Focus Areas
- Machine learning in the field of agriculture and vegetation
- Classification methods, such as supervised learning, unsupervised learning, self-taught learning
- Feature learning, representation learning
- Interpretability and explainability for deep neural networks
- Analysis techniques, such as clustering or archetypal analysis
- Pattern recognition in large data sets
Topic Propositions
Viewpoint Planning for Fruit Mapping in Occluded Regions by Generative Models
Advanced precision agriculture applications such as fruit counting for yield estimation are expected to play a key role to meet rising demands on agricultural production. However, autonomous targeted sensing of fruit is challenging due to the cluttered, unstructured nature of greenhouse environments where fruits are typically grown. Inherent occlusions and inaccessible areas make it difficult to precisely maneuver a robotic sensor to inspect the fruit from different viewpoints. Actively planning viewpoints in a way that yet unobserved but interesting regions of fruit are sensed efficiently by guiding the sensor targeted to informative waypoints remains subject of ongoing research.
This master thesis will investigate using deep learning-based methods for viewpoint planning. Specifically, it will explore generative neural networks, which aim at learning the underlying data structure in order to create then new or reconstruct hidden data out of it (see here). This allows for reasoning about the scene in currently occluded regions as a basis for active viewpoint planning(see here). Fruit mapping and planning performance will be assessed in terms of appropriate evaluation criteria compared to less advanced planners. The student can use an existing simulation environment and planning approach as a starting point.
Supervision: Ahmed Emam, Ribana Roscher
The Effect of Cutmix and Mixup On Attribution Methods (Master Thesis)
German title: Der Effekt von cutmix und Mixup auf Attribution Methoden
There exist several attribution methods to interpret the decision of Convolutional Neural Networks (CNN) such as Saliency Maps, GradCAM, Occlusion Sensitivity Maps, etc. Our previous experiments suggest, that there are relations between attribution maps derived from classification tasks and the pixel-wise occurrence in the corresponding remote sensing image, especially for binary classification problems. The task of this thesis would be to analyze what effect CutMix and MixUp have in this regard.
Supervision: Timo Stomberg, Ribana Roscher
Combine Attributions of all Layers in Convolutional Neural Networks (Bachelor Or Master Thesis)
German titlE: Kombination der Attributionen aller Layer in Convolutional Neural Networks
There exist several attribution methods to interpret the decision of Convolutional Neural Networks (CNN) such as Saliency Maps, GradCAM, Occlusion Sensitivity Maps, etc. Depending on the layer the method is applied to, the result can be very different. This is due to the fact that the deeper the layer, the more complex the features the model is considering: The first layers consider low-level features such as edges, and the last layers consider high-level features. The task of this thesis would be to combine the attributions of multiple layers to make more general statements about the decision process of a CNN.
Supervision: Timo Stomberg, Ribana Roscher
The Effect of Activation Functions on Saliency Maps (Bachelor or Master Thesis)
German titlE: DeR Effekt Von Aktivierungsfunktionen Auf Saliency Maps
There exist several attribution methods to interpret the decision of Convolutional Neural Networks (CNN). Gradient-based methods such as Saliency Maps often produce very noisy results. There exist methods to compensate for that noisiness such as SmoothGrad. We believe, that the commonly used activation function ReLU contributes to the noise and that other activation functions, such as tanh, might result in less noise.
Supervision: Timo Stomberg, Ribana Roscher
Renaturation detection using esa Worldcover Maps of 2020 and 2021, and Sentinel-2 imagery (Bachelor Thesis)
German title: Detektion von Renaturierung unter Verwendung der ESA WorldCover karten von 2020 und 2021, und Sentinel-2 Bildern
ESA WorldCover is a land cover dataset with 10 meters resolution which has been published for the years 2020 and 2021. Land cover changes within these two maps might reveal renaturation areas. However, since different algorithms have been used to generate the maps, differences might not only occur due to real land cover changes. Therefore, Sentinel-2 images must be used to validate the land cover changes in the ESA WorldCover maps.
Supervision: Timo Stomberg, Ribana Roscher
weakly supervised Land cover prediction using the activation space of a neural network (master Thesis)
German title: Schwach überwachtes lernen von landbedeckungs-klassen unter verwendung des aktivierungsraumes eines neuronalen netzwerks
Analyzing the activation space of a neural network allows a better understanding of the decision-making process of the network. We found a separation of land cover classes in the activation space, although our model has not been trained on these classes. The task of the master thesis is to analyze this separation for several network architectures. Further, the relationship between activations and land cover classes allows the prediction of land cover classes in satellite images. Experiments shall show to which extent this is possible. For more information on the activation space, see here.
Supervision: Timo Stomberg, Ribana Roscher
Erfassung landwirtschaftlicher Praktiken in Deutschland – Analyse von Bodenbearbeitung und Glyphosateinsatz im deutschen Ackerbau (Master Thesis)
Die Art und Weise, wie wir unsere Nahrungsmittel herstellen, ist unumstritten ein wichtiger
Einflussfaktor für das Erreichen globaler und lokaler Umweltziele. Um jedoch entsprechende agrarpolitische Maßnahmen effizient und zielführend auszugestalten, ist detailliertes Wissen über die Verbreitung und Nutzung verschiedener landwirtschaftlicher Praktiken fundamental – das Sammeln und Aufbereiten eben solcher Informationen stellt Forschende und Entscheidungsträger*innen jedoch bisher oft vor viele Probleme, da Daten nur eingeschränkt verfügbar und/oder schwer zu erheben sind.
Das TrAgS-Projekt (Tracking the use and adoption of Agricultural technologies through Satellite remote sensing and self-supervised deep learning) adressiert diese Schwierigkeit der Datenverfügbarkeit und hat das Ziel, die Analyse von Landwirtschaftstechnologien mithilfe von Satellitenfernerkundungsdaten und maschinellem Lernen zu verbessern.
Die Masterarbeit sollte umfassen:
- Das Ausformulieren der Umfrage und Verteilen dieser unter deutschen Ackerbau-Landwirten
- Unterstützung in der Koordination für die technische Umsetzung der Befragung
- Das Ordnen und Filtern der erhobenen Daten
- Deskriptive Auswertung der Daten hinsichtlich des Nutzens besagter Praktiken
- Eine erste Analyse und Interpretation der gefundenen Nutzungsmuster
!!! Diese Arbeit wird nicht von unserer Gruppe betreut. Die Ergebnisse der Arbeit sind jedoch auch für unsere Arbeit im TrAgS Projekt relevant.
Supervision: Alexa Leyens, Hugo Storm
Analyzing Internal Waves with multi-modal deep learning
German title: Analyse interner Wellen mit multi-modalem deep learning
The world’s oceans are constantly in motion due to various influences caused by natural and anthropogenic factors. In order to better quantify and understand these changes, monitoring tools need to be developed which analyze different types of observations with global coverage. Against this background, satellite observations provide comprehensive multi-modal data, representing a valuable complement to sparsely distributed in-situ observing systems. In combination with automatic analysis and interpretation tools, they are useful to continuously monitor the ocean.
The aim of this thesis is to derive statistics of internal waves over time using deep-learning algorithms. Multi-modal Sentinel-3 observations will be used and analyzed jointly in a multi-modal framework.
Supervision: Ribana Roscher, Jürgen Kusche
Sparse Representation for root segmentation in Hyperspectral images
German title: Sparse Representation für die Segmentierung von Wurzeln in Hyperspektralbildern
The segmentation of roots in images is challenging due to their similar color appearance and their thin structure. Sparse representation can be used for segmentation by representing each image patch by a weighted linear combination of basis patches (e.g., [2]). These basis patches are learned from annotated images and will be collected in discriminative dictionaries, which contain the most distinct structures of roots and background. The suitability of the approach will be evaluated by means of hyperspectral images of plants grown in flat rhizotrons [1].
[1] Steier, A., Cendrero-Mateo, P., Malolepszy, T., Gioia, T., Briese, C., & Rascher, U. (2013). HIDeR – Hyperspectral Imaging Device for Rhizotrons Simultaneously Scanning of Shoot and Root in Plant Phenotyping, DPPN.
[2] Berkels, B., Kotowski, M., Rumpf, M., & Schaller, C. (2011, May). Sulci detection in photos of the human cortex based on learned discriminative dictionaries. In International Conference on Scale Space and Variational Methods in Computer Vision, 326-337.
Supervision: Ribana Roscher, Uwe Rascher
Detection Of Iceberg Rafted Debris in X-Ray Cores
German title: Detektion von Eisberg-Geröllen in Röntgenaufnahmen von Bohrkernen
Analyzing the current and past behavior of the Antarctic ice sheet is a major research area to get information about the long-term climate history of Antarctica and to predict the future behavior of the ice sheet. As soon as wandering icebergs reach warmer waters, the debris they carry (so-called iceberg rafted debris) falls down through the ocean as consequence of the melting process. The debris is deposited as sediment on the seafloor and gives, when analyzed, valuable information about changes in the ice sheet.
This thesis is concerned with an automatic algorithm to detect debris in x-ray images of drill cores, which were collected during an Antarctica expedition in 2019. Supervised learning with convolutional neural networks will be utilized to detect debris and estimate position and size in the drill core. The algorithm will be analyzed by manual annotations provided by experts.
Supervision: Ribana Roscher, Mike Weber (Institute for Geosciences Department of Geochemistry and Petrology)