Spiral Jetty imagery leads Great Salt Lake's decline by 3 years

Robert Smithson's 1970 land artwork Spiral Jetty sits in the north arm of Utah's Great Salt Lake, where severe lake decline has repeatedly submerged and exposed it. Researchers assembled 1,744 co-registered satellite image chips from Landsat 4 through 9 and Sentinel-2, covering every year and calendar month from 1984 to 2025, and computed a 14-feature complexity signature for each one: Shannon entropy, multiscale permutation entropy, fractal dimension, lacunarity, gray-level co-occurrence texture, intensity statistics, and features from an ImageNet-pretrained ResNet50 network. They compared these measurements against a 42-year monthly climate and hydrology dataset drawn from NASA GISTEMP, the USGS National Water Information System, Open-Meteo, and the Global Carbon Budget.

A bootstrap analysis found that Shannon entropy, the metric behind an earlier small-sample claim of a positive correlation with global temperature, is a weak proxy and does not support that claim. Two other measures held up better: coarse-scale permutation entropy and mean image intensity tracked lake elevation strongly, with Spearman correlations of 0.85 to 0.88 and 95 percent confidence intervals that excluded zero. Separately, the third principal component of the ResNet50 embeddings emerged, without any supervision, as what the authors call an AI climate axis: it correlated 0.86 with cumulative CO2 and -0.83 with lake elevation.

The strongest finding concerns timing. Image complexity leads lake stage by about three years, with a Pearson correlation of 0.58 at that lag and a 95 percent confidence interval of 0.40 to 0.73. An STL decomposition of the complexity signal shows a non-monotonic trend: it rises from 1984 to 2015, then declines sharply as the lake approaches record-low elevations. Partial correlations that control for calendar month and sensor confirm the pattern survives after removing seasonal and instrument effects. The authors say the results refine the familiar art-as-thermometer metaphor for Spiral Jetty into a more specific one: art as a leading indicator of hydrological state. They have released the dataset, feature pipeline, and analysis code as a public benchmark.

Key facts

  • 1,744 co-registered Landsat 4-9 and Sentinel-2 image chips of Spiral Jetty, spanning 1984 to 2025, feed a 14-feature complexity signature compared against a 42-year climate and hydrology panel.
  • Coarse-scale permutation entropy and mean intensity correlate with lake elevation at 0.85 to 0.88 (Spearman); Shannon entropy, used in an earlier small-sample study, does not hold up under bootstrap testing.
  • An unsupervised principal component of ResNet50 image embeddings correlates 0.86 with cumulative CO2 and -0.83 with lake elevation.
  • Image complexity leads lake stage by about three years, with a Pearson correlation of 0.58 and a 95 percent confidence interval of 0.40 to 0.73 at that lag.
  • The dataset, feature pipeline, and analysis code are released as a public benchmark.

Why it matters

The core claim is that satellite photographs of a specific artwork carry a genuine climate signal, not just a curiosity. An unsupervised principal component of pretrained ResNet50 image features, extracted with no climate labels involved, correlates 0.86 with cumulative CO2 emissions and -0.83 with lake elevation. That reframes Spiral Jetty from a passive thermometer recording ambient change into what the authors call a leading indicator: its visual complexity moves about three years ahead of the lake's own water level, with a Pearson correlation of 0.58 over that lag.

Who it affects

The direct audience is researchers working across remote sensing, hydrology and cultural heritage: anyone studying the Great Salt Lake's record decline, anyone building complexity or texture metrics from Landsat and Sentinel-2 archives, and anyone interested in using ResNet50 or similar pretrained vision models as off-the-shelf feature extractors for environmental time series.

How to use it

The authors release the full 1,744-image dataset, the 14-feature complexity pipeline, and the analysis code as a public benchmark, which lets other groups rerun the correlations, test the method against other lake systems or land artworks, or extend the feature set. No price, paywall or licence terms are mentioned.

How solid is it

The analysis spans 42 years of monthly data across Landsat 4 through 9 plus Sentinel-2, compared against a climate and hydrology panel drawn from four sources: NASA GISTEMP, USGS NWIS, Open-Meteo, and the Global Carbon Budget. Bootstrap resampling stress-tests the entropy result, which is how the paper catches that Shannon entropy, the basis of an earlier small-sample claim, does not actually hold up. Partial correlations controlling for calendar month and sensor confirm the main results survive those checks. The main limitation is scope: this is a single-site case study of one lake and one artwork.

Risks and caveats

The findings are correlational: the paper reports Spearman and Pearson coefficients, not a proven causal mechanism, and it does not explain why image complexity would lead lake stage by roughly three years. It also gives no numeric lake elevation figures, only correlation coefficients, and does not name the paper's authors, institutions, or a link to the released dataset.