arxiv: 2604.18881 · v1 · submitted 2026-04-20 · 💻 cs.CV · cs.AI

Recognition: unknown

A Proxy Consistency Loss for Grounded Fusion of Earth Observation and Location Encoders

Esther Rolf, Kevin Lane, Levi Cai, Morteza Karimzadeh, Zhongying Wang

Pith reviewed 2026-05-10 04:10 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords proxy consistency losslocation encoderearth observation fusiongeographic priorair quality predictionpoverty mappingsparse labelsdata imputation

0 comments

The pith

A proxy consistency loss trains location encoders on abundant geographic proxies to improve fusion with earth observation data for sparse-label predictions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a proxy consistency loss that aligns a trainable location encoder with proxy variables correlated to the target, allowing the encoder to absorb rich geographic information independently of scarce training labels. This produces embeddings that fuse more effectively with earth observation inputs than direct proxy concatenation or frozen pretrained priors. Experiments on air quality prediction and poverty mapping demonstrate gains in both in-sample accuracy, showing the loss successfully transfers proxy information, and out-of-sample generalization to regions without labels.

Core claim

By applying a proxy consistency loss to a location encoder, proxy data can be integrated implicitly as a geographic prior; the resulting embeddings outperform both joint input of proxies to an observation encoder and fusion with frozen pretrained location embeddings, delivering superior in-sample and out-of-sample performance on air quality and poverty tasks.

What carries the argument

The proxy consistency loss (PCL), a regularization term that enforces agreement between the location encoder's output and proxy variable values, which grounds the embeddings with proxy information while remaining trainable with limited labels.

If this is right

Proxy data sampled independently of labels can still improve prediction accuracy through the location encoder.
The learned embeddings support generalization beyond the spatial extent of training labels.
The approach beats both naive concatenation of proxies with observations and use of static geographic priors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same consistency-loss pattern could apply to any domain that pairs sparse labels with dense but imperfect proxies, such as species distribution or disease incidence mapping.
If the location encoder is made deeper or conditioned on additional metadata, the PCL might capture finer-scale geographic structure without extra labeled data.
Deployment in regions with rapidly changing proxies would require periodic re-training of the encoder to avoid stale consistency signals.

Load-bearing premise

Proxy variables must be correlated with the target in a way that transfers useful signal without systematic bias or proxy-specific overfitting.

What would settle it

Re-running the air quality or poverty experiments with the PCL disabled and observing no drop in out-of-sample accuracy relative to the frozen-embedding or direct-input baselines would indicate the loss adds no benefit.

Figures

Figures reproduced from arXiv: 2604.18881 by Esther Rolf, Kevin Lane, Levi Cai, Morteza Karimzadeh, Zhongying Wang.

**Figure 1.** Figure 1: Our model architecture fuses a trainable location encoder (blue) with an EO-based encoder model for prediction of a supervised [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Performance (R 2 ) on air quality task across checkerboard sizes δ ( ◦ ). As δ increases, performance generally degrades. Curves show the mean over 8 checkerboard partitions (4 spatial offsets × train/test swap), and shaded regions denote ±1 SE across partitions [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: PCL scaling with proxy sampling ratio ρ. Performance of trained location-encoder fusion with proxy consistency loss (PCL) as we vary the sampling ratio ρ. Left: UAR 50/50 spatial split. Right: systematically offset checkerboard split (δ = 8◦ ). Solid curves report mean R 2 and shaded regions denote ±1 standard error (SE) across runs/partitions. We compare two PCL sampling strategies: sampling proxy tar… view at source ↗

**Figure 5.** Figure 5: In the air quality task the EPA-only model fails to capture finer-grain temporal trends. The PCL-trained model instead shows temporal trends are dominant, as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

read the original abstract

Supervised learning with Earth observation inputs is often limited by the sparsity of high-quality labeled or in-situ measured data to use as training labels. With the abundance of geographic data products, in many cases there are variables correlated with - but different from - the variable of interest that can be leveraged. We integrate such proxy variables within a geographic prior via a trainable location encoder and introduce a proxy consistency loss (PCL) formulation to imbue proxy data into the location encoder. The first key insight behind our approach is to use the location encoder as an agile and flexible way to learn from abundantly available proxy data which can be sampled independently of training label availability. Our second key insight is that we will need to regularize the location encoder appropriately to achieve performance and robustness with limited labeled data. Our experiments on air quality prediction and poverty mapping show that integrating proxy data implicitly through the location encoder outperforms using both as input to an observation encoder and fusion strategies that use frozen, pretrained location embeddings as a geographic prior. Superior performance for in-sample prediction shows that the PCL can incorporate rich information from the proxies, and superior out-of-sample prediction shows that the learned latent embeddings help generalize to areas without training labels.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper introduces a proxy consistency loss to train location encoders from independently sampled geographic proxies, but the abstract's performance claims lack any supporting numbers or details.

read the letter

The paper's main contribution is a proxy consistency loss that trains a location encoder using independently sampled proxy geographic data for better Earth observation predictions under label scarcity. It outperforms direct fusion and frozen embedding approaches on air quality and poverty tasks according to the abstract. The novelty lies in the PCL setup, which allows the location encoder to learn from abundant proxies without requiring them to be paired with every training label. This is a practical angle for tasks like mapping where proxies such as other environmental variables are plentiful. The approach shows promise for improving both in-sample accuracy and generalization to unseen areas. It handles the problem of limited high-quality labels by leveraging correlated but distinct variables through the location prior. The idea of regularizing the encoder for robustness with scarce data is sound in concept. However, the abstract provides no actual metrics, baseline details, or ablation studies, so the superiority claims can't be evaluated from the summary alone. If the full paper has those, fine, but as presented, the central results lack support. There's also the risk that proxies introduce bias if their correlation with the target isn't strong or consistent across regions, which could undermine the out-of-sample gains the method aims for. The paper needs to demonstrate that the loss doesn't overfit to proxy-specific patterns. This paper targets researchers working on multimodal geospatial models and label-efficient learning in Earth observation. Readers dealing with air quality, poverty, or similar mapping problems might find the technique useful if the experiments are reproducible and the proxies are well-justified. It has enough substance to go to peer review rather than a desk reject. I'd recommend sending it for review, with the expectation that referees will want to see the quantitative results, code if available, and analysis of proxy correlation assumptions.

Referee Report

3 major / 0 minor

Summary. The manuscript proposes a Proxy Consistency Loss (PCL) to integrate proxy geographic variables into a trainable location encoder for Earth Observation (EO) fusion tasks. The central claim is that implicit integration of proxies via the location encoder with PCL regularization outperforms both direct concatenation of proxies and EO inputs to an observation encoder and fusion strategies that rely on frozen pretrained location embeddings; this is asserted to yield superior in-sample fits (showing rich proxy information capture) and out-of-sample generalization (showing better performance in label-sparse regions) on air quality prediction and poverty mapping.

Significance. If the experimental results hold under rigorous validation, the approach could meaningfully advance label-efficient EO modeling by offering a flexible mechanism to leverage abundant proxy data products for geographic priors without explicit fusion or additional labeling. The method targets a practical bottleneck in supervised EO learning and, if reproducible, would provide a useful regularization strategy for location encoders.

major comments (3)

Abstract: the claim of experimental superiority on air quality and poverty mapping tasks is asserted without any quantitative metrics, baseline details, ablation studies, or error analysis, leaving the central claim without verifiable support from the given text.
Method (PCL formulation): without the explicit loss equation, sampling procedure for proxies, or training details, it remains unclear whether the PCL introduces circularity via hyperparameters or proxy selection tuned on the same data used for final evaluation, undermining the positioning of PCL as an independent regularizer.
Experiments section: the central claim requires that chosen proxies are sufficiently correlated with the target (air quality, poverty) to transfer useful signal; no analysis of proxy-target correlation strength, regional variation, or potential bias from sensor artifacts is provided, which is load-bearing for the generalization argument.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback on our manuscript. We address each major comment point by point below, with revisions incorporated where they strengthen the presentation without altering the core contributions.

read point-by-point responses

Referee: Abstract: the claim of experimental superiority on air quality and poverty mapping tasks is asserted without any quantitative metrics, baseline details, ablation studies, or error analysis, leaving the central claim without verifiable support from the given text.

Authors: We agree that the abstract would benefit from explicit quantitative support. We have revised the abstract to include key performance metrics from the experiments section (specific RMSE reductions and correlation improvements relative to direct fusion and frozen-embedding baselines), along with brief references to the ablation studies and error analysis that demonstrate the superiority claims. revision: yes
Referee: Method (PCL formulation): without the explicit loss equation, sampling procedure for proxies, or training details, it remains unclear whether the PCL introduces circularity via hyperparameters or proxy selection tuned on the same data used for final evaluation, undermining the positioning of PCL as an independent regularizer.

Authors: The full manuscript presents the PCL loss equation in Section 3.2, the proxy sampling procedure in Section 3.1, and training details in Section 4.1. To eliminate any ambiguity regarding circularity, we have added a clarifying paragraph in the revised Section 3.3 stating that proxy variables were chosen from established geographic data products based on domain literature, independent of the target labels, and that hyperparameters were selected exclusively via cross-validation on training splits. revision: partial
Referee: Experiments section: the central claim requires that chosen proxies are sufficiently correlated with the target (air quality, poverty) to transfer useful signal; no analysis of proxy-target correlation strength, regional variation, or potential bias from sensor artifacts is provided, which is load-bearing for the generalization argument.

Authors: We acknowledge that an explicit correlation analysis would further substantiate the generalization argument. We have added a new subsection (4.3) in the revised experiments that reports proxy-target correlation strengths, examines regional variations, and discusses potential sensor artifacts, showing how the PCL consistency regularization helps transfer useful signal while mitigating bias effects in out-of-sample settings. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation and claims rest on independent empirical comparisons

full rationale

The paper introduces a trainable location encoder with a proxy consistency loss (PCL) to incorporate proxy variables as a geographic prior, then evaluates it via direct comparisons to baselines (joint input to observation encoder, frozen pretrained embeddings). No equations, loss formulations, or results in the provided text reduce the reported performance gains to the inputs by construction, fitted parameters renamed as predictions, or load-bearing self-citations. The superiority claims for in-sample and out-of-sample prediction are grounded in task-specific experiments on air quality and poverty mapping rather than definitional equivalence or imported uniqueness theorems.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that proxies can be sampled independently of labels and that consistency regularization will produce generalizable embeddings; no free parameters or invented entities are explicitly quantified in the abstract.

free parameters (1)

PCL weighting hyperparameter
A scalar balancing the proxy consistency term against the primary supervised loss is expected to be present and tuned on validation data.

axioms (1)

domain assumption Proxy variables are correlated with but distinct from the target variable of interest
Explicitly stated as the basis for leveraging abundant geographic data products when labels are sparse.

invented entities (1)

Proxy Consistency Loss (PCL) no independent evidence
purpose: Regularize the location encoder to incorporate proxy information implicitly
Newly introduced loss term whose effectiveness is the load-bearing element of the method.

pith-pipeline@v0.9.0 · 5517 in / 1378 out tokens · 47955 ms · 2026-05-10T04:10:58.038260+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

38 extracted references · 3 canonical work pages · 1 internal anchor

[1]

Alphaearth foundations: An embedding field model for accurate and efficient global mapping from sparse label data.arXiv preprint arXiv:2507.22291, 2025

Christopher F Brown, Michal R Kazmierski, Valerie J Pasquarella, William J Rucklidge, Masha Samsikova, Chenhui Zhang, Evan Shelhamer, Estefania Lahera, Olivia Wiles, Simon Ilyushchenko, et al. Alphaearth foundations: An embedding field model for accurate and efficient global mapping from sparse label data. arXiv preprint arXiv:2507.22291, 2025. 1

work page arXiv 2025
[2]

Geo-aware networks for fine-grained recognition

Grace Chu, Brian Potetz, Weijun Wang, Andrew Howard, Yang Song, Fernando Brucher, Thomas Leung, and Hartwig Adam. Geo-aware networks for fine-grained recognition. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion Workshops, pages 0–0, 2019. 2

2019
[3]

Considine, Jiayuan Hao, Priyanka deSouza, Danielle Braun, Colleen E

Ellen M. Considine, Jiayuan Hao, Priyanka deSouza, Danielle Braun, Colleen E. Reid, and Rachel C. Neth- ery. Evaluation of model-based pm2.5 estimates for exposure assessment during wildfire smoke episodes in the western u.s.Environmental Science & Technol- ogy, 57(5):2031–2041, 2023. PMID: 36693177. 1

2031
[4]

Latent domain model- ing improves robustness to geographic shifts.arXiv preprint arXiv:2503.02036, 2025

Ruth Crasto and Esther Rolf. Latent domain model- ing improves robustness to geographic shifts.arXiv preprint arXiv:2503.02036, 2025. 1, 2

work page arXiv 2025
[5]

Range: Retrieval augmented neural fields for multi-resolution geo-embeddings

Aayush Dhakal, Srikumar Sastry, Subash Khanal, Adeel Ahmad, Eric Xing, and Nathan Jacobs. Range: Retrieval augmented neural fields for multi-resolution geo-embeddings. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pages 24680–24689, 2025. 2

2025
[6]

Climplicit: Climatic implicit embeddings for global ecological tasks.International Conference on Learning Representations (ICLR) Workshops, 2025

Johannes Dollinger, Damien Robert, Elena Plekhanova, Lukas Drees, and Jan Dirk Wegner. Climplicit: Climatic implicit embeddings for global ecological tasks.International Conference on Learning Representations (ICLR) Workshops, 2025. 5

2025
[7]

Vi- irs night-time lights.International journal of remote sensing, 38(21):5860–5879, 2017

Christopher D Elvidge, Kimberly Baugh, Mikhail Zhizhin, Feng Chi Hsu, and Tilottama Ghosh. Vi- irs night-time lights.International journal of remote sensing, 38(21):5860–5879, 2017. 4

2017
[8]

Long short-term memory.Supervised sequence labelling with recurrent neural networks, pages 37–45, 2012

Alex Graves. Long short-term memory.Supervised sequence labelling with recurrent neural networks, pages 37–45, 2012. 5

2012
[9]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 5

2016
[10]

Momentum contrast for unsupervised visual representation learning

Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020. 5

2020
[11]

Enhancing and interpreting deep learning for sea ice charting using the autoice benchmark.Remote Sensing Applications: Society and Environment, 38:101538,

Sepideh Jalayer, Samira Alkaee Taleghan, Rafael Pires de Lima, Behzad Vahedi, Nick Hughes, Farnoush Banaei-Kashani, and Morteza Karimzadeh. Enhancing and interpreting deep learning for sea ice charting using the autoice benchmark.Remote Sensing Applications: Society and Environment, 38:101538,
[12]

Combining satellite imagery and machine learning to predict poverty.Science, 353(6301):790–794, 2016

Neal Jean, Marshall Burke, Michael Xie, W Matthew Alampay Davis, David B Lobell, and Stefano Ermon. Combining satellite imagery and machine learning to predict poverty.Science, 353(6301):790–794, 2016. 1, 2, 3

2016
[13]

Performance and generalizability impacts of incorporating location encoders into deep learning for dynamic PM2

Morteza Karimzadeh, Zhongying Wang, and James L Crooks. Performance and generalizability impacts of incorporating location encoders into deep learning for dynamic PM2. 5 estimation.GIScience & Remote Sensing, 62(1):2594797, 2025. 1, 2, 5

2025
[14]

Satclip: Global, general-purpose location embeddings with satellite imagery

Konstantin Klemmer, Esther Rolf, Caleb Robinson, Lester Mackey, and Marc Rußwurm. Satclip: Global, general-purpose location embeddings with satellite imagery. InProceedings of the AAAI Conference on Artificial Intelligence, pages 4347–4355, 2025. 2, 3

2025
[15]

Earth embeddings: To- wards ai-centric representations of our planet

Konstantin Klemmer, Esther Rolf, Marc Russwurm, Gustau Camps-Valls, Mikolaj Czerkawski, Stefano Ermon, Alistair Francis, Nathan Jacobs, Hannah Rae Kerner, Lester Mackey, et al. Earth embeddings: To- wards ai-centric representations of our planet. 2025. 1

2025
[16]

Rajesh Kumar, Piyush Bhardwaj, Cenlin He, Jennifer Boehnert, Forrest Lacey, Stefano Alessandrini, Kevin Sampson, Matthew Casali, Scott Swerdlin, Olga Wil- helmi, et al. A long-term high-resolution air quality reanalysis with a public-facing air quality dashboard over the contiguous united states (conus).Earth Sys- tem Science Data, 17(5):1807–1834, 2025. 4, 1

2025
[17]

Decoupled Weight Decay Regularization

Ilya Loshchilov and Frank Hutter. Decou- pled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 1

work page internal anchor Pith review Pith/arXiv arXiv 2017
[18]

Effective approaches to attention-based neural machine translation

Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. InProceedings of the 2015 conference on empirical methods in natural language processing, pages 1412–1421, 2015. 5

2015
[19]

Presence-only geographical priors for fine-grained im- age classification

Oisin Mac Aodha, Elijah Cole, and Pietro Perona. Presence-only geographical priors for fine-grained im- age classification. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV),
[20]

Csp: Self-supervised contrastive spatial pre-training for geospatial-visual representa- tions

Gengchen Mai, Ni Lao, Yutong He, Jiaming Song, and Stefano Ermon. Csp: Self-supervised contrastive spatial pre-training for geospatial-visual representa- tions. InInternational Conference on Machine Learn- ing, pages 23498–23515. PMLR, 2023. 2

2023
[21]

Mmearth: Exploring multi-modal pretext tasks for geospatial representation learning

Vishal Nedungadi, Ankit Kariryaa, Stefan Oehm- cke, Serge Belongie, Christian Igel, and Nico Lang. Mmearth: Exploring multi-modal pretext tasks for geospatial representation learning. InEuropean Con- ference on Computer Vision, pages 164–182. Springer,
[22]

Using multiple in- put modalities can improve data-efficiency and OOD generalization for ML with satellite imagery

Arjun Rao and Esther Rolf. Using multiple in- put modalities can improve data-efficiency and OOD generalization for ML with satellite imagery. In TerraBytes-Towards global datasets and models for Earth Observation, pages 166–188. PMLR, 2025. 1, 2, 3

2025
[23]

A generaliz- able and accessible approach to machine learning with global satellite imagery.Nature communications, 12 (1):4392, 2021

Esther Rolf, Jonathan Proctor, Tamma Carleton, Ian Bolliger, Vaishaal Shankar, Miyabi Ishihara, Ben- jamin Recht, and Solomon Hsiang. A generaliz- able and accessible approach to machine learning with global satellite imagery.Nature communications, 12 (1):4392, 2021. 4

2021
[24]

Chlorophyll-a and total suspended solids retrieval and mapping us- ing sentinel-2a and machine learning for inland wa- ters.Ecological Indicators, 113:106236, 2020

Mohammadmehdi Saberioon, Jakub Brom, V ´aclav Nedbal, Pavel Sou˘cek, and Petr C ´ısa˘r. Chlorophyll-a and total suspended solids retrieval and mapping us- ing sentinel-2a and machine learning for inland wa- ters.Ecological Indicators, 113:106236, 2020. 1

2020
[25]

Geosynth: Contextually-aware high- resolution satellite image synthesis

Srikumar Sastry, Subash Khanal, Aayush Dhakal, and Nathan Jacobs. Geosynth: Contextually-aware high- resolution satellite image synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 460–470, 2024. 2

2024
[26]

Vicedo-Cabrera, Francesco Sera, Pierre Masselot, Massimo Stafoggia, Kees de Hoogh, Itai Kloog, Stefan Reis, Massimo Vieno, and Antonio Gasparrini

Rochelle Schneider, Ana M. Vicedo-Cabrera, Francesco Sera, Pierre Masselot, Massimo Stafoggia, Kees de Hoogh, Itai Kloog, Stefan Reis, Massimo Vieno, and Antonio Gasparrini. A satellite-based spatio-temporal machine learning model to recon- struct daily pm2.5 concentrations across great britain. Remote Sensing, 12(22), 2020. 1

2020
[27]

Bidirectional re- current neural networks.IEEE transactions on Signal Processing, 45(11):2673–2681, 1997

Mike Schuster and Kuldip K Paliwal. Bidirectional re- current neural networks.IEEE transactions on Signal Processing, 45(11):2673–2681, 1997. 5

1997
[28]

Multimae meets earth observation: Pre-training multi-modal multi-task masked autoen- coders for earth observation tasks

Jose Sosa, Danila Rukhovich, Anis Kacem, and Djamila Aouada. Multimae meets earth observation: Pre-training multi-modal multi-task masked autoen- coders for earth observation tasks. In2025 IEEE In- ternational Conference on Image Processing (ICIP), pages 797–802. IEEE, 2025. 1

2025
[29]

Ssl4eo-l: Datasets and foundation models for landsat imagery.Advances in Neural Information Processing Systems, 36:59787–59807, 2023

Adam Stewart, Nils Lehmann, Isaac Corley, Yi Wang, Yi-Chia Chang, Nassim Ait Ait Ali Braham, Shradha Sehgal, Caleb Robinson, and Arindam Banerjee. Ssl4eo-l: Datasets and foundation models for landsat imagery.Advances in Neural Information Processing Systems, 36:59787–59807, 2023. 5

2023
[30]

Geoclip: Clip-inspired alignment be- tween locations and images for effective worldwide geo-localization.Advances in Neural Information Processing Systems, 36:8690–8701, 2023

Vicente Vivanco Cepeda, Gaurav Kumar Nayak, and Mubarak Shah. Geoclip: Clip-inspired alignment be- tween locations and images for effective worldwide geo-localization.Advances in Neural Information Processing Systems, 36:8690–8701, 2023. 1, 2, 4, 5

2023
[31]

Mtp: Advancing re- mote sensing foundation model via multitask pretrain- ing.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:11632–11654,

Di Wang, Jing Zhang, Minqiang Xu, Lin Liu, Dong- sheng Wang, Erzhong Gao, Chengxi Han, Haonan Guo, Bo Du, Dacheng Tao, et al. Mtp: Advancing re- mote sensing foundation model via multitask pretrain- ing.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:11632–11654,
[32]

High-resolution es- timation of daily pm2

Zhongying Wang, James L Crooks, Elizabeth Anne Regan, and Morteza Karimzadeh. High-resolution es- timation of daily pm2. 5 levels in the contiguous us using bi-lstm with attention.Remote Sensing, 17(1): 126, 2025. 2, 4, 5, 1

2025
[33]

Torchspatial: A location encoding framework and benchmark for spatial repre- sentation learning.Advances in Neural Information Processing Systems, 37:81437–81460, 2024

Nemin Wu, Qian Cao, Zhangyu Wang, Zeping Liu, Yanlin Qi, Jielu Zhang, Joshua Ni, Xiaobai Yao, Hongxu Ma, Lan Mu, et al. Torchspatial: A location encoding framework and benchmark for spatial repre- sentation learning.Advances in Neural Information Processing Systems, 37:81437–81460, 2024. 4

2024
[34]

Fifty years of landsat sci- ence and impacts.Remote Sensing of Environment, 280:113195, 2022

Michael A Wulder, David P Roy, V olker C Radeloff, Thomas R Loveland, Martha C Anderson, David M Johnson, Sean Healey, Zhe Zhu, Theodore A Scam- bos, Nima Pahlevan, et al. Fifty years of landsat sci- ence and impacts.Remote Sensing of Environment, 280:113195, 2022. 4

2022
[35]

Downscaling SMAP soil moisture using a wide & deep learning method over the Continen- tal United States.Journal of Hydrology, 609:127784,

Mengyuan Xu, Ning Yao, Haoxuan Yang, Jia Xu, An- nan Hu, Luis Gustavo Goncalves de Goncalves, and Gang Liu. Downscaling SMAP soil moisture using a wide & deep learning method over the Continen- tal United States.Journal of Hydrology, 609:127784,
[36]

Using publicly available satellite imagery and deep learning to understand eco- nomic well-being in africa.Nature communications, 11(1):2583, 2020

Christopher Yeh, Anthony Perez, Anne Driscoll, George Azzari, Zhongyi Tang, David Lobell, Stefano Ermon, and Marshall Burke. Using publicly available satellite imagery and deep learning to understand eco- nomic well-being in africa.Nature communications, 11(1):2583, 2020. 1

2020
[37]

Sus- tainbench: Benchmarks for monitoring the sustainable development goals with machine learning

Christopher Yeh, Chenlin Meng, Sherrie Wang, Anne Driscoll, Erik Rozi, Patrick Liu, Jihyeon Lee, Mar- shall Burke, David Lobell, and Stefano Ermon. Sus- tainbench: Benchmarks for monitoring the sustainable development goals with machine learning. InThirty- fifth Conference on Neural Information Processing Systems, Datasets and Benchmarks Track (Round 2),
[38]

Air quality prediction details A.1

4 A Proxy Consistency Loss for Grounded Fusion of Earth Observation and Location Encoders Supplementary Material A. Air quality prediction details A.1. Dataset details Earth observation features used in the air quality prediction task include time series of: Moderate Resolution Imaging Spectroradiometer (MODIS) Multi-Angle Implementation of Atmospheric Co...

2010