{
"markdown": "# rs-dl-architect\n\n**Expert guidance for designing, analyzing, and reviewing deep learning architectures in remote sensing and satellite imagery.**\n\n\n\n[](LICENSE)\n[](https://github.com/anthropics/skills)\n[](https://your-username.github.io/rs-dl-architect)\n\nA comprehensive Claude AI skill covering segmentation, change detection, multi-modal fusion, temporal analysis, foundation models (SatMAE, Prithvi, GeoSAM), object detection, and uncertainty quantification. Supports both **PyTorch** and **TensorFlow** with production-ready code templates. ([ Link to website ](https://rupesh4604.github.io/rs-dl-architect/))\n\n---\n\n## šÆ What This Skill Does\n\n- **Designs** architectures tailored to your remote sensing task (task-shape diagnosis first, not reflexive model recommendations)\n- **Analyzes** existing models, papers, and checkpoints (explains what's actually happening, not just listing layers)\n- **Reviews** your code/architecture with a 10-section rigorous checklist (blocking issues + concrete fixes, not vague suggestions)\n\n**Key differentiator:** Not generic computer vision adapted to RS ā built from the ground up for satellite/aerial imagery physics, GSD considerations, geographic generalization, and multi-spectral/multi-temporal/multi-modal challenges.\n\n---\n\n## š Quick Start\n\n### Installation\n\n**Option 1: Via Claude Code Plugin** _(Recommended)_\n\n```bash\nclaude plugin install rs-dl-architect@Rupesh4604\n```\n\n**Option 2: Manual Installation**\n\n```bash\n# Clone the repository\ngit clone https://github.com/Rupesh4604/rs-dl-architect.git\n\n# Copy to Claude skills directory\ncp -r rs-dl-architect/rs-dl-architect ~/.claude/skills/\n\n# For project-specific installation\ncp -r rs-dl-architect/rs-dl-architect ./.claude/skills/\n```\n\n**Option 3: Direct Download**\nDownload the [latest .skill file](https://github.com/Rupesh4604/rs-dl-architect/releases) and upload via Claude.ai or Claude Code.\n\n### Usage\n\nJust ask Claude naturally ā the skill triggers automatically on relevant queries:\n\n```\n\"I need to detect crop field boundaries from 10m Sentinel-2 with 500 labels\"\n\n\"How should I fuse SAR and optical for flood mapping?\"\n\n\"Review my U-Net ā using ImageNet weights with 13 Sentinel-2 bands\"\n\n\"Should I use SatMAE or DINOv2 for EuroSAT land cover classification?\"\n```\n\n---\n\n## š Coverage\n\n### Architectures & Tasks\n\n| Domain | Coverage |\n| ------------------------- | ------------------------------------------------------------------------------------------------ |\n| **Segmentation** | U-Net family, SegFormer, UNetFormer, Mask2Former, DC-Swin, RS-Mamba, boundary-aware patterns |\n| **Change Detection** | Siamese (FC-Siam-diff), BIT, ChangeFormer, ChangeMamba, semantic CD, continuous monitoring |\n| **Data Fusion** | SAR+optical, RGB+DEM, hyperspectral+LiDAR, early/mid/late fusion, modality dropout, CMX, DOFA |\n| **Temporal Analysis** | ConvLSTM, U-TAE, TSViT, Presto, TempCNN, irregular time series, cloud masking |\n| **Foundation Models** | SatMAE, Prithvi-EO, Clay, DOFA, GeoSAM, RemoteCLIP, LoRA/PEFT adaptation, multispectral adapters |\n| **Object Detection** | RetinaNet, FCOS, DETR/DINO, oriented bbox (DOTA), small-object patterns, YOLOv8-OBB |\n| **Interpretability & UQ** | Grad-CAM, MC dropout, deep ensembles, evidential learning, conformal prediction, calibration |\n\n### Code Templates (PyTorch + TensorFlow)\n\n- Multispectral input adapters for ImageNet weights\n- Minimal U-Net with shape annotations\n- Siamese change-detection (FC-Siam-diff)\n- Combined Dice + CE loss\n- Training loop with mixed precision + gradient accumulation\n- Tile-based inference with cosine-weighted blending\n- Deep ensemble patterns\n\n---\n\n## š” Example Use Cases\n\n\nArchitecture Selection\n\n**Query:**\n\n```\nI need to detect crop field boundaries from 10m Sentinel-2 data\nwith only 500 labeled samples. What architecture should I use?\n```\n\n**Response:**\n\n- Recommends foundation-model fine-tuning (Prithvi-EO with LoRA)\n- Explains why label efficiency matters\n- Provides PyTorch implementation with proper multispectral input handling\n- Lists failure modes (GSD mismatch, small fields <2 pixels)\n- Suggests alternatives (SatMAE++, from-scratch U-Net as baseline)\n\n\n\nMulti-Modal Fusion Design\n\n**Query:**\n\n```\nHow do I combine SAR and optical data for flood mapping?\nShould I fuse early or late?\n```\n\n**Response:**\n\n- Analyzes modality differences (speckle vs additive noise, cloud penetration)\n- Recommends mid-level cross-attention fusion (not early concat)\n- Flags modality dropout requirement for robustness\n- Provides dual-encoder architecture with shape annotations\n- Cites CMX, DeCUR as reference implementations\n\n\n\nCode Review\n\n**Query:**\n\n```\nReview my U-Net for building extraction ā using ImageNet weights\nwith 13 Sentinel-2 bands via channel concat.\n```\n\n**Response:**\n\n- **Blocking issue:** Silently dropping spectral information (red-edge, SWIR)\n- **Fix:** Use multispectral adapter pattern from code templates\n- Checks receptive field vs GSD (10m ā need larger context)\n- Suggests boundary-aware loss for cadastral precision\n- Verifies train/val split is region-based, not random tiles\n\n\n\nFoundation Model Selection\n\n**Query:**\n\n```\nShould I use SatMAE or DINOv2 for land cover classification on EuroSAT?\n```\n\n**Response:**\n\n- Compares both: SatMAE native multispectral vs DINOv2 stronger features\n- Recommends SatMAE++ with linear probe baseline first\n- Provides LoRA fine-tuning template for parameter efficiency\n- Explains when each wins (SatMAE: <10k labels; DINOv2: RGB-only high-res)\n- Notes EuroSAT is small (27k tiles) ā foundation models mandatory\n\n\n---\n\n## š§ How It Works\n\n### Three-Mode Workflow\n\nThe skill automatically identifies your intent:\n\n1. **Build Mode** ā You want a new architecture or implementation\n - Task-shape diagnosis (what's being predicted, input modality, GSD, label budget)\n - Routes to relevant subdomain reference (segmentation, fusion, temporal, etc.)\n - Produces 5-part recommendation: What / Why / Cost / Failure Modes / Alternatives\n\n2. **Analyze Mode** ā You want to understand an existing architecture/paper\n - Structured walkthrough: data flow ā tensor shapes ā key decisions ā novelty\n - Comparisons to baselines and related work\n - Diagrams when branches/attention is non-trivial\n\n3. **Review Mode** ā You want critique or debugging help\n - 10-section checklist (task-architecture match, data flow, loss, evaluation, etc.)\n - Blocking issues with concrete fixes (not \"consider regularization\")\n - Acknowledges strengths + flags weaknesses with evidence\n\n### Cross-Cutting Principles\n\nBuilt into every recommendation:\n\n- **Channels ā RGB** ā multispectral input needs proper adapters, never silent band dropping\n- **GSD matters** ā receptive field, augmentation, class definitions all depend on ground sampling distance\n- **Geographic generalization** ā held-out region tests, not random tile splits\n- **Foundation models changed the calculus** ā for <50k labels, fine-tune beats from-scratch\n- **Temporal context is often free signal** ā multi-date stacks almost always beat single-date\n- **Class imbalance is the norm** ā weighted loss / focal loss / sampling strategies required\n\n---\n\n## š Repository Structure\n\n```\nrs-dl-architect/\nāāā README.md # This file\nāāā docs/ # GitHub Pages documentation site\nā āāā index.html\nā āāā privacy.html\nāāā rs-dl-architect/ # The actual skill\nā āāā SKILL.md # Main entry point (~120 lines)\nā āāā references/ # Loaded on-demand\nā āāā segmentation.md\nā āāā change-detection.md\nā āāā data-fusion.md\nā āāā temporal-analysis.md\nā āāā foundation-models.md\nā āāā object-detection.md\nā āāā interpretability-uncertainty.md\nā āāā review-checklist.md\nā āāā code-templates.md\nā āāā design-principles.md\nāāā LICENSE\nāāā .gitignore\n```\n\n**Size:** ~1,460 lines of expert guidance across 10 files \n**Main file:** 120 lines (routes to references) \n**References:** 100ā300 lines each (loaded only when relevant)\n\n---\n\n## š Target Users\n\n- Remote sensing researchers\n- GIS engineers and geospatial AI practitioners\n- Graduate students in earth observation / geospatial ML\n- Satellite imagery ML engineers\n- Anyone working with Sentinel-1/2, Landsat, PlanetScope, aerial imagery, hyperspectral, SAR, or multi-modal EO data\n\n**Assumes:** MS/PhD-level technical depth, familiarity with PyTorch or TensorFlow, working knowledge of remote sensing fundamentals.\n\n---\n\n## š¤ Contributing\n\nContributions welcome! Areas of particular interest:\n\n- Additional code templates (e.g., TensorFlow equivalents, JAX patterns)\n- Newer foundation models (as they're released)\n- Domain-specific extensions (agriculture, disaster response, urban planning)\n- More review checklist items\n- Dataset recommendations and benchmark updates\n\n**To contribute:**\n\n1. Fork the repository\n2. Create a feature branch (`git checkout -b feature/new-reference`)\n3. Make your changes (follow the existing reference structure)\n4. Test with realistic queries\n5. Submit a Pull Request\n\nSee [CONTRIBUTING.md](CONTRIBUTING.md) for detailed guidelines.\n\n---\n\n## š License\n\nThis skill is released under the [MIT License](LICENSE).\n\nYou are free to:\n\n- Use commercially\n- Modify and redistribute\n- Use privately\n- No warranty or liability\n\n---\n\n## š Links\n\n- **Documentation:** [https://Rupesh4604.github.io/rs-dl-architect](https://Rupesh4604.github.io/rs-dl-architect)\n- **Issues & Discussions:** [GitHub Issues](https://github.com/your-username/rs-dl-architect/issues)\n- **Privacy Policy:** [rs-dl-architect/privacy](https://rupesh4604.github.io/rs-dl-architect/privacy.html)\n- **Claude Skills Specification:** [Anthropic Skills Repo](https://github.com/anthropics/skills)\n- **Related Work:** [Awesome Remote Sensing Change Detection](https://github.com/wenhwu/awesome-remote-sensing-change-detection)\n\n---\n\n## š Stats\n\n\n\n\n\n**Coverage:** Segmentation ā¢ Change Detection ā¢ Data Fusion ā¢ Temporal ā¢ Foundation Models ā¢ Detection ā¢ Interpretability ā¢ Review\n\n---\n\n## š Acknowledgments\n\nBuilt for the remote sensing and geospatial AI community. Special thanks to:\n\n- The teams behind SatMAE, Prithvi-EO, Clay, GeoSAM, and other RS foundation models\n- Authors of BIT, ChangeFormer, U-TAE, and other open-source RS architectures\n- The broader earth observation ML research community\n\n---\n\n## š® Contact\n\n**Maintainer:** [M Rupesh Kumar Yadav](<[@Rupesh4604](https://github.com/Rupesh4604)>) \n**Email:** rupesh32003@gmai.com \n**Institution:** IIT Bombay, M.Tech\n\nFor questions, feedback, or collaboration opportunities, open an issue or reach out directly.\n\n---\n\n
\n Built with ā¤ļø for the remote sensing community\n \n Making deep learning architecture decisions clearer, one satellite image at a time\n