Vision Transformers for Remote-Sensing Images

Vision Transformers for Remote-Sensing Images is a cutting-edge international workshop designed to teach participants how to apply state-of-the-art Vision Transformer architectures to satellite and aerial imagery. With growing applications in climate research, defense, agriculture, and urban planning, transformers are enabling a leap forward in geospatial image analysis.

This hands-on program will introduce the theory of transformers, their adaptation to vision tasks (e.g., ViT, Swin Transformer), and how they outperform traditional CNNs in capturing long-range dependencies and spatial relationships in high-resolution imagery. Participants will work on real datasets (Sentinel, Landsat, DOTA, etc.) using frameworks like PyTorch, Hugging Face, and TIMM.

• Introduce Vision Transformers and their application in satellite image analysis

• Enable hands-on experimentation with publicly available geospatial datasets

• Teach model customization and fine-tuning techniques

• Promote responsible AI usage in environmental and humanitarian applications

• Foster interdisciplinary innovation at the intersection of AI and Earth science

Day 1: Transformers vs CNNs in Remote Sensing

Session Title: “Beyond CNNs: Vision Transformers for Scene Classification”

🔹 Topics:

• Review of CNN architectures in remote sensing (ResNet, UNet, etc.)
• Introduction to Vision Transformers (ViT): How they work
• Why ViTs are suited for remote-sensing imagery (large context, less inductive bias)
• Comparison: ViT vs CNN in scene classification

🔹 Hands-on/Demo:

• Colab demo using pretrained ViT and CNN for a sample land scene classification task using EuroSAT or BigEarthNet dataset

Day 2: Land-Cover Change Detection Using Transformers

Session Title: “Tracking the Earth: Transformers for Change Detection”

🔹 Topics:

• Problem of land-cover change detection (LCCD) and its applications (urbanization, deforestation)
• Architectures adapted for temporal change detection (Siamese ViTs, TimeSFormer)
• Pipeline: Preprocessing → Patch Embedding → Transformer Blocks → Classification head

🔹 Hands-on/Case Study:

• Visual result comparison (before/after images and heatmaps)

Day 3: Fine-Tuning Vision Transformers on Small Labeled Sets

Session Title: “Efficient Learning: Adapting ViTs with Limited Data”

🔹 Topics:

• Challenges of training ViTs with small labeled data
• Strategies: Transfer learning, self-supervised learning (DINO, MAE), adapter layers
• Case studies in remote sensing: Agriculture crop mapping, disaster response

🔹 Hands-on/Discussion:

• Colab demo: Fine-tuning a ViT model on a small custom dataset
• Open discussion: “Where can we take this next?” (research directions, publication ideas)

• Geospatial and remote-sensing professionals

• AI/ML engineers and computer vision researchers

• Earth scientists, environmental engineers, and urban planners

• Students and researchers in space science, climate, or deep learning

• Government/NGO professionals working with Earth observation data