acceptodds
Under review as a conference paper at ICLR 2027

Planetary Feature Fields are Scalable Earth Representations

40 views

Abstract

Satellite observations, precomputed embeddings, and map products describe the same evolving Earth, yet are stored as independent, petabyte-scale data products. Their continued growth calls for compact representations of multiple products while preserving spatial and temporal detail. We introduce Planetary Feature Fields (PFFs), which exploit redundancy across data products by modeling them jointly as continuous functions of space and time at planetary scale. PFFs are spatially local explicit–implicit (hybrid) neural fields. Each field shares a factored feature volume––a decomposition of an explicit 3D grid with smaller factors––across products, while lightweight implicit decoders reconstruct individual products across multiple timesteps. PFFs reconstruct EO products over space and time more accurately than single-product fields at matched compression rates. At compression relative to the uncompressed source data, reconstructed features retain approximately 90% or more of the performance achieved with the original features on pixel-level segmentation, change detection, and patch-level classification tasks. PFFs can add new timesteps by extending their factored feature volumes and add new products by attaching new decoders, while leaving existing outputs unchanged. PFFs reduce end-to-end feature access latency by an order of magnitude relative to evaluated API and cloud-storage pipelines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Neural Fields, Representation Learning, Earth Observation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Y4CPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents Planetary Feature Fields (PFFs), which jointly compress heterogeneous Earth-observation products using local, factored spatiotemporal neural fields with product-specific decoders. PFFs allow new timesteps or products to be added while preserving existing outputs bit-exactly. Evaluation on 14 products spanning 2017–2025 shows that PFFs achieve better average reconstruction quality than storage-matched single-product neural fields, generally retain at least 90% of source-feature performance across several downstream tasks, and provide faster feature retrieval than the evaluated API and cloud-storage pipelines. Continual-learning experiments further demonstrate efficient adaptation to new timesteps and products without changing previously supported outputs.

Critical strengths

Strong empirical validation at a meaningful scale: the paper evaluates 14 heterogeneous EO products spanning 2017–2025, with reconstruction experiments covering 100 global tiles and 16 TB of source data, plus downstream-task and retrieval-latency evaluations. This breadth makes the approach's practical relevance more convincing than a small-scale neural-field demonstration.

Critical weaknesses
  1. The presented evidence does not fully substantiate the claim of a general-purpose Earth representation. Most downstream evaluations rely on reconstructed products instead of the shared latent representation, and each local field is fitted specifically to its region without demonstrating transferability to previously unseen areas. Consequently, the results establish PFF more convincingly as a multi-product neural compression system rather than as a broadly transferable Earth representation, which undermines the central framing of the paper.
  2. The compression evaluation does not include sufficiently competitive non-neural-field baselines. The comparisons focus primarily on storage-matched coordinate networks and neural fields, without direct evaluation against multivariate autoencoders, tensor compression methods, scientific-data codecs, or related latent-field approaches. Furthermore, because the reported compression ratio depends on product dimensionality and redundancy, the current experiments do not show whether PFF competes with the most robust alternative compression paradigms.
Questions for authors
  1. Most downstream results evaluate reconstructed product features rather than the shared PFF latent, while the direct probe of the shared feature appears substantially weaker. What evidence supports framing PFF as a general-purpose Earth representation rather than primarily a multi-product compression system?
  2. How is the reported ~1800x compression ratio calculated, including product decoders, metadata, numerical precision, and missing coverage? Do the reconstruction and storage advantages hold when you exclude redundant high-dimensional embeddings and compare PFF with competitive conventional or learned codecs under identical data composition and quality targets?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Y4CP (1/2)

Thank you for critically engaging with our work, and the positive feedback.

Responses to weaknesses and questions

[W1,Q1] On claims surrounding a general-purpose Earth representation vs multi-product compression system

Thank you for the insightful question. We do not claim that PFFs provide a new general-purpose embedding or transfer to unobserved regions without fitting. Existing methods that provide reusable Earth representations are in plentiful supply (for example, AlphaEarth, TESSERA, Copernicus-FM, or CROMA). Materializing our shared latent as another dense, global embedding archive would reintroduce the storage and serving costs that motivate PFFs. Each regional field is fitted independently, without requiring its latent channels to have consistent meanings across tiles. The product decoders map these internal features into consistent product feature spaces, enabling downstream use across regions.

Multi-product neural compression is central to our approach. PFF’s neural compression is however achieved by learning a shared Earth representation. As in neural radiance fields and hybrid INRs [1,2], the data are encoded in a fitted field and recovered through learned decoding. PFFs use this formulation to represent heterogeneous Earth products within a common regional field. Figure 4 shows that sharing this representation improves reconstruction over separate fields at matched storage and preserves important predictive information (Table 1). These results support a learned Earth representation for compression and retrieval. A general-purpose embedding that transfers to unseen regions without fitting would be a stronger claim that we do not make.

[W2]

[W2a] Direct evaluation against multivariate autoencoders, tensor compression methods, scientific-data codecs, or related latent-field approaches

Thank you for prompting us to incorporate additional non-neural field baselines to our manuscript. We added SZ3, a prediction-based scientific-data compressor, and TTHRESH, a tensor compressor to Appendix Table 3a of our manuscript. In addition, we add several lossless compression methods and tabulate their final storage per-tile in Table 3b. The comparison below uses the same ten-tile PFFBench subset of all 14 products, and the same reconstruction metrics, with a total budget of 128 MB per tile (canonical PFF size).

MethodCompression approachReconstruction score (%)
JPEG2000Wavelet-based image compression86.69 +/- 4.58
SZ3Error-bounded scientific-data compression21.30 +/- 5.56
TTHRESHTensor decomposition83.97 +/- 5.08
PFFShared neural field95.20 +/- 0.22

We generally find that the evaluated lossy compression methods require approximately 16 times the bit budget to beat PFFs on the 14-product reconstruction task. At matched storage, PFFs beat these lossy codecs by ~9 reconstruction points.

[W2b] Furthermore, because the reported compression ratio depends on product dimensionality and redundancy, the current experiments do not show whether PFF competes with the most robust alternative compression paradigms.

Exploiting redundancy, both within a product and across products, is an explicit goal of PFFs. We discuss how product composition affects the compression ratio in Section 6. Additionally, to ensure fair comparisons, we evaluate reconstruction at matched storage budgets using the same products, spatial extent, and observed years for each method. This measures how well each method represents the same data within the same budget (Table 6). Because all methods represent the same source data within the same storage budget, PFFs’ reconstruction advantage cannot be attributed to a larger compression-ratio numerator.

AuthorsOpenReview ↗

Response to Reviewer Y4CP (2/2)

[Q2]

[Q2a] How is the reported ~1800x compression ratio calculated, including product decoders, metadata, numerical precision, and missing coverage?

The exact reported ratio in our paper is 227.712 GB / 0.128 GB = 1,779x. The numerator uses the products’ native (quantized) shipped precision, including the int8 quantization of AlphaEarth and TESSERA precomputed embeddings. Absent product-years and missing products as well as nodata values are excluded from the numerator calculation.

The denominator includes the final PFF checkpoint with fp16 weights, all product decoders, and metadata. Across over 1000 PFFs trained, our checkpoints are of the following structure:

1. 121.07 MB of feature-volume and basis parameters.
2. 6.86 MB of decoders.
3. 0.07 MB of metadata (for example, the coordinate transform needed to map a geographic location into a spatial location for a specific product within a PFF).

Note that as mentioned in Section 6 of our paper, our compression ratios can change significantly depending on product composition, as well as the processing stage we consider to be the numerator. For example, the dequantized products at float32 precision (normally recommended before being used for analysis [3,4]), increase the numerator to 534.45 GB, giving approximately 4,175x compression. We therefore state the source precision explicitly in Table 2 and report separate comparisons against lossless compression techniques applied to the quantized embeddings. These comparisons are tabulated below (and have been added to Table 3b of the revised manuscript).

MethodStorage per tile (GB)Size relative to PFF
DEFLATE84.86 +/- 4.65(663 +/- 36)x
ZSTD84.28 +/- 4.64(658 +/- 36)x
LZMA83.45 +/- 4.49(652 +/- 35)x
LERC, zero error139.14 +/- 3.86(1,087 +/- 30)x
JPEG2000, lossless85.32 +/- 4.98(667 +/- 39)x
PFF0.1281x

Overall, PFFs are ~650x smaller than the best lossless codec evaluated, and are ~9 reconstruction score percentage points better than the best lossy codec evaluated at matched storage.

[Q2b] Do the reconstruction and storage advantages hold when you exclude redundant high-dimensional embeddings and compare PFF with competitive conventional or learned codecs under identical data composition and quality targets?

Precomputed embeddings are part of the storage problem that motivates PFFs. Our evaluation therefore includes them alongside raw satellite observations and derived maps. Each codec receives the same product stack, so exploiting redundancy within these products is an opportunity available to every method. Our results cover a reasonably large, heterogeneous, and representative product stack. We leave a systematic study of which products compress best together, including subsets without AlphaEarth and TESSERA, to future work.

[1] Mildenhall et al. “NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis.” ECCV, 2020.

[2] Fridovich-Keil et al. “K-Planes: Explicit Radiance Fields in Space, Time, and Appearance.” CVPR, 2023.

[3] Google. (2026, July 7). AlphaEarth Foundations GCS data. Google for Developers.

[4] GeoTessera. (n.d.). Architecture guide: Quantization system. GeoTessera documentation.

Reviewer_RH44OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PFFs to jointly compress different EO data. The experimental results show a good reconstruction-storage budget balance and faster feature retrieval.

Critical strengths
  1. The evaluation of this paper is comprehensive. The paper reports over 6 million data points for the evaluation across various tiles, time and EO products.
  2. The authors propose to freeze existing parameters and preserve temporal interpolation coordinates. I think this is a smart design for adding new years without changing previous outputs.
Critical weaknesses
  1. In table 5, bit-proportional baselines actually outperform PFF on embedding rec, but the authors didn't evaluate their downstream performance. I think this should be added.
  2. I think the compression baselines should also include conventional methods like codecs.
Questions for authors
  1. Have the authors tested whether a classifier trained on original features can use PFF-rec features directly?
  2. Have you evaluated several successive yearly updates?
2 comments
AuthorsOpenReview ↗

Response to Reviewer RH44

Thank you for engaging with our work, and the excellent suggestions for improvement.

Response to weaknesses and questions

[W1] On evaluating bit-proportional unimodal baselines on downstream tasks.

We evaluated bit-proportional unimodal TensoRF-VM fields on PASTIS segmentation. PFFs achieve higher downstream performance on both precomputed embedding products, despite the baseline’s higher embedding reconstruction scores (Table 5; Table 6 in the revised manuscript). The bit-proportional baselines’ improvement on the precomputed embedding product, however, comes at a significant retention % cost on the other products.

ProductRaw mIoU (%)Uniform allocation retention (%)PFF retention (%)Bit-proportional retention (%)
AlphaEarth53.5667.090.186.8
TESSERA61.3769.08988.0
Sentinel-213.3392.0136.384.7
Landsat13.3594.6124.086.2
LST3.3399.2100.994.8
Sentinel-1 RTC5.0091.3159.180.6
PALSAR3.3888.6133.276.4

We plan on expanding this evaluation of the bit-proportional unimodal field and adding additional results to the camera-ready version of our manuscript. Thank you for prompting us to run this evaluation.

[W2] Adding compression baselines (standard codecs) to PFFs

Thank you for this suggestion. We evaluated both lossy and lossless codecs, and have updated Appendix Table 3 in the revised manuscript with our results. At a matched storage budget of 128 MB per tile, PFFs outperform the best tested lossy codec by ~9 points in mean reconstruction score. PFFs also require approximately 650x less storage than the smallest lossless representation tested.

Lossy compression: We evaluated several lossy codecs used in scientific machine learning such as JPEG2000 (https://gdal.org/en/stable/drivers/raster/jp2openjpeg.html), SZ3 (https://github.com/szcompressor/SZ3), and TTHRESH (https://arxiv.org/abs/1806.05952).

Method128 MB2,048 MB8,192 MB
JPEG200086.69 +/- 4.5896.03 +/- 2.2297.12 +/- 0.05
SZ321.30 +/- 5.5663.02 +/- 7.3186.05 +/- 1.92
TTHRESH83.97 +/- 5.0896.50 +/- 0.5897.53 +/- 0.82
PFF95.20 +/- 0.22——

This evaluation was performed on a 10-tile PFFBench subset. We generally find that traditional codecs evaluated are poor compressors at high compression ratios.

Lossless compression: These codecs preserve the source values exactly. We therefore tabulate the minimum file size obtained with these codecs compared to PFFs.

MethodStorage per tile (GB)
DEFLATE84.86 ± 4.65
ZSTD84.28 ± 4.64
LZMA83.45 ± 4.49
LERC, zero error139.14 ± 3.86
JPEG2000, lossless85.32 ± 4.98
PFF (lossy reference)0.128

[Q1] Have the authors tested whether a classifier trained on original features can use PFF-rec features directly?

We test direct classifier reuse on PASTIS crop-type segmentation using AlphaEarth, Tessera, and Sentinel-2 features. For each source, we train a linear classifier on original features and evaluate it on both original features and their PFF reconstructions at the same held-out pixels. In this setup, the linear classifier is neither recalibrated nor refit on the PFF-reconstructed features.

FeaturesRaw classifier on originalRaw classifier on PFF-recPFF-trained classifier
AlphaEarth53.5646.1948.20
Tessera61.3752.3154.62
Sentinel-213.3313.2318.17

We observe a ~4% drop in performance for the precomputed embedding products when a classifier trained on the original embeddings is tested on PFF-reconstructed features of the PASTIS segmentation task (compared to PFF-reconstructed features). The drop is higher for Sentinel-2 (27.2%), but the reconstructed features still recover the raw classifier’s performance.

[Q2] On several successive yearly updates

We evaluated successive yearly updates on ten PFFBench tiles. After pretraining without 2017, 2024 and 2025, we adapted each PFF to 2024, then 2025, and finally to an earlier historical year (2017). The table below reports AlphaEarth reconstruction after all three updates.

Added yearAdapted AlphaEarth skillJoint-pretraining AlphaEarth skill
2024: first forward update79.5 +/- 3.0688.1 +/- 2.91
2025: second forward update79.4 +/- 4.0788.1 +/- 3.30
2017: backward update82.0 +/- 5.6390.1 +/- 4.40

Across all 3 successive updates, newly adapted timesteps track the pre-trained reconstruction score. The second consecutive temporal update performs comparably to the first. By construction (Figure 3), these successive updates do not change previously added timesteps’ reconstruction scores.

The response has addressed my concerns. Thank you for your efforts, and I will keep my acceptance score.

Reviewer_e4CdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a method designed to store representations of geographic locations across space and time. The motivation is convincing: Earth observation is a field where the amount of available data is growing rapidly, with new satellites being launched every year. However, making effective use of all this data is far from trivial.

While the amount of data is enormous, reaching the petabyte scale, and the structure of the data can vary significantly between products (some products have different GSDs, while others have very different physical interpretations, such as SAR versus optical observations), there is an important amount of redundancy between products, since they all capture information about the same Earth. In addition, spatio-temporal sampling is far from uniform, with some geographic regions much more extensively observed than others.

The paper proposes a natural and effective way to aggregate these heterogeneous products into a compact, shared representation. The method is interesting because of its computational efficiency and its ability to progressively incorporate new timesteps with very few additional parameters, while preserving previously learned information. It also provides a mechanism for adding new products through lightweight decoders, although the generality of this approach remains less clear.

The experimental evaluation is comprehensive, covering reconstruction quality, downstream tasks such as semantic segmentation and change detection, continual learning, and retrieval efficiency. The experiments demonstrate the practical relevance of the approach, while the ablations provide useful insights into its behavior and limitations.

Overall, while the architectural novelty is somewhat limited, the paper addresses a relevant problem with a simple and effective solution, supported by a convincing experimental evaluation.

Critical strengths
  1. The paper is well motivated and easy to read. The amount of available EO data, as well as its diversity, continues to increase over time. However, dealing with this diversity is challenging, both computationally, as datasets can reach petabyte scale, and methodologically, as integrating products with different structures (varying image sizes, GSDs, numbers of channels, etc.) is far from trivial. Pretrained embeddings also represent another type of data, different from traditional satellite observations, which the proposed method successfully integrates.
  2. Authors evaluate the usefulness of the resulting representations on a convincing set of downstream applications.
  3. The proposed approach is a simple but effective solution. It builds on existing methods, namely TensoRF and K-Planes. While the architectural novelty is limited, the approach is particularly well suited to EO:
  • Regional fields allow the representation to scale geographically.
  • Factorized volumes exploit spatial and temporal structure while reducing storage requirements.
  • Shared latent features exploit redundancy across EO products, improving compression efficiency.
  • Product-specific decoders accommodate different feature spaces and physical quantities.
  1. The continual-learning aspect is interesting. Adding new timesteps is efficient, requiring the optimization of only about 0.2% of the existing model's parameters while preserving previous predictions. Continual learning seems particularly relevant in EO, where recent models such as AlphaEarth and TESSERA are trained on massive amounts of data but do not necessarily provide mechanisms to update their representations with new observations of a given location. In this context, the ability to efficiently incorporate new temporal information without retraining the entire representation is appealing. To me, this is one of the most interesting research directions opened up by the paper.
Critical weaknesses
  1. The continual-learning aspect is interesting, but the ability to incorporate genuinely new products is not fully demonstrated. When adding new timesteps, the authors extend the representation with additional temporal parameters, which are then trained while keeping the existing representation frozen. However, when adding a new product, the entire shared representation is frozen, and only a product-specific decoder is trained to reconstruct the new product from the information already encoded in the field. While this strategy effectively avoids forgetting, I find the continual-learning claim somewhat limited in this setting, as the representation itself cannot incorporate new information from the additional product. This could be particularly restrictive when introducing modalities with substantially different physical properties. For instance, consider a PFF initially trained only on Sentinel-2 multispectral observations. If hyperspectral observations are subsequently introduced, they may contain spectral information that was not captured by the original data. Since the shared representation remains frozen, the new decoder cannot recover information that is absent from it.
  2. The Galileo experiment is also interesting, as the newly trained decoder achieves downstream performance comparable to the original Galileo embeddings (42.2% versus 42.1% mIoU). However, Galileo is another pretrained EO embedding product, and the original PFF was already trained on 14 heterogeneous products, including AlphaEarth and TESSERA embeddings. Therefore, the new product may contain substantial information already represented in the shared field.
  3. The fine-tuning experiment shows that updating the shared representation can improve new-product reconstruction but causes significant forgetting of previously learned products. This highlights a trade-off between incorporating new information and preserving existing knowledge that is not fully addressed by the proposed approach.
  4. The compression evaluation could be better contextualized. While the authors report impressive compression ratios, these are measured relative to uncompressed source data. However, some of the products considered, such as AlphaEarth and TESSERA, already provide compact, quantized embeddings.
Questions for authors
  1. What happens when introducing a genuinely different sensing modality into a representation initially trained on a restricted set of products? This would help clarify the generality of the proposed new-product adaptation strategy.
  2. How do the reported the storage requirements compare to those of the original products in their actual distributed formats? This would help clarify the practical storage savings achieved by PFF beyond existing compression and quantization strategies.
3 comments
AuthorsOpenReview ↗

Response to Reviewer e4Cd (1/2)

Thank you for your detailed review and accurate summary of our work.

Response to weaknesses and questions

[W1, Q1] On PFFs’ ability to learn modalities with substantially different physical properties/ introducing a genuinely different sensing modality.

We agree with the highlighted limitation that under our new-product continual learning setting, freezing the shared field can limit adaptation when the new product requires new structure (physical or otherwise) that the shared field does not capture during pretraining. As described in Section 2.3 (l. 267–269), new-product adaptation trains a decoder on the existing shared field. Our work studies whether the new product can reuse structure captured by this field.

To examine this limitation, we train PFFs with either purely optical/multispectral input (Sentinel-2, Landsat), or only SAR input (Sentinel-1, PALSAR). Once trained, we attempt to adapt an optical-only PFF to SAR, and a SAR-only PFF to optical modalities. Reconstruction (R^2) is substantially lower than when the target modality is included during pretraining.

Optical PFF -> SAR adaptation

TargetOptical -> SARSAR-only pretrainingOptical + SAR pretraining
Sentinel-10.597 ± 0.1290.862 ± 0.0600.846 ± 0.064
PALSAR0.620 ± 0.1220.876 ± 0.0400.847 ± 0.053

SAR PFF -> Optical adaptation

TargetSAR -> OpticalOptical-only pretrainingOptical + SAR pretraining
Sentinel-20.520 ± 0.1340.927 ± 0.0330.901 ± 0.042
Landsat0.609 ± 0.1320.955 ± 0.0210.938 ± 0.028

However, adaptation improves when we exclude SAR or optical products from the full stack while retaining the other products, including embeddings and maps. In the table below, all raw products of the target product type (either optical or SAR) are excluded during pretraining.

TargetAll except SAR -> SARFull 14-product PFF
Sentinel-10.739± 0.0900.828 ± 0.066
PALSAR0.732 ± 0.0990.840 ± 0.050
TargetAll except optical -> opticalFull 14-product PFF
Sentinel-20.790 ± 0.0810.883 ± 0.047
Landsat0.851 ± 0.0640.926 ± 0.032

Product adaptation does not necessarily require pretraining on a product with the same physical properties as the target. A shared representation pretrained on diverse EO products can capture enough relevant structure to support adaptation, even when the target sensing modality is absent as a raw product (note that AlphaEarth, TESSERA and some map products are themselves derived partly from optical and SAR inputs, so the held-out modality is not entirely absent). We have added this discussion to our manuscript as Appendix Table 10, referenced in both Section 4 and (limitations) Section 6.

[W2] The new product may contain substantial information already represented in the shared field

We highlight PFFs’ ability to exploit the inherent redundancy in several EO products as a core strength of our method. In lines 53-79, we detail this inherent redundancy which is particularly strong in precomputed embedding products. Galileo, AlphaEarth, and TESSERA are learned representations of highly overlapping locations and input modalities, yet are stored and served independently. This overlap allows us to support an additional product through a small decoder, without retraining the existing field or changing the existing representation.

[W3] On addressing tradeoffs between incorporating new information and preserving existing knowledge (finetuning vs PFF adaptation)

Thank you for the insightful question. Indeed, adapting a pre-trained PFF to a new timestep or a new product requires carefully balancing learning versus forgetting. In our experiments, we find our adaptation approach to provide a strong balance between both objectives, especially since finetuning a PFF on new data leads to catastrophic forgetting – where the existing representation can no longer be reliably used. We address this tradeoff in the caption of Figure 6 (l. 285-291), and Section 4. For Galileo, for example, freezing also achieves higher new-product downstream performance than fine-tuning while preserving existing products (Figure 6b). We believe further optimization techniques to balance between the two objectives when adding a new timestep or product (such as methods derived from [1]) would be a valuable direction for future work. We have added this discussion to section 6 in our revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer e4Cd (2/2)

[W4, Q2] Reporting PFF’s compression ratio and storage requirements compared

Our compression ratio (approximately 1800x) uses the uncompressed 14-product stack, with AlphaEarth and TESSERA already stored as quantized int8 (Appendix Table 2). We use this convention since different products use different compressors for their released products. For example, AlphaEarth COGs are shipped with ZSTD on the quantized int8 embeddings, TESSERA uses the LZW codec, and several other products use DEFLATE (from Microsoft Planetary Computer). However, to fully utilize the embeddings, these GeoTIFFs must be decompressed and dequantized into analysis-ready products (in float32). The table below shows the raw and compressed size of the embedding products, along with PFFs’ compression ratio calculated relative to these previously used methods.

RepresentationSize per tile (GB)PFF smaller by
Raw float32, all 14 products (analysis-ready)534.454175x
Raw products, default shipped precision.227.711779x
DEFLATE84.86 ± 4.65663 ± 36x
ZSTD84.28 ± 4.64658 ± 36x
LZMA83.45 ± 4.49652 ± 35x
LERC, zero error139.14 ± 3.861087 ± 30x
JPEG2000, lossless85.32 ± 4.98667 ± 39x
PFF0.1281x

PFFs require approximately 650× less storage than the smallest lossless codec tested. Note that these lossless codecs take between 4-7 hours on a 16 vCPU node. PFFs generally take 2 MI250x GPU hours to train, and can be queried on demand (no need to fully decompress the stack to fetch only a few/sparse features). We thank you for prompting us to make these comparisons and have added this as Appendix Table 3 and Table 9 along with a reference to it in Section 4. We’ve also modified the abstract phrase “uncompressed source data” into “source data at shipped precisions.”

[1] Li, Zhizhong, and Derek Hoiem. "Learning without forgetting." Proceedings of the European Conference on Computer Vision (ECCV) 2016.

I'm happy with the author's responses. The first set of new results more clearly delimit the limitations of the method, while the results on compression really help highlighting the impressive compression capabilities of the implicit model.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response by the Authors to all Reviewers

We thank all reviewers for their thoughtful and constructive feedback on our work.

We are encouraged to read that the paper is “well motivated and easy to read” (R-e4Cd) and that our evaluation is “comprehensive” (R-e4Cd, R-RH44), with “strong empirical validation at a meaningful scale” (R-Y4CP) that demonstrates the “practical relevance” of our approach (R-e4Cd, R-Y4CP). While we build on existing factorized feature volumes like TensoRF and K-Planes, we are happy to hear that our approach is “particularly well suited to EO” (R-e4Cd), and that our continual-learning approach is “a smart design for adding new years without changing previous outputs” (R-RH44) as well as “one of the most interesting research directions opened up by the paper” (R-e4Cd).

We have responded to each reviewer’s comments individually, and we are confident that we can address the remaining questions and concerns during the rest of the discussion period. Based on the feedback, we have uploaded a revised manuscript with changes marked in green text. We are happy to answer any remaining questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.