Inferring Height from Earth Embeddings: First insights using Google AlphaEarth

Alireza Hamoudzadeh; Valeria Belloni; Roberta Ravanelli

cs.CVFeb 19, 2026

Inferring Height from Earth Embeddings: First insights using Google AlphaEarth

Authors: Alireza Hamoudzadeh, Valeria Belloni, Roberta Ravanelli

Organizations: Geodesy and Geomatics Division, DICEA, Sapienza University of Rome, Rome, 00184, Italy · Geomatics Unit, Department of Geography, Faculty of Sciences, University of Liège, Liège, 4000, Belgium

Abstract

This study investigates whether the geospatial and multimodal features encoded in 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.

Figures & tables

Attribute	AlphaEarth	TESSERA
Spatial coverage	Global	Global (sparse)
Temporal coverage	2017-2024	2017-2024 (not continuous)
Time granularity	Annual	Annual
Spatial resolution	10 m	10 m
Embedding dimension	64 bands	128 bands
Modality	Multi-modal	Multi-modal

Table 1: Features of main Earth Embedding datasets.

Author	Task	Area	Resolution
Khan and Ahmad (2025)	Land Use classification	19.3 km 2	10 m
Alvarez et al. (2025)	Urban Air Quality prediction	60 points *	-
Ma et al. (2025)	Agricultural tasks	47709 points *	-
Fang et al. (2025)	Yield estimation	992 points *	-
Murakami (2025)	Crop classification	270 points *	-
Lucero et al. (2026)	Biomass estimation	13 km 2	10 m

Table 2: Details of the area and resolution of the study for each downstream task investigated using AlphaEarth Embeddings. * The analysis is done in vectorial form and outside the native raster domain.

Figure 1: Study area: training and validation region in blue ( $\approx 5550$ km 2 ) and testing region in red ( $\approx 2315$ km 2 ).

Panel of Figure 2 — Figure 2: Simplified overview of the proposed approach for training/validation (a) and testing (b).

Parameter	Value
Input data	AlphaEarth Embedding at native 10 m resolution
Target data	IGN RGE Alti DSM at 10 m resampled resolution
Architecture	U-Net / U-Net++
Encoder	ResNet18
Input channels	64
Output channels	1

Table 3: Parameters and hyperparameters of the U-Net and U-Net++ models.

Figure 3: Training and validation losses and best model (U-Net) along with the corresponding loss, learning rate, and epoch number.

Figure 4: Training and validation losses and best model (U-Net++) along with the corresponding loss, learning rate, and epoch number.

Panel of Figure 5 — Figure 5: Correlation between inferred height and reference height for each model (training). The black diagonal line y = x represents the best possible fit; the blue line is the actual fit to the plot.

Panel of Figure 6 — Figure 6: Qualitative results: Reference height and inferred height (meters) from the tested models (training).

Panel of Figure 7 — Figure 7: Correlation between inferred height and reference height for each model (testing). The black diagonal plot y=x represents the best possible fit; the blue line is the actual fit to the plot.

Panel of Figure 8 — Figure 8: Qualitative results: Reference height and inferred height (meters) from the tested models (testing).

	U-Net		U-Net++		Ridge Regression
	Train	Test	Train	Test	Train	Test
Performance metrics
$R^{2}$	0.97	0.78	0.97	0.84	0.64	0.38
Pearson correlation	0.99	0.91	0.99	0.92	0.80	0.73
Spearman correlation	0.99	0.90	0.99	0.91	0.81	0.69
Difference metrics

Table 4: Model performance and difference statistics for U-Net, U-Net++, and Ridge Regression on training and test datasets. All statistics are computed between the inferred heights and the reference DSM.

Figure 9: Histogram of the height distribution of DSM over the test and train sets.

Remote Sensing

Remote sensing uses satellite, aerial, and ground-based observations to measure and map Earth and other planetary surfaces.

See all 528 papers

May 20, 2026

Mapping Tomato Cropping Systems in California Using AlphaEarth Geospatial Embeddings and Deep Learning Analysis

Field-scale crop maps support supply-chain forecasting and policy, yet statewide crop identification still often depends on retrospective surveys or remote-sensing workflows built around hand-engineered spectral features. Those pipelines can be accurate, but they require repeated preprocessing and often lose robustness across years. This study evaluated whether Google DeepMind's AlphaEarth geospatial embeddings can serve as an analysis-ready alternative for mapping processing tomato systems in California. LandIQ 2018 crop polygons were used to assemble a balanced reference dataset of 4,742 tomato and 4,742 non-tomato fields. For each polygon, 64-band AlphaEarth embedding chips were extracted and aligned with binary masks, then divided into spatially independent training (n = 6,638), validation (n = 1,422), and test (n = 1,424) sets. A U-Net segmentation model was trained on AWS SageMaker using a composite masked binary cross-entropy and soft Dice loss. To complement hard predictions, Monte Carlo dropout was retained at inference and repeated 100 times per chip to estimate predictive mean and variance. On the independent test set, the model achieved 99.19% pixel accuracy, 98.69% precision, 99.40% recall, 99.04% F1 score, 98.11% intersection over union, and 99.02% chip accuracy. Uncertainty maps were consistently highest near field edges and low within field interiors. The results show that AlphaEarth embeddings retain crop-relevant spatial and temporal structure and can support accurate, field-scale tomato mapping without manual feature engineering.

Mohammadreza Narimani, Alireza Pourreza, Parastoo Farajpoor

Department of Biological and Agricultural Engineering, University of California, Davis, Davis, CA, 95616, USA

Remote Sensing Image Segmentation Agricultural Remote Sensing Geospatial Foundation Models

Ask PDF

Mar 26, 2026

GeoHeight-Bench: Towards Height-Aware Multimodal Reasoning in Remote Sensing

Current Large Multimodal Models (LMMs) in Earth observation are predominantly evaluated on planar optical tasks and often neglect the vertical dimension, although vertical geometric structure can be critical in applications such as disaster response and urban-morphology analysis. Progress on height-aware reasoning is also hindered by the absence of systematic evaluation: few benchmarks pair optical imagery with height products across pixel-, object-, and scene-level reasoning. To address this gap, we introduce GeoHeight-Bench, a large-scale benchmark for height-aware remote sensing understanding, together with a more challenging terrain-oriented extension, GeoHeight-Bench+. The benchmark is constructed through a scalable, VLM-driven generation pipeline that combines metadata extraction with prompt engineering, and its quality is assessed through a human-in-the-loop verification protocol. To examine whether height-aware reasoning can be learned from optical imagery, we further provide GeoHeightChat, a height-aware baseline that transfers implicit height-related geometric representations into an optical LMM. Evaluations of a broad range of closed- and open-source LMMs show that current models remain limited in their ability to reason about height information, while aligning implicit height priors improves most height-dependent tasks. However, several tasks, particularly slope reasoning and terrain-based flood-susceptibility mapping, remain largely unsolved, highlighting concrete open problems for height-aware GeoAI. Dataset and Code will be released here.

Xuran Hu, Zhitong Xiong, Zhongcheng Hong +3

KTH Royal Institute of Technology · Technical University of Munich · The Hong Kong University of Science and Technology (Guangzhou)

Remote Sensing Image Understanding Large Vision-Language Models Remote Sensing Geospatial Reasoning Remote Sensing Benchmarks Multimodal Model Evaluation

Ask PDF

Aug 3, 2026

Deep Evidential Regression for Sparse Forest Height Estimation from Multimodal Satellite Imagery

Accurate estimation of forest height from satellite imagery is essential for applications such as carbon accounting, biodiversity monitoring, and ecosystem management. While recent deep learning approaches provide accurate predictions, they typically do not quantify predictive uncertainty. This limitation is particularly relevant in geospatial settings characterized by sparse supervision and geographic distribution shift. In this work, we investigate Deep Evidential Regression (DER) for forest height estimation on the TreeUQ benchmark, a large-scale dataset designed for the joint estimation of tree count and average tree height at 10 m resolution, based on Sentinel-1/-2 data as well as tree inventory data over the federal state of Bavaria. To account for the extreme label sparsity of the tree inventory data, we introduce a masked evidential loss for dense geospatial prediction. Using a U-Net architecture with multimodal Sentinel-1 and Sentinel-2 inputs, the proposed approach jointly predicts tree height and associated uncertainty estimates in a single forward pass. Experimental results show that DER achieves predictive performance comparable to a deterministic U-Net while additionally providing well-calibrated uncertainty estimates. These findings demonstrate the potential of evidential learning as an efficient framework for uncertainty-aware forest structure estimation from Earth observation data.

Laura Bader, Muhammad Ammar Ahmed, Xiao Xiang Zhu +1

LMU Munich, 80539 Munich, Germany · Munich Center for Machine Learning (MCML) · TU Munich, 80333 Munich

Remote Sensing Image Understanding Evidential Deep Learning Uncertainty Quantification Remote Sensing Uncertainty Calibration

Ask PDF

Abstract

Remote Sensing

Explore similar work

Mapping Tomato Cropping Systems in California Using AlphaEarth Geospatial Embeddings and Deep Learning Analysis

GeoHeight-Bench: Towards Height-Aware Multimodal Reasoning in Remote Sensing

Deep Evidential Regression for Sparse Forest Height Estimation from Multimodal Satellite Imagery