Björn Waske

The agricultural landscapes of Germany have undergone significant intensification and land consolidation. Both contribute to the decline of non-productive landscape elements, such as hedgerows, tree rows, field margins, and ditches. Mapping these elements is essential for monitoring policy targets, yet comprehensive cadastral data is lacking. This study investigates whether landscape elements in agricultural areas in Germany can be mapped using freely available multispectral Sentinel-2 imagery. First, we generate a nationwide, least-vegetated image composite for the 2023 and 2024 vegetation periods. For each pixel, we select the observation with the maximum bare soil index from available Sentinel-2 scenes. This maximizes the contrast between permanently vegetated landscape elements and the surrounding arable land. We fine-tuned a transformer-based semantic segmentation model (SegFormer MiT-B4) on this data and compared it with a proposed added refinement head designed to promote the spatial continuity of linear structures. We validate the model on independent data, covering various terrestrial landscapes in Germany. The refinement head improves the model's overall performance (macro IoU: 0.63 to 0.65), with the largest gains observed for woody elements (IoU: 0.67 to 0.71). However, improvements for herbaceous elements are marginal (IoU: 0.58 to 0.59). Qualitative maps and a 1 km² hexagon aggregation demonstrate the model's applicability at the national scale and reveal pronounced regional patterns in landscape element density and composition.

Earth observation missions such as Sentinel and Landsat support the large-scale identification of agricultural crops by providing free radar and multispectral satellite images. The extraction of representative image information as well as the combination of different image sources for improved feature selection still represent a major challenge in the field of remote sensing. In this paper, we propose a novel three-dimensional (3D) deep learning U-Net model to fuse multi-level image features from multispectral and synthetic aperture radar (SAR) time series data for seasonal crop-type mapping at a regional scale. For this purpose, we used a dual-stream U-Net with a 3D squeeze-and-excitation fusion module applied at multiple stages in the network to progressively extract and combine multispectral and SAR image features. Additionally, we introduced a distinctive method for generating patch-based multitemporal multispectral composites by selective image sampling within a 14-day window, prioritizing those with minimal cloud cover. The classification results showed that the proposed network provided the best overall accuracy (94.5%) compared to conventional two-dimensional (2D) and three-dimensional U-Net models (2D: 92.6% and 3D: 94.2%). Our network successfully learned multi-modal dependencies between the multispectral and SAR satellite images, leading to improved field mapping of spectrally similar and heterogeneous classes while mitigating the limitations imposed by persistent cloud coverage. Additionally, the feature representations extracted by the proposed network demonstrated their transferability to a new cropping season, providing a reliable mapping of spatio-temporal crop type patterns.

In organic farming, clover is an important basis for green manure in crop rotation systems due to its nitrogen-fixing effect. However, clover is often sown in mixtures with grass to achieve a yield-increasing effect. In order to determine the quantity and distribution of clover and its influence on the subsequent crops, clover plants must be identified at the individual plant level and spatially differentiated from grass plants. In practice, this is usually done by visual estimation or extensive field sampling. High-resolution unmanned aerial vehicles (UAVs) offer a more efficient alternative. In the present study, clover and grass plants were classified based on spectral information from high-resolution UAV multispectral images and texture features using a random forest classifier. Three different timestamps were observed in order to depict the phenological development of clover and grass distributions. To reduce data redundancy and processing time, relevant texture features were selected based on a wrapper analysis and combined with the original bands. Including these texture features, a significant improvement in classification accuracy of up to 8% was achieved compared to a classification based on the original bands only. Depending on the phenological stage observed, this resulted in overall accuracies between 86% and 91%. Subsequently, high-resolution UAV imagery data allow for precise management recommendations for precision agriculture with site-specific fertilization measures.

The documentation of historical remains and cultural heritage is of great importance to preserve historical knowledge. Many studies use low-resolution airplane-based laser scanning and manual interpretation for this purpose. In this study, a concept to automatically detect terrain anomalies in a historical conflict landscape using high-resolution UAV-LiDAR data was developed. We applied different ground filter algorithms and included a spline-based approximation step in order to improve the removal of low vegetation. Due to the absence of comprehensive labeled training data, a one-class support vector machine algorithm was used in an unsupervised manner in order to automatically detect the terrain anomalies. We applied our approach in a study site with different densities of low vegetation. The morphological ground filter was the most suitable when dense near-ground vegetation is present. However, with the use of the spline-based processing step, all filters used could be significantly improved in terms of the F1-score of the classification results. It increased by up to 42 percentage points in the area with dense low vegetation and by up to 14 percentage points in the area with sparse low vegetation. The completeness (recall) reached maximum values of 0.8 and 1.0, respectively, when taking into account the results leading to the highest F1-score for each filter. Therefore, our concept can support on-site field prospection.