[Paper Review] Inferring Height from Earth Embeddings: First insights using Google AlphaEarth

Alireza Hamoudzadeh, Valeria Belloni+1 more|arXiv (Cornell University)|Feb 19, 2026

Remote Sensing and LiDAR Applications0 citations

TL;DR

This paper evaluates whether AlphaEarth Embeddings can guide lightweight DL models (U-Net and U-Net++) to infer regional surface height from 10 m AlphaEarth data, comparing against a ridge baseline and assessing generalization to a new area.

ABSTRACT

This study investigates whether the geospatial and multimodal features encoded in extit{Earth Embeddings} can effectively guide deep learning (DL) regression models for regional surface height mapping. In particular, we focused on AlphaEarth Embeddings at 10 m spatial resolution and evaluated their capability to support terrain height inference using a high-quality Digital Surface Model (DSM) as reference. U-Net and U-Net++ architectures were thus employed as lightweight convolutional decoders to assess how well the geospatial information distilled in the embeddings can be translated into accurate surface height estimates. Both architectures achieved strong training performance (both with $R^2 = 0.97$), confirming that the embeddings encode informative and decodable height-related signals. On the test set, performance decreased due to distribution shifts in height frequency between training and testing areas. Nevertheless, U-Net++ shows better generalization ($R^2 = 0.84$, median difference = -2.62 m) compared with the standard U-Net ($R^2 = 0.78$, median difference = -7.22 m), suggesting enhanced robustness to distribution mismatch. While the testing RMSE (approximately 16 m for U-Net++) and residual bias highlight remaining challenges in generalization, strong correlations indicate that the embeddings capture transferable topographic patterns. Overall, the results demonstrate the promising potential of AlphaEarth Embeddings to guide DL-based height mapping workflows, particularly when combined with spatially aware convolutional architectures, while emphasizing the need to address bias for improved regional transferability.

Motivation & Objective

• Assess whether AlphaEarth Embeddings encode decodable height-related signals for regional surface height inference.
• Evaluate lightweight CNN decoders (U-Net and U-Net++) to map embeddings to high-quality DSM-derived height.
• Compare DL approaches against a linear Ridge baseline on training and unseen regions.
• Investigate generalization and transferability of embedding-guided height mapping across a large study area.

Proposed method

• Use AlphaEarth Embeddings at 10 m resolution as input features.
• Train two lightweight decoders (U-Net and U-Net++) with a ResNet-18 encoder on 512x512 patches.
• Normalize embedding inputs from uint8 to [0,1] and align with a 5x5 m DSM target resampled to 10 m.
• Optimize with AdamW, MSE loss, ReduceLROnPlateau, and early stopping; 80/20 train/validation split and a separate test area.
• Compare against Ridge regression baseline using identical inputs and evaluation metrics.

Experimental results

Research questions

• RQ1 Can AlphaEarth Embeddings effectively guide DL regression to predict regional terrain height from a DSM reference?
• RQ2 Do lightweight, spatially aware decoders preserve height-related spatial patterns and generalize to new regions?
• RQ3 How does embedding-guided height inference compare to a traditional linear baseline?
• RQ4 What are the strengths and limitations of transferring embedding-based height models across areas with different height distributions?

Key findings

• Both U-Net and U-Net++ achieve strong training performance (R2 ≈ 0.97) indicating embeddings encode decodable height signals.
• On testing data, U-Net++ generalizes better (R2 ≈ 0.84) than U-Net (R2 ≈ 0.78), showing improved robustness to distribution shifts.
• U-Net++ yields lower RMSE (~16 m) on the test set and exhibits closer height support to reference DSM than the Ridge baseline.
• DL models outperform Ridge regression on both training and test sets, demonstrating the value of embedding-driven height mapping.
• Height inference remains challenged by biases and residual distribution differences, underscoring transferability limitations across regions.