Theses and Projects: Difference between revisions

(One intermediate revision by the same user not shown)

Line 58:

~~=== * '''New!''' Context Specific Self-supervised Pre-Training for Remote Sensing Applications (Semantic Segmentation, Change Detection, Socio-Economic Indicator Estimation, ...) (B/M/P) ===~~

Satellite images in combination with Machine/Deep Learning models have shown to be an effective tool for analysis and monitoring tasks regarding disasters, deforestation, climate change, socio-economic estimation and others. The training of these models usually rely on labelled ground-truth data, which is labour intensive and therefore often scarcely available. To overcome this limitation, models are often trained in self-supervised approaches with unlabelled data, such as Contrastive Learning or Masked Autoencoders. However, these approaches are completely independent and not related to the intended downstream task. In this project/thesis the relationship between the pre-training task and the model performance on the downstream task will be explored and self-supervised training approaches tailored for a selected remote sensing downstream task (semantic segmentation of trees, tree crown share per pixel estimation, change detection of disasters, socio-economic estimation, ...) will be developed.

~~Please contact Fabian Wölk [fabian.woelk@cs.uni-goettingen.de]~~

~~=== [Occupied] Tree Growth Detection using Satellite Images and Computer Vision Methods (B/M/P) ===~~

A tree planting project in Madagascar was initiated several years ago. The outcomes of this project shall now be evaluated by analyzing satellite images of the study area with Computer Vision methods. In a first step, very high resolution (VHR) satellite images from 2023 with a resolution of 0.5m will be used to identify trees with object detection / semantic segmentation. In the next step a lower resolution (5m) satellite image time series starting in 2015 will be used for change detection to identify, in which locations the project was (un)successful.

~~Please contact Fabian Wölk [fabian.woelk@cs.uni-goettingen.de]~~

=== * '''New!''' Emotional Support Conversation Generation based on LLM (B/M/P)===

Line 83:

Line 71:

Please contact Fei Gao [fei.gao@cs.uni-goettingen.de]

~~=== [Occupied] Image-to-Image Translation of Different Nightlight Image Types (B/M/P) ===~~

Nightlight intensities have been proven to be a good indicator for socio-economic status. However, for long-term temporal analyses their use can be challenging, as different satellites for sensing nightlight intensities operated at different times (DMSP OLS 1992-2014 and VIIRS 2012-2023). Both types differ not only in resolution, but there is also a big discrepancy in the optical appearance and value ranges. To obtain consistent nightlight images for temporal analysis, Image-to-Image Translation methods shall be used in this project/thesis for the conversion between both types. Finally the performance of the translated and original nightlight images for a regression on socio-economic indicators shall be evaluated.

~~Please contact Fabian Wölk [fabian.woelk@cs.uni-goettingen.de]~~

~~=== [Occupied] Satellite Image Indices and Machine Learning for Socio-economic Estimation (B/M/P) ===~~

There are several indices, which can be derived from satellite images. For example the Normalized Difference Vegetation Index (NDVI) indicates the presence and condition of vegetation, while the Normalized Difference Built-up Index (NDBI) indicates the presence of built-up areas such as buildings or roads. The distributions of these and other indices may have different explanatory power to estimate the socio-economic status of locations. Therefore in this project/thesis the regression performance of machine learning models - using statistics of these indices as features - to estimate socio-economic indicators shall be evaluated for the individual and also combined indices. Optionally, Convolutional Neural Networks (CNNs) can be applied additionally, which take the derived index images as input.

~~Please contact Fabian Wölk [fabian.woelk@cs.uni-goettingen.de]~~

=== * 3D natural hazard simulator ===

@@ Line 58: / Line 58: @@
-===  * '''New!''' Context Specific Self-supervised Pre-Training for Remote Sensing Applications (Semantic Segmentation, Change Detection, Socio-Economic Indicator Estimation, ...) (B/M/P) ===
-Satellite images in combination with Machine/Deep Learning models have shown to be an effective tool for analysis and monitoring tasks regarding disasters, deforestation, climate change, socio-economic estimation and others. The training of these models usually rely on labelled ground-truth data, which is labour intensive and therefore often scarcely available. To overcome this limitation, models are often trained in self-supervised approaches with unlabelled data, such as Contrastive Learning or Masked Autoencoders. However, these approaches are completely independent and not related to the intended downstream task. In this project/thesis the relationship between the pre-training task and the model performance on the downstream task will be explored and self-supervised training approaches tailored for a selected remote sensing downstream task (semantic segmentation of trees, tree crown share per pixel estimation, change detection of disasters, socio-economic estimation, ...) will be developed.
-Please contact Fabian Wölk [fabian.woelk@cs.uni-goettingen.de]
-===  [Occupied] Tree Growth Detection using Satellite Images and Computer Vision Methods (B/M/P) ===
-A tree planting project in Madagascar was initiated several years ago. The outcomes of this project shall now be evaluated by analyzing satellite images of the study area with Computer Vision methods. In a first step, very high resolution (VHR) satellite images from 2023 with a resolution of 0.5m will be used to identify trees with object detection / semantic segmentation. In the next step a lower resolution (5m) satellite image time series starting in 2015 will be used for change detection to identify, in which locations the project was  (un)successful.
-Please contact Fabian Wölk [fabian.woelk@cs.uni-goettingen.de]
 ===  * '''New!''' Emotional Support Conversation Generation based on LLM (B/M/P)===
@@ Line 83: / Line 71: @@
 Please contact Fei Gao [fei.gao@cs.uni-goettingen.de]
-===  [Occupied]  Image-to-Image Translation of Different Nightlight Image Types (B/M/P) ===
-Nightlight intensities have been proven to be a good indicator for socio-economic status. However, for long-term temporal analyses their use can be challenging, as different satellites for sensing nightlight intensities operated at different times (DMSP OLS 1992-2014 and VIIRS 2012-2023). Both types differ not only in resolution, but there is also a big discrepancy in the optical appearance and value ranges. To obtain consistent nightlight images for temporal analysis, Image-to-Image Translation methods shall be used in this project/thesis for the conversion between both types. Finally the performance of the translated and original nightlight images for a regression on socio-economic indicators shall be evaluated.
-Please contact Fabian Wölk [fabian.woelk@cs.uni-goettingen.de]
-===  [Occupied] Satellite Image Indices and Machine Learning for Socio-economic Estimation (B/M/P) ===
-There are several indices, which can be derived from satellite images. For example the Normalized Difference Vegetation Index (NDVI) indicates the presence and condition of vegetation, while the Normalized Difference Built-up Index (NDBI) indicates the presence of built-up areas such as buildings or roads. The distributions of these and other indices may have different explanatory power to estimate the socio-economic status of locations. Therefore in this project/thesis the regression performance of machine learning models - using statistics of these indices as features - to estimate socio-economic indicators shall be evaluated for the individual and also combined indices. Optionally, Convolutional Neural Networks (CNNs) can be applied additionally, which take the derived index images as input.
-Please contact Fabian Wölk [fabian.woelk@cs.uni-goettingen.de]
 ===  * 3D natural hazard simulator  ===