ahmed aboutaib

Research internship focused on contrastive self-supervised learning (SSL) applied to multispectral Sentinel-2 time series for crop representation without labeled data. Developed and evaluated ten temporal data augmentation techniques (e.g., Resampling, Temporal Dropout, Window Slice, StiefelGen) integrated within SimCLR, MoCo, BYOL, and VICReg frameworks using PyTorch. Conducted large-scale experiments on the FranceCrops dataset, analyzing how temporal structures affect representation robustness and downstream crop classification performance. This work contributes to advancing unsupervised representation learning for remote sensing time series, reducing annotation needs by up to 5× while improving model generalization.

Developed an automated system for multilingual news processing, including web scraping, paraphrasing, and translation in French, Arabic, and English. The pipeline combined Selenium for data extraction, a T5 model fine-tuned with LoRA for paraphrase generation, and Google Translate for automatic translation. The entire process was automated using a task scheduler, enabling continuous and scalable content publication.