This thesis addresses the problem of semantic change detection in satellite image time series (SITS-SCD), which represents one of the key research challenges in modern remote sensing. Existing architectures, predominantly based on convolutional encoders, often achieve limited performance, particularly when dealing with rare classes and when generalizing to out-of-distribution data. The objective of this work is to explore improvements to these architectures through the use of pretrained models, which can provide richer representations and better capture spatio-temporal patterns.
We first conduct a detailed analysis of a reference SITS-SCD architecture and evaluate its performance on the DynamicEarthNet and MUDS datasets. We then systematically investigate architectural modifications, including the replacement of convolutional blocks with pretrained components (e.g., ResNet blocks), as well as parameter tuning. Special attention is given to assessing the impact of these modifications on overall accuracy, rare class prediction, and cross-domain robustness.
The experimental results aim to contribute to a deeper understanding of the advantages and limitations of pretrained models in the context of multi-temporal semantic change detection. Finally, we provide an evaluation of the effectiveness of the proposed modifications and formulate recommendations for further development of SITS-SCD architectures.
|
