{"id":"d09ab58b-e376-474a-85c6-a91f8992d248","arxiv_id":"2506.19656","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Encoding multichannel Earth system data with standard video codecs (x265, VP9) yields 8 to 250 times smaller files at 40 to 66 dB PSNR, with no measured loss on two downstream deep learning tasks.","lead":"This paper presents xarrayvideo, a Python library that compresses large Earth observation datasets by encoding them as standard videos, shrinking files by 8 to 250 times while keeping image quality high. Readers may care because terabyte-scale satellite and climate datasets become much cheaper to store, download, and reuse, with benchmark evidence that deep learning models perform the same on the compressed data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DeepExtremeCubes downstream equivalence test is confounded: Table IV compares original data against a version that also drops band B07, forward-fills missing values, and adds an invalid mask, so 'no performance loss' is not attributable to the codec alone.","rationale":"The paper is a well-executed engineering contribution: the PSNR/bpppb benchmarks across four real datasets are credible, the library is released, and the DynamicEarthNet downstream test is clean and supports the utility claim for that dataset. The most load-bearing soft spot is not the universal 'any downstream application' phrasing by itself (no finite evaluation can prove that), but the fact that the DeepExtremeCubes downstream equivalence, which is the main evidence for the larger data release, is confounded by simultaneously changing the band set and missing-data policy. The reader's weakest_assumption identified both this confound and the missing quality threshold; I agree that both matter, but the confound is the more specific and testable issue. If the codec-only ablation shows identical performance, the paper's central claim is substantially supported and only the overgeneralized Discussion language would need softening. If it does not, then the 56 dB equivalence is an artifact of preprocessing, and the paper must either qualify the claim or re-run the release. This does not warrant rejection, because the compression ratios, PSNR results, and one clean downstream experiment stand; it does warrant a conditional verdict requiring the ablation and explicit error bars before the interchangeability claim is accepted.","tokens_in":20621,"tokens_out":4299,"duration_ms":47659,"concrete_test":"Run a codec-only ablation on the DeepExtremeCubes downstream benchmark. Encode the same ten cubes with xarrayvideo/libx265 12-bit at the 'Very high' 56 dB setting while preserving all seven bands (e.g., encode B07 in a second video) and keep missing-value handling identical between the original and compressed branches: either evaluate both with the original validity mask and no forward-fill, or apply the same forward-fill to both. Then apply the pretrained model from [42] and report test R2, L1, and MSE with per-cube standard errors over the ten cubes. If the compressed-with-all-bands/no-preprocessing-change scores match the original within one standard error, the codec is exonerated and the confound is resolved; if they reproduce the difference only when B07 is dropped and values are forward-filled, the paper's equivalence claim is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section III-B states that the released DeepExtremeCubes version 'channel B07 ... is not included from the original dataset to improve the compression ratio ... missing data has been filled forward ... and an additional invalid mask has been added since compressed data cannot hold missing values.' Table IV then reports that the Video 56dB (270 Gb) version achieves essentially identical downstream performance to the original (test R2 0.905 vs 0.906, L1 0.037 vs 0.036). Because the compared artifacts differ in three ways simultaneously (lossy x265 compression, removal of a spectral band, and a different missing-data policy), the same-score result cannot be assigned to the codec. Forward-filling and mask addition change the actual inputs to the next-step reflectance model, and dropping B07, while acknowledged as redundant with B06/B8A, is still an information loss that is not part of the video codec. Thus the abstract claim that 'no performance loss is observed' for DeepExtremeCubes, and the Discussion's stronger 'completely interchangeable' claim, rest on only one unconfounded downstream experiment (DynamicEarthNet). This matters because the 270 GB redistribution is advertised as 'without compromising quality,' and Table IV is the evidence for that on the larger release. The confound is acknowledged in the manuscript but not controlled. The same table also shows a small but measurable drop at 47 dB, so the 'sufficiently high quality' threshold is not pinned down; however, the primary issue is that the 56 dB equivalence is not a codec-only comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces xarrayvideo, an open-source Python library that compresses multichannel spatiotemporal Earth system datasets by encoding them as standard videos via ffmpeg. The method maps xarray variables to video files, with options for codec (x265, VP9, FFV1, JPEG2000), bit depth, PCA preprocessing, and quality presets. The authors benchmark the library on four datasets: DynamicEarthNet (Planet imagery), DeepExtremeCubes (Sentinel-2), ERA5 (reanalysis), and SimpleS2, reporting rate-distortion curves in terms of bits per pixel per band (bpppb), PSNR, SSIM, and spectral angle. They also redistribute compressed versions of DynamicEarthNet (525 GB to 8.5 GB) and DeepExtremeCubes (2.3 TB to 270 GB) through HuggingFace in the TACO format. Downstream evaluations apply pretrained models from the original dataset publications without modification: landcover semantic segmentation for DynamicEarthNet and next-step reflectance prediction for DeepExtremeCubes. The paper reports no performance loss for these tasks at the high-quality compression levels and argues that lossy compressed data at sufficiently high quality is completely interchangeable with original data for any downstream application.","tokens_in":20734,"tokens_out":5299,"duration_ms":54663,"significance":"If the central claims hold, the contribution is practically significant for the Earth observation and machine learning communities. xarrayvideo provides a simple, standardized, and accessible way to compress large spatiotemporal datasets using mature video codecs, with fully disclosed encoder configurations (Table A.1) and reversible mapping to xarray structure. The released compressed datasets are substantial real artifacts: DynamicEarthNet at 62x compression and 60.15 dB PSNR, and DeepExtremeCubes at 8.5x compression and about 56 dB PSNR. The downstream tests use pretrained models from prior publications without modification, which is an appropriate external benchmark. The paper is also transparent about limitations (full-cube memory loading, decompression overhead) and about the specific codec configurations used. However, the strength of the 'no performance loss' and 'completely interchangeable' claims is not yet matched by the evidence, which is based on only two downstream tasks and, for DeepExtremeCubes, on a comparison that is confounded by simultaneous preprocessing changes.","major_comments":[{"comment":"The DeepExtremeCubes downstream comparison is confounded. The 'Video 56dB' version differs from the original in at least three ways simultaneously: lossy x265 compression, omission of band B07, and a different missing-data policy (forward-filling plus an added invalid mask), as acknowledged in Section III-B. The near-identical test metrics (R2 0.905 vs 0.906, L1 0.037 vs 0.036) therefore cannot be attributed to the video codec alone. Since this experiment is the only evidence for the abstract's claim that 'No performance loss is observed' for DeepExtremeCubes and for the advertised redistribution 'without compromising quality,' the authors should provide an ablation that isolates the codec effect, for example by compressing the full 7-band data with unchanged missing-data handling, or by comparing codec-only versus codec-plus-preprocessing variants.","section":"Section III-B, Table IV"},{"comment":"The claim that a dataset 'compressed lossily at sufficiently high quality is completely interchangeable with the original data for any downstream application' is substantially overgeneralized relative to the evidence. Only two downstream tasks are tested, one of which is confounded as noted above, and the study does not characterize the quality threshold below which performance degrades. Table IV itself shows a measurable drop at 47 dB PSNR (test R2 from 0.906 to 0.899, L1 from 0.036 to 0.040), so the 'sufficiently high quality' regime is not pinned down. The authors should soften the interchangeability claim to the tested tasks and quality levels, or add a rate-performance study that maps downstream accuracy as a function of compression level.","section":"Section I and Section IV"},{"comment":"The PSNR definition used here is non-standard and complicates comparisons. PSNR is computed with respect to the maximum intensity present in the original image sequence for each channel, rather than a fixed bit-depth maximum. This can inflate headline numbers relative to conventional PSNR and makes the comparisons with Das et al. and Cra5 in the Discussion questionable, since those results are based on different error metrics and PSNR conventions. The authors should report a standard fixed-range PSNR as well, or at least provide a conversion, so that the reported 55-65 dB figures can be interpreted by readers outside the paper.","section":"Section II-C and Section IV"}],"minor_comments":[{"comment":"Row 'Original (3,2 Tb)' appears to be a typo; the text and elsewhere state DeepExtremeCubes is 2.3 TB, so '2,3 Tb' is likely intended.","section":"Table IV"},{"comment":"The word 'respectivley' should be 'respectively' in the sentence on median compression and decompression times.","section":"Section III-A"},{"comment":"The comment '(# input tyxc coords)' contains a typo; the coordinate tuple is ('time', 'y', 'x'), so 'tyx' would be the correct abbreviation.","section":"Listing 1"},{"comment":"The phrase 'channel B07 ... is not included from the original dataset' could be rephrased as 'channel B07 is omitted from the compressed version of the original dataset' for clarity.","section":"Section III-B"},{"comment":"The paper uses both 'Principal Components (PC) Analysis' and 'PCA transformation'; please standardize the terminology (e.g., PCA).","section":"Section I"}],"recommendation":"major_revision","confidential_remarks":"The core engineering contribution, xarrayvideo, is solid and the benchmarks are useful. The main issue is that the downstream equivalence and interchangeability claims are broader than what the evidence supports, particularly for DeepExtremeCubes where a confound exists. I would recommend requesting an ablation or reanalysis to isolate the codec effect, and a more measured phrasing of the claims, before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the artifact is real: xarrayvideo is a simple, well-engineered library that encodes multichannel spatiotemporal xarray data as standard video codecs via ffmpeg, and the authors release two large compressed datasets (DeepExtremeCubes at 270 GB, DynamicEarthNet at 8.5 GB) with fully disclosed encoder configurations. Second, the headline claim that compressed data are 'completely interchangeable' with originals is not supported by the evidence presented; only one downstream test is unconfounded, and the paper's own 47 dB results show measurable degradation.\n\nThe novelty is modest but legitimate. The core move, running H.265 over image sequences, was already applied to hyperspectral data with bands-as-time by Santos et al. What is new is the recipe for true spatiotemporal EO data, with careful per-channel-max PSNR, SSIM, spectral angle, bpppb, and full configuration tables in the appendix. The rate-distortion curves on four real datasets are useful, and the finding that x265 and vp9 dominate JPEG2000 at low bpppb is a practical data point. The DynamicEarthNet downstream test is clean: three pretrained models evaluated on two compressed versions give essentially identical mIoU and accuracy to the original. The PCA ablation is honest, showing PCA did not help. The code, datasets, and configs are public, which strengthens the contribution.\n\nThe soft spot is the DeepExtremeCubes equivalence test. As Section III-B discloses, the compressed 'Video 56 dB' version also drops band B07, forward-fills missing values, and adds an invalid mask. Table IV's same-score result therefore cannot be attributed to the codec alone. That leaves only one unconfounded downstream experiment, which is thin support for the Discussion's 'completely interchangeable for any downstream application.' The small but real drop at 47 dB (test R2 from 0.906 to 0.899, L1 from 0.036 to 0.040) shows a quality threshold exists, but the paper never characterizes it. Minor issues: the DeepExtremeCubes size is given as 2.3 Tb in the abstract and Figure 1 but 3.2 Tb in Table IV, and no error bars are reported for the small cube samples (five to ten cubes).\n\nThe paper deserves serious peer review. It is a practical engineering contribution with reproducible measurements and public artifacts, and the right referee can push for toned-down claims and a controlled codec-only comparison on DeepExtremeCubes. I would accept with revisions, not desk-reject.","headline":"Practical library plus solid rate-distortion measurements, but the 'completely interchangeable' claim is overreaching and the DeepExtremeCubes downstream test is confounded by preprocessing changes.","tokens_in":21515,"tokens_out":2093,"would_cite":true,"duration_ms":21343,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims standard video codecs compress Earth system data up to 250x without degrading downstream machine learning performance.","keywords":["video-compression","Earth-system-data","xarrayvideo","H.265","multispectral-imagery","ERA5","lossy-compression","downstream-tasks"],"falsifier":"Run a third downstream model, such as a regression of a biophysical variable or a rare-event detection model, on the released compressed DynamicEarthNet and DeepExtremeCubes versions and on the original data, and check whether performance differs beyond noise.","tokens_in":20211,"feed_emoji":"🗜️","tokens_out":8501,"duration_ms":80278,"temperature":0.7,"pith_summary":"The paper tries to establish that the spatial, temporal, and spectral redundancies of large Earth system datasets are close enough to ordinary video that standard, off-the-shelf video codecs can compress them dramatically without compromising scientific use. It presents xarrayvideo, a library that writes xarray data cubes as standard video files (H.265 or VP9 through ffmpeg) and reads them back, reporting compression ratios up to 250x and PSNRs between roughly 40 and 66 dB on four real datasets. The load-bearing conclusion is interchangeability: a dataset compressed lossily at sufficiently high quality should behave identically to the original in any downstream task. As evidence, pretrained models for landcover segmentation and next-step reflectance forecasting score essentially identically on original and compressed data, and two compressed datasets (DeepExtremeCubes at 270 GB from 2.3 TB, DynamicEarthNet at 8.5 GB from 525 GB) are publicly released. If this holds, terabyte-scale Earth observation becomes distributable and trainable at a fraction of current storage and bandwidth costs.","feed_headline":"Video codecs shrink Earth datasets up to 250x without hurting AI models","feed_subtitle":"Two terabyte-scale satellite datasets are reissued at 8-62x smaller size with identical downstream model scores.","key_machinery":"The machinery is the xarray2video mapping: a spatiotemporal DataArray with axes (bands, time, y, x) is reshaped so that frames are (time, y, x) slices; bands are packed three per video (an extra band is triplicated to fill the third channel, or PCA reduces the band count); each video is encoded by ffmpeg with libx265 or VP9 at 12-bit depth, with JPEG2000 via GDAL as the image-codec baseline; lossless auxiliary masks use ffv1; and all non-video metadata goes into a small NetCDF file (x.nc). The same mapping rules choose codec, bit depth, quality preset, and whether PCA is applied. This turns decades of video-codec engineering for motion compensation, transform coding, and rate control into a generic compressor for Earth system data, and it is what allows compressing clouds, misregistration, missing data, and smooth reanalysis fields with the same tool.","core_discovery":"On its own terms, the central claim is that the time axis of Earth system data carries exactly the kind of redundancy that video codecs are built to remove, so a standard encoder can serve as a general, off-the-shelf compressor for multispectral spatiotemporal cubes. xarrayvideo encodes the (time, y, x) array as a video, packs bands into three-channel videos, and reports PSNRs of 55.86, 40.60, 46.58, and 43.23 dB at 0.1 bpppb and 65.91, 54.28, 62.90, and 55.04 dB at 1 bpppb on DynamicEarthNet, DeepExtremeCubes, ERA5, and SimpleS2. It then supplies the key evidence for interchangeability: pretrained models for landcover segmentation and next-step reflectance prediction, applied without modification, score essentially identically on the original and compressed data at high quality (60 dB for DynamicEarthNet, 56 dB for DeepExtremeCubes), and two such compressed datasets are publicly released at 270 GB and 8.5 GB. The paper states the strong version of the conclusion: no performance loss is observed when compressed versions are used in the respective deep learning downstream tasks.","pith_inferences":["If the interchangeability claim is accepted, benchmark and foundation-model corpora could ship one high-quality lossy master plus checksums, with the originals archived, cutting distribution cost by one to two orders of magnitude; this extension is not developed in the paper.","The paper's two downstream tests leave open whether codec artifacts such as blocking, ringing, temporal flicker, and chroma subsampling are inert for other tasks like rare-event detection, biophysical parameter retrieval, or change tracking; testing the released cubes on a broader task battery would settle that.","Because the released DeepExtremeCubes version also drops band B07, forward-fills missing values, and adds an invalid mask, part of the reported performance equivalence may come from those preprocessing changes rather than from the codec itself; an ablation on a preprocessed-but-lossless baseline would separate the two effects.","The finding that PCA before compression was neutral or detrimental suggests that the simple route of splitting bands into three-channel videos is the more robust design choice for more than three bands, at least with the tested x265 settings."],"forward_implications":["DynamicEarthNet can be distributed at 62x smaller size (8.5 GB at 60.15 dB) and DeepExtremeCubes at 8.5x smaller (270 GB at 56.65 dB) while reproducing the original datasets' scores on the tested segmentation and reflectance-forecasting models.","At more aggressive settings the same pipeline still yields usable data: a 2.1 GB DynamicEarthNet at 54 dB (about 249x smaller) matches the original metrics, while DeepExtremeCubes at 47 dB and 85 GB costs only a small drop in test R2 from 0.906 to 0.899.","Because xarrayvideo builds on ffmpeg and standard codecs, any xarray-shaped dataset can be compressed without neural-model training, specialized hardware, or per-dataset tuning.","Lossless layers such as cloud masks, scene classification, and validity masks compress about 34x with the ffv1 codec, so mixed lossy/lossless packaging is practical.","Codec choice is data-dependent: x265 and VP9 beat JPEG2000 for natural-looking Earth observation scenes at low bitrates, while JPEG2000 wins on smooth reanalysis fields at high bitrates."],"supporting_citations":[{"why":"Defines the TACO cloud-optimized format used to redistribute the two compressed datasets.","marker":"[1]"},{"why":"Supplies the DynamicEarthNet dataset and the pretrained segmentation models used for downstream evaluation.","marker":"[18]"},{"why":"Supplies the DeepExtremeCubes dataset used for compression benchmarks and release.","marker":"[20]"},{"why":"Defines the ERA5 reanalysis product, whose subset is compressed in the weather benchmark.","marker":"[21]"},{"why":"Provides Cra5, a neural-compressed ERA5 corpus, as the main comparison for extreme compression and downstream skill retention.","marker":"[28]"},{"why":"The only prior work the authors find on compressing multichannel video, providing the baseline PSNR at 1 bpppb.","marker":"[30]"},{"why":"Closest prior idea: applying a video codec to hyperspectral images by treating bands as frames, motivating the present approach.","marker":"[35]"},{"why":"Supplies the WeatherBench2 ERA5 subset used for the reanalysis compression benchmark.","marker":"[36]"},{"why":"Supplies the pretrained next-step reflectance model used to evaluate the DeepExtremeCubes downstream task.","marker":"[42]"}],"fun_headline_variants":["Video codecs compress Earth data 250x, AI models unaffected","Watch: video compression slashes Earth dataset sizes 250x","Earth data as video: 250x compression, same AI performance","xarrayvideo: turn Earth cubes into videos for 250x smaller files","AI-ready Earth data in 250x less space via video encoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument collapses if two pretrained models are not enough to certify that codec artifacts are inert for all downstream Earth system applications, since the paper never identifies the quality threshold at which losses begin.","fun_headline_variants_meta":{"raw":{"variants":["Video codecs compress Earth data 250x, AI models unaffected","Watch: video compression slashes Earth dataset sizes 250x","Earth data as video: 250x compression, same AI performance","xarrayvideo: turn Earth cubes into videos for 250x smaller files","AI-ready Earth data in 250x less space via video encoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":3031,"prompt_tokens":1196,"completion_tokens":1835,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":812,"completion_tokens_details":{"reasoning_tokens":1743}},"tokens_in":812,"tokens_out":1835,"duration_ms":10945,"temperature":1.0,"reasoning_tokens":1743,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:31:11.678626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a third downstream model, such as a regression of a biophysical variable or a rare-event detection model, on the released compressed DynamicEarthNet and DeepExtremeCubes versions and on the original data, and check whether performance differs beyond noise.","supporting_citations":[{"cited_title":"Taco: Transparent access to cloud-optimized spatio-temporal datasets,","cited_arxiv_id":null,"evidence_quote":"Defines the TACO cloud-optimized format used to redistribute the two compressed datasets."},{"cited_title":"Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the DynamicEarthNet dataset and the pretrained segmentation models used for downstream evaluation."},{"cited_title":"DeepExtremeCubes: Integrating Earth system spatio-temporal data for impact assessment of climate extremes","cited_arxiv_id":"2406.18179","evidence_quote":"Supplies the DeepExtremeCubes dataset used for compression benchmarks and release."},{"cited_title":"Hyperspectral image, video compression using sparse tucker tensor decomposition,","cited_arxiv_id":null,"evidence_quote":"The only prior work the authors find on compressing multichannel video, providing the baseline PSNR at 1 bpppb."},{"cited_title":"Performance evaluation of the h. 264/avc video coding standard for lossy hyperspectral image compression,","cited_arxiv_id":null,"evidence_quote":"Closest prior idea: applying a video codec to hyperspectral images by treating bands as frames, motivating the present approach."},{"cited_title":"Weather- bench 2: A benchmark for the next generation of data-driven global weather models,","cited_arxiv_id":null,"evidence_quote":"Supplies the WeatherBench2 ERA5 subset used for the reanalysis compression benchmark."},{"cited_title":"Explainable Earth Surface Forecasting under Extreme Events","cited_arxiv_id":"2410.01770","evidence_quote":"Supplies the pretrained next-step reflectance model used to evaluate the DeepExtremeCubes downstream task."}],"review_version":2}