Skip to content

Latest commit

 

History

History
203 lines (154 loc) · 11.2 KB

File metadata and controls

203 lines (154 loc) · 11.2 KB

AlphaEarth Foundations vs Prithvi-100M: Technical Comparison & Ecosystem Analysis

1. Architecture Comparison

1.1 Prithvi-100M (IBM/NASA)

Paper: arXiv:2310.18660

Architecture: Standard Temporal ViT + Masked AutoEncoder (MAE)

Parameter Value
Model type Encoder-only ViT with MAE pre-training
Parameters ~100M (encoder 12 layers + decoder 8 layers)
embed_dim 768
depth (encoder) 12
num_heads 12
decoder_embed_dim 512
decoder_depth 8
decoder_num_heads 16
mlp_ratio 4
patch_size (1, 16, 16) — Conv3d with temporal tubelet
img_size 224 x 224
in_chans 6 (Blue, Green, Red, Narrow NIR, SWIR1, SWIR2)
num_frames 3 (temporal)
mask_ratio 0.75

Pre-training methodology:

  • Self-supervised MAE: mask 75% of 3D patches, reconstruct pixels
  • Single MSE reconstruction loss
  • Training data: Contiguous US HLS (Harmonized Landsat Sentinel-2) dataset only
  • Single-source optical data (no SAR, no elevation, no climate)

Key characteristics:

  • All patches processed at single resolution (14x14 = 196 tokens per frame)
  • 3D positional embedding for spatial + temporal dimensions
  • Outputs 768-dim token features — requires task head for downstream use
  • No geometric constraints on embedding space
  • Weights fully open-source (Apache 2.0) via HuggingFace

1.2 AlphaEarth Foundations (Google DeepMind)

Paper: arXiv:2507.22291

Architecture: Custom STP (Space-Time-Precision) multi-resolution encoder

Parameter Value
Model type Custom multi-resolution pyramid (NOT standard ViT)
Parameters ~480M (selected) / ~1B (largest variant)
Embedding dimension 64 (8 float32 values)
Spatial resolution 10m globally
Embedding constraint L2-normalized on S^63 unit hypersphere
Quantization 8-bit (4x compression, negligible quality loss)

STP Architecture — three simultaneous operators per block:

Operator Resolution Mechanism Purpose
Space 1/16 L ViT-like spatial self-attention Global context
Time 1/8 L Time-axial self-attention + sinusoidal timecodes Temporal dynamics
Precision 1/2 L 3x3 convolutions Local detail

Each block terminates with learned Laplacian pyramid rescaling, enabling cross-resolution state passing. Final output at precision operator resolution (1/2 L).

Pre-training methodology:

  • Teacher-Student framework with shared parameters
    • Teacher video embedding model with implicit decoders
    • Student video embedding model (identical architecture)
    • Text alignment model
  • 4 loss functions (vs Prithvi's 1):
    1. Reconstruction: per-source conditional decoding (sensor-specific)
    2. Batch uniformity: minimizes absolute dot products between batch-rotated embeddings
    3. Consistency: temporal/spatial coherence across observations
    4. Text-contrastive: aligns satellite embeddings with geotagged Wikipedia articles

Training data: ~3 billion observations across ~1.1% of Earth's land surface, from 9+ gridded datasets:

Type Dataset Bands/Channels
Optical Sentinel-2 L1C 5 bands
Optical/Thermal Landsat 8/9 L1C 7 bands
C-band SAR Sentinel-1 GRD 5 channels
L-band SAR ALOS PALSAR-2 3 channels
Elevation Copernicus DEM GLO-30 1
LiDAR GEDI L2A 1
Climate ERA5-Land monthly multiple
Gravity GRACE monthly multiple
Land cover NLCD 2019/2021 1
Text Wikipedia geocoded articles N/A
Text GBIF research-grade observations N/A

Key characteristics:

  • Multi-resolution simultaneous pathways (1/16L, 1/8L, 1/2L) vs Prithvi's single resolution
  • 64-dim output constrained to unit hypersphere — directly usable for k-NN/linear probe
  • Handles 9+ heterogeneous data sources natively (no channel mismatch problem)
  • Produces pre-computed global annual embedding datasets (2017-2024) via GEE
  • Model weights NOT open-sourced — only embeddings are public

2. Side-by-Side Comparison

Dimension Prithvi-100M AlphaEarth Foundations
Architecture Standard ViT + MAE (single resolution) Custom STP multi-resolution pyramid (NOT ViT)
Parameters 100M 480M - 1B
Pre-training MAE self-supervised (mask & reconstruct pixels) Teacher-Student distillation + 4-loss joint training
Loss functions 1 (MSE reconstruction) 4 (reconstruction + uniformity + consistency + text-contrastive)
Training data US-only HLS (Landsat+S2), single optical source Global 3B observations, 9+ sources (optical+SAR+LiDAR+climate+gravity+text)
Input format Fixed 6 bands, (B,6,T,224,224) 9+ heterogeneous sources, any combination
Temporal 3D positional embedding, T frames stacked Dedicated Time Operator: time-axial self-attention + sinusoidal timecodes
Multi-resolution None — all patches equal Three-path pyramid with Laplacian rescaling
Output 768-dim token sequence (needs task head) 64-dim L2-normalized embedding (directly usable)
Embedding constraint None Unit hypersphere S^63 (von Mises-Fisher)
Text alignment None Yes — geotagged Wikipedia contrastive learning
Weights open Yes (Apache 2.0, HuggingFace) No — only pre-computed embeddings via GEE
Fine-tunable Yes — load weights, add head, PEFT/full tune No — can only consume pre-computed embeddings
Global coverage No pre-computed results 2017-2024 annual 10m global embedding dataset
Geographic bias US-centric (trained on CONUS only) Global (but 1.1% land surface)

3. Has Anyone Reproduced AlphaEarth?

3.1 Reproduction attempts

Only one known attempt: Brayden-Zhang/alphaearth-foundations (207 GitHub stars)

This unofficial PyTorch implementation:

  • Implements the STP three-path architecture
  • Implements the Teacher-Student-Text framework
  • BUT trained on only 1/40th of the Landsat subset
  • BUT used batch size 16 (paper uses 256)
  • BUT far from reaching paper's training scale (100K steps)

Verdict: Architecture verification only, not a true reproduction. No other team has attempted a full reproduction.

3.2 Why no one has reproduced it

Barrier 1 — Data: 3 billion observations from 9+ sources requires PB-scale storage and massive GEE compute quotas. Academic teams cannot access this at scale.

Barrier 2 — Compute: 480M-1B parameter model trained at 256 batch size x 100K steps requires large GPU/TPU clusters. Google DeepMind-level infrastructure. Not feasible with a few A100s.

Barrier 3 — Architecture complexity: STP three-path pyramid + Learned Laplacian Pyramid + Teacher-Student distillation + 4 losses = far more complex than standard ViT-MAE. Even if implemented, without sufficient data and compute, the trained model won't match Google's results.

3.3 Papers using AlphaEarth embeddings (without reproducing the model)

Instead of reproducing the model, researchers are consuming the pre-computed embedding dataset:

Pattern: The community treats AlphaEarth as an "embedding service" — consume the outputs, don't replicate the model.

4. Is AlphaEarth's Architecture Revolutionary?

4.1 What IS genuinely novel

Innovation Why it matters
Multi-resolution STP three-path processing (1/16L + 1/8L + 1/2L) Standard ViT processes all patches at one resolution. STP captures macro context and pixel detail simultaneously.
S^63 hypersphere embedding constraint Standard embeddings have no geometric structure. AlphaEarth's embeddings are directly comparable via cosine distance without additional training.
Native 9-source heterogeneous fusion Prithvi/SatMAE can only handle fixed-band inputs. AlphaEarth handles optical+SAR+LiDAR+climate natively at the architecture level.
Vision-language alignment First geospatial FM to do satellite-text contrastive learning, enabling semantic interpretability.
Global pre-computed embedding dataset Other models release weights and expect users to run inference. AlphaEarth releases the results directly, lowering the barrier to zero.

4.2 What is NOT revolutionary (just scale)

Aspect Assessment
MAE-style reconstruction loss Standard, used by SatMAE/Prithvi
Teacher-Student distillation Standard in SSL (BYOL, DINO)
Uniformity loss Known technique from contrastive learning literature
Transformer backbone Ubiquitous

4.3 Why traditional methods still dominate

Despite AlphaEarth's innovations, most real-world projects still use traditional approaches (ResNet/UNet + supervised training):

  1. When you have labels, you don't need a foundation model. Most operational remote sensing tasks (e.g., crop classification in a known region) have sufficient labeled data. A simple UNet trained on 10K labeled patches often beats a foundation model with a linear probe.

  2. Inference cost. Running a 480M model per-pixel at 10m resolution globally is expensive. A lightweight CNN is orders of magnitude cheaper.

  3. Foundation model advantage appears in specific scenarios: few-shot learning (very few labels), cross-domain transfer (train on US, deploy in China), and embedding-based retrieval (find similar locations). If you don't need these, traditional methods suffice.

5. Implications for Our Project

5.1 Why we chose Prithvi over AlphaEarth for GeoAdapter

Reason Detail
Prithvi weights are open-source Can load, freeze, and insert GeoAdapter in front. AlphaEarth weights are closed.
Prithvi has the "fixed input" problem 6-band HLS input lock-in = exactly the problem GeoAdapter solves
AlphaEarth has no input problem 9-source native fusion means GeoAdapter is irrelevant to AlphaEarth
Prithvi is fine-tunable Supports PEFT, which is the basis of our benchmark
AlphaEarth is consumption-only Can only use pre-computed embeddings, cannot fine-tune

5.2 How our two papers cover the ecosystem

Paper Foundation Model Approach Contribution
Paper 5 (world model) AlphaEarth (frozen embeddings) Consume embedding dataset, learn temporal dynamics on top Application layer — "what can you do with FM embeddings?"
GeoAdapter paper Prithvi-100M (frozen backbone) Adapt the model itself to accept heterogeneous inputs via PEFT Infrastructure layer — "how to make FMs work with real-world data?"

Together, these two papers provide a complete story: one shows how to efficiently use closed FM outputs (AlphaEarth), the other shows how to adapt open FM internals to real deployment constraints (Prithvi + GeoAdapter). This is a strong, non-overlapping portfolio.