@article{CherifHellBrandmeier, author = {Cherif, Eya and Hell, Maximilian and Brandmeier, Melanie}, title = {DeepForest: novel deep learning models for land use and land cover classification using multi-temporal and -modal sentinel data of the amazon basin}, series = {Remote Sensing}, volume = {14}, journal = {Remote Sensing}, number = {19}, issn = {2072-4292}, doi = {10.3390/rs14195000}, abstract = {Land use and land cover (LULC) mapping is a powerful tool for monitoring large areas. For the Amazon rainforest, automated mapping is of critical importance, as land cover is changing rapidly due to forest degradation and deforestation. Several research groups have addressed this challenge by conducting local surveys and producing maps using freely available remote sensing data. However, automating the process of large-scale land cover mapping remains one of the biggest challenges in the remote sensing community. One issue when using supervised learning is the scarcity of labeled training data. One way to address this problem is to make use of already available maps produced with (semi-) automated classifiers. This is also known as weakly supervised learning. The present study aims to develop novel methods for automated LULC classification in the cloud-prone Amazon basin (Brazil) based on the labels from the MapBiomas project, which include twelve classes. We investigate different fusion techniques for multi-spectral Sentinel-2 data and synthetic aperture radar Sentinel-1 time-series from 2018. The newly designed deep learning architectures—DeepForest-1 and DeepForest-2—utilize spatiotemporal characteristics, as well as multi-scale representations of the data. In several data scenarios, the models are compared to state-of-the-art (SotA) models, such as U-Net and DeepLab. The proposed networks reach an overall accuracy of up to 75.0\%, similar to the SotA models. However, the novel approaches outperform the SotA models with respect to underrepresented classes. Forest, savanna and crop were mapped best, with F1 scores up to 85.0\% when combining multi-modal data, compared to 81.6\% reached by DeepLab. Furthermore, in a qualitative analysis, we highlight that the classifiers sometimes outperform the inaccurate labels.}, language = {en} } @article{HellBrandmeierBriechleetal., author = {Hell, Maximilian and Brandmeier, Melanie and Briechle, Sebastian and Krzystek, Peter}, title = {Classification of Tree Species and Standing Dead Trees with Lidar Point Clouds Using Two Deep Neural Networks: PointCNN and 3DmFV-Net}, series = {PFG - Journal of Photogrammetry, Remote Sensing and Geoinformation Science}, volume = {90}, journal = {PFG - Journal of Photogrammetry, Remote Sensing and Geoinformation Science}, issn = {2512-2819}, doi = {10.1007/s41064-022-00200-4}, pages = {103 -- 121}, abstract = {Knowledge about tree species distribution is important for forest management and for modeling and protecting biodiversity in forests. Methods based on images are inherently limited to the forest canopy. Airborne lidar data provide information about the trees' geometric structure, as well as trees beneath the upper canopy layer. In this paper, the potential of two deep learning architectures (PointCNN, 3DmFV-Net) for classification of four different tree classes is evaluated using a lidar dataset acquired at the Bavarian Forest National Park (BFNP) in a leaf-on situation with a maximum point density of about 80 pts/m2. Especially in the case of BFNP, dead wood plays a key role in forest biodiversity. Thus, the presented approaches are applied to the combined classification of living and dead trees. A total of 2721 single trees were delineated in advance using a normalized cut segmentation. The trees were manually labeled into four tree classes (coniferous, deciduous, standing dead tree with crown, and snag). Moreover, a multispectral orthophoto provided additional features, namely the Normalized Difference Vegetation Index. PointCNN with 3D points, laser intensity, and multispectral features resulted in a test accuracy of up to 87.0\%. This highlights the potential of deep learning on point clouds in forestry. In contrast, 3DmFV-Net achieved a test accuracy of 73.2\% for the same dataset using only the 3D coordinates of the laser points. The results show that the data fusion of lidar and multispectral data is invaluable for differentiation of the tree classes. Classification accuracy increases by up to 16.3\% points when adding features generated from the multispectral orthophoto.}, language = {en} } @article{HellBrandmeier, author = {Hell, Maximilian and Brandmeier, Melanie}, title = {Identifying Plausible Labels from Noisy Training Data for a Land Use and Land Cover Classification Application in Amaz{\^o}nia Legal}, series = {remote sensing}, volume = {26}, journal = {remote sensing}, number = {12}, doi = {10.3390/rs16122080}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:863-opus-57532}, pages = {24}, abstract = {Most studies in the field of land use and land cover (LULC) classification in remote sensing rely on supervised classification, which requires a substantial amount of accurate label data. However, reliable data are often not immediately available, and are obtained through time-consuming manual labor. One potential solution to this problem is the use of already available classification maps, which may not be the true ground truth and may contain noise from multiple possible sources. This is also true for the classification maps of the MapBiomas project, which provides land use and land cover (LULC) maps on a yearly basis, classifying the Amazon basin into more than 24 classes based on the Landsat data. In this study, we utilize the Sentinel-2 data with a higher spatial resolution in conjunction with the MapBiomas maps to evaluate a proposed noise removal method and to improve classification results. We introduce a novel noise detection method that relies on identifying anchor points in feature space through clustering with self-organizing maps (SOM). The pixel label is relabeled using nearest neighbor rules, or can be removed if it is unknown. A challenge in this approach is the quantification of noise in such a real-world dataset. To overcome this problem, highly reliable validation sets were manually created for quantitative performance assessment. The results demonstrate a significant increase in overall accuracy compared to MapBiomas labels, from 79.85\% to 89.65\%. Additionally, we trained the L2HNet using both MapBiomas labels and the filtered labels from our approach. The overall accuracy for this model reached 93.75\% with the filtered labels, compared to the baseline of 74.31\%. This highlights the significance of noise detection and filtering in remote sensing, and emphasizes the need for further research in this area.}, language = {en} } @inproceedings{HellBrandmeierNuechter, author = {Hell, Maximilian and Brandmeier, Melanie and N{\"u}chter, Andreas}, title = {Transferability of Deep Learning Models for Land Use/Land Cover Classification}, series = {43. Wissenschaftlich-Technische Jahrestagung der DGPF in M{\"u}nchen. 22.-23. M{\"a}rz 2023 in M{\"u}nchen - Publikationen der DGPF}, volume = {31}, booktitle = {43. Wissenschaftlich-Technische Jahrestagung der DGPF in M{\"u}nchen. 22.-23. M{\"a}rz 2023 in M{\"u}nchen - Publikationen der DGPF}, editor = {Kersten, Thomas P. and Tilly, Nora}, doi = {10.24407/KXP:1841078182}, pages = {142 -- 149}, language = {en} }