Satellite imagery: detection and segmentation

Research question

Since August 2025, I have been working under Dr. Ujjwal Verma at Manipal Institute of Technology on object detection and semantic segmentation in satellite imagery. The goal is to compare deep-learning approaches, including transformer-based methods, and understand how architecture, preprocessing, loss functions, and training choices affect predictions on remote-sensing data.

Dataset and preparation

The original dataset exceeds 10 GB, with more than 8 GB after cleaning. Images are processed at 512 × 512 pixels. My work has included annotation handling, cleaning, preprocessing, dataset preparation, and splitting. Exact dataset names, classes, annotation formats, and sample imagery are not published here because the project is subject to an NDA.

Experimental methodology

The research has involved more than eleven detection architectures and six segmentation architectures, spanning conventional baselines and transformer-based approaches. My responsibilities include training and fine-tuning, changing hyperparameters and loss configurations, and analyzing outputs with task-appropriate metrics: accuracy, precision, recall, F1, IoU, mIoU, and mAP.

I have worked with PyTorch and TensorFlow workflows, TensorBoard, Google Colab Pro, and RTX 3050 compute resources. Approximately 50–100 experiments have been undertaken across the project. Architecture names, detailed comparisons, and unpublished outcomes are omitted pending approval.

My role

As Lead Student Researcher, I coordinate seven students while remaining hands-on with implementation, evaluation, and analysis. I organize dataset and modeling tasks, review progress, maintain consistency across experiments, and contribute to documentation, figures, presentations, and discussions of research direction with the supervising researcher.

Current status and disclosure

The project is ongoing, with work progressing through data preparation, model experimentation, evaluation, and comparative analysis toward intended academic outputs. The NDA limits what can be shared publicly. The 91% accuracy and 0.73 mIoU figures previously listed on my CV are not reproduced here until their disclosure and evaluation context are approved.

What I am learning

Comparing models fairly requires more than obtaining a good score. Dataset consistency, metric definitions, reproducible training configurations, and careful interpretation of failure cases all matter. I am interested in what architectural comparisons can teach us about model behavior, rather than treating the highest single metric as the complete answer.

Possible public artifacts

An approved high-level pipeline diagram, sanitized model comparison, permitted prediction visualizations, or an authorized research presentation may be added later. Restricted data, internal tables, and unpublished findings will not be posted without approval.

← All projects