some majority land use and land cover types dominating over rare ones. In cultivated ecosystems minority classes are often the target as they might indicate a beginning land use change. Most standard classifiers perform best on a balanced distribution of classes, and fail to detect minority classes. We used the synthetic minority oversampling technique (smote) with Random Forest to classify land cover classes in a small agricultural catchment in South Korea using modis time series. This area faces a major soil erosion problem and policy measures encourage farmers to replace annual by perennial crops to mitigate this issue. Our major goal was therefore to improve the classification performance on annual and perennial crops. We compared four different classification scenarios on original imbalanced and synthetically oversampled balanced data to quantify the effect of smote on classification performance. smote substantially increased the true positive rate of all oversampled minority classes. However, the performance on minor classes remained lower than on the majority class. We attribute this result to a class overlap already present in the original data set that is not resolved by smote. Our results show that resampling algorithms could help to derive more accurate land use and land cover maps from freely available data. These maps can be used to provide information on the distribution of land use classes in heterogeneous agricultural areas and could potentially benefit decision making.

No abstract

A representation of the world as a 3D model is a common necessity in robotics and automation. In previous work, we developed a concept to generate boundary representation (B-Rep) models from multiple point clouds using a hand-held depth-camera and to register them without a prior known pose. During the online reconstruction, properties of the sensor and the system (like noise) lead to small holes in the B-Rep. To prevent tedious post-processing, holes should be closed during the reconstruction. Our goal is to automatically close identified holes. However, not every hole can be closed automatically, as it may be unreasonable. For this case we develop a visual indication for the user, so he can close the hole by recording another depth image. In an experimental validation, we conclude the usefulness of the addition to the system.

A complete object database containing a model (representing geometric and texture information) of every possible workpiece is a common necessity e.g. for different object recognition or task planning approaches. The generation of these models is often a tedious process. In this paper we present a fully automated approach to tackle this problem by generating complete workpiece models using a robotic manipulator. A workpiece is recorded by a depth sensor from multiple views for one side, then turned, and captured from the other side. The resulting point clouds are merged into one complete model. Additionally, we represent the information provided by the object’s texture using keypoints. We present a proof of concept and evaluate the precision of the final models. In the end we conclude the usefulness of our approach showing a precision of around 1 mm for the resulting models.

No abstract