Deep Learning Architecture Expert for Remote Sensing

Design, analyze, and review deep learning models for satellite imagery, aerial photos, and geospatial AI — with production-ready code and honest engineering tradeoffs.

[ PYTORCH • TENSORFLOW • FOUNDATION MODELS • MULTISPECTRAL ]

Install Plugin See Examples

Built for Real Remote Sensing Challenges

🛰️

Multi-Modal Fusion

SAR + optical, RGB + DEM, hyperspectral + LiDAR. Handles early/mid/late fusion patterns, modality dropout, and resolution alignment — not just naive channel stacking.

📡

Foundation Models

SatMAE, Prithvi-EO, Clay, DOFA, GeoSAM, RemoteCLIP. Covers LoRA/PEFT adaptation, multispectral input adapters, and when to use each vs training from scratch.

🔍

Change Detection

Bi-temporal siamese, BIT, ChangeFormer, ChangeMamba. Distinguishes binary vs semantic CD, handles registration errors, and flags pseudo-change patterns.

📊

Temporal Analysis

ConvLSTM, U-TAE, TSViT, Presto. Processes irregular time series, handles cloud masking, and separates per-pixel vs per-tile temporal modeling strategies.

🎯

Code Templates

PyTorch + TensorFlow starter patterns. Multispectral adapters, U-Net, siamese networks, Dice+CE loss, tile inference with blending, deep ensembles — all with shape annotations.

🔬

Rigorous Reviews

10-section checklist covering task-architecture match, data flow, loss design, geographic generalization, and reproducibility. Identifies blocking issues with concrete fixes.

Example Use Cases

01. Architecture Selection
"I need to detect crop field boundaries from 10m Sentinel-2 data with only 500 labeled samples. What architecture should I use?"
→ Recommends foundation-model fine-tuning (Prithvi-EO with LoRA), explains why label efficiency matters, provides PyTorch implementation with proper multispectral input handling.
02. Multi-Modal Design
"How do I combine SAR and optical data for flood mapping? Should I fuse early or late?"
→ Analyzes modality differences (speckle vs additive noise), recommends mid-level cross-attention fusion, flags modality dropout requirement, provides dual-encoder architecture.
03. Code Review
"Review my U-Net for building extraction — using ImageNet weights with 13 Sentinel-2 bands via channel concat."
→ Identifies blocking issue (silently dropping spectral information), provides specific fix (multispectral adapter pattern), checks receptive field vs GSD.
04. Uncertainty Quantification
"My crop classifier fails catastrophically on some tiles. How do I know when not to trust it?"
→ Recommends deep ensembles (5 seeds) + temperature scaling, explains epistemic vs aleatoric uncertainty, provides selective prediction pattern.

Installation

$ claude plugin install rs-dl-architect
# or via GitHub
$ git clone https://github.com/Rupesh4604/rs-dl-architect
$ cp -r rs-dl-architect ~/.claude/skills/

Works in Claude Code, Claude.ai, and any environment with the SKILL.md specification

Coverage Across Remote Sensing Deep Learning

Segmentation

  • U-Net family
  • SegFormer, UNetFormer
  • Mask2Former
  • DC-Swin, RS-Mamba

Object Detection

  • RetinaNet, FCOS
  • DETR, DINO
  • Oriented bbox (DOTA)
  • Small-object patterns

Foundation Models

  • SatMAE, Prithvi-EO
  • Clay, DOFA
  • GeoSAM, RemoteCLIP
  • LoRA/PEFT adaptation

Interpretability

  • Grad-CAM, attention
  • MC dropout, ensembles
  • Evidential learning
  • Conformal prediction
1,460+
Lines of Expert Guidance
10
Specialized References
2
Frameworks (PyTorch + TF)
100%
Remote Sensing Focused