Land Art as a Big-Data Climate Sensor

# Land Art as a Big-Data Climate Sensor: A Multi-Feature Complexity Signature for Robert Smithson’s Spiral Jetty from 1,744 Landsat & Sentinel-2 Scenes (1984–2025)
Source: [https://arxiv.org/html/2609.13182](https://arxiv.org/html/2609.13182)
Alev CinbarcıSean S\. KalaycıoğluDepartment of Aerospace Engineering, Toronto Metropolitan University, Toronto, ON, CanadaDepartment of Mechanical Engineering, York University, Toronto, ON, CanadaDirector, Space, Robotics and AI, Dr\. Robot Inc\., Toronto, ON, CanadaCorrespondence: skalay@torontomu\.ca

###### Abstract

Robert Smithson’s 1970 land artworkSpiral Jetty, anchored at the north arm of Utah’s Great Salt Lake \(GSL\), has been alternately submerged and exposed across a half\-century of catastrophic lake decline \(peak elevation 4,210\.22 ft NGVD29 in 1987; historic low 4,189\.25 ft in 2023; a 20\.96 ft / 6\.39 m drop in 36 years\)\. We treat the work as a geographically fixed remote\-sensing target and analyse1,744 co\-registered Landsat 4–9 and Sentinel\-2 chipsspanningevery year and every calendar month from 1984 to 2025\. A 14\-feature complexity signature combining Shannon entropy, multi\-scale permutation entropy, box\-counting fractal dimension, gliding\-box lacunarity, gray\-level co\-occurrence texture, first\-order intensity statistics, and ImageNet\-pretrained ResNet50 avg\-pool deep features is computed at scale and validated against a 42\-year monthly climate panel built from NASA GISTEMP, USGS NWIS, Open\-Meteo, and the Global Carbon Budget\. Year\-aggregated bootstrap analysis shows \(i\) Shannon entropy alone is a weak proxy, refuting an earlier small\-sample claim of positive correlation with global temperature; \(ii\) coarse\-scale permutation entropy and mean intensity track lake elevation strongly \(Spearmanρ≈\+0\.85\\rho\\approx\+0\.85to\+0\.88\+0\.88, 95 % CIs excluding zero\); \(iii\) the third principal component of the ResNet50 avg\-pool embeddings emerges*without supervision*as an “AI climate axis” withρ=\+0\.86\\rho=\+0\.86to cumulative CO2andρ=−0\.83\\rho=\-0\.83to lake elevation; \(iv\) image complexityleads lake stage by approximately three years\(Pearsonr=\+0\.58r=\+0\.58at lag\+3\+3, 95 % CI\[\+0\.40,\+0\.73\]\[\+0\.40,\+0\.73\]\); and \(v\) STL decomposition reveals the long\-term trend isnon\-monotonic, rising 1984–2015 then collapsing post\-2015 in coincidence with the lake’s historic\-record\-low elevations\. Scene\-level partial correlations controlling for calendar month and sensor identity confirm that the year\-aggregated signal is robust to seasonal and sensor confounds\. These findings refine the popular “art\-as\-thermometer” framing into a defensible “art\-as\-leading\-indicator\-of\-hydrological\-state” reading\. The dataset, feature pipeline, and reproducible analysis code are released as a public benchmark\.

Keywords:remote sensing; image complexity; permutation entropy; fractal dimension; deep features; ResNet50; Great Salt Lake; cultural heritage; climate change; time\-series analysis; unsupervised feature learning; digital twin\.

## 1Introduction

In April 1970 the American artist Robert Smithson rolled 6,650 tons of black basalt and salt\-encrusted earth into the north\-eastern shoulder of Utah’s Great Salt Lake \(GSL\) and builtSpiral Jetty\[[1](https://arxiv.org/html/2609.13182#bib.bib1)\]: a 460 m anti\-clockwise curl of stone extending from Rozel Point into shallow, hyper\-saline water\. Smithson conceived the work in the explicit vocabulary of thermodynamic entropy — a sculpture whose meaning was inseparable from rust, crystal growth, and the disordering of form\. “The work has to do with”, he wrote two years later, “a self\-contained sealed\-off place where everything happens at the same time\. The eye is the eye of a hurricane”\[[1](https://arxiv.org/html/2609.13182#bib.bib1)\]\. Within two years of its completion the lake had risen and submerged the spiral; it remained underwater for almost three decades\.

Over the half\-century since, the GSL has experienced a catastrophic decline\. USGS gauges record the north\-arm elevation falling from a flood\-cycle peak of4,210\.22 ft NGVD29 in 1987 to 4,189\.25 ft in 2023— a 20\.96 ft \(6\.39 m\) drop in 36 years and the lowest reading in the gauge’s 60\-year instrumental record\[[2](https://arxiv.org/html/2609.13182#bib.bib2),[3](https://arxiv.org/html/2609.13182#bib.bib3),[4](https://arxiv.org/html/2609.13182#bib.bib4)\]\. Over the same window the NASA GISTEMP global mean temperature anomaly rose by approximately 1\.0∘C\. Each rise and fall of the lake exposed or drownedSpiral Jettyand left a record in satellite imagery of basalt, salt crust, algal mat, and pinkDunaliella salinabrine\.

Spiral Jettyhas thus become, by accident, an unintended scientific instrument\. Its half\-century history places a single geographically stable artwork at the precise intersection of \(a\) a hyper\-saline endorheic lake collapsing under climate stress, \(b\) the longest continuous record of free, multi\-decadal satellite Earth observation ever assembled \(Landsat since 1972; Sentinel\-2 since 2015\), and \(c\) a salt\-crust regime whose crystallisation–dissolution–recrystallisation cycle integrates climatic history on annual to multi\-year time scales\. Few sites in the world combine this triad\. The question this paper asks is whether the resulting visual record carries a measurable climate signal — and, if so, which features carry which signals, at what statistical strength, and on what time scale\.

### 1\.1A pilot study, its limitations, and what this paper does

A pilot analysis by one of the present authors\[[5](https://arxiv.org/html/2609.13182#bib.bib5)\]computed Shannon entropy on 42 archival satellite images of the artwork 1970–2022 and reported correlations with global temperature, carbon emissions, lake flow, and salinity, advancing the popular hypothesis that the artwork functions as a “strategic thermometer” of climate change\. The pilot, however, suffered from two limitations that we now revisit at scale:

1. 1\.It used asingle complexity feature\(Shannon entropy\) where many are available; image complexity has been studied in remote sensing for half a century through GLCM texture\[[6](https://arxiv.org/html/2609.13182#bib.bib6)\], fractal dimension\[[7](https://arxiv.org/html/2609.13182#bib.bib7),[8](https://arxiv.org/html/2609.13182#bib.bib8)\], gliding\-box lacunarity\[[9](https://arxiv.org/html/2609.13182#bib.bib9),[10](https://arxiv.org/html/2609.13182#bib.bib10)\], multi\-scale entropy\[[11](https://arxiv.org/html/2609.13182#bib.bib11)\], and permutation entropy\[[12](https://arxiv.org/html/2609.13182#bib.bib12)\], each capturing a distinct facet of visual organisation\. Modern deep CNN features add another dimension\[[13](https://arxiv.org/html/2609.13182#bib.bib13),[14](https://arxiv.org/html/2609.13182#bib.bib14),[15](https://arxiv.org/html/2609.13182#bib.bib15)\]\.
2. 2\.Itseffective sample sizeon the climate axis was 13 years, well below the threshold at which bootstrap statistics become reliable, and the climate variables were repeated across multiple images of the same year without proper aggregation\.

In this paper we revisit the question with a dataset two orders of magnitude larger and a 14\-feature complexity signature\. We retrieve all publicly available Landsat 4–9 Collection\-2 Level\-2 and Sentinel\-2 Level\-2A scenes intersecting a 5 km×\\times5 km bounding box centred on Rozel Point for 1984\-01\-01 to 2025\-12\-31 —1,744 co\-registered chips, every year, every calendar month\. We assemble a monthly climate panel for the same window from NASA GISTEMP, the Salt Lake City station record \(Open\-Meteo Historical Reanalysis\), the USGS gauges for both arms of the GSL \(Saltair and Saline\), the Global Carbon Budget, and the pilot study’s salinity series\. We compute a complexity signature combining six classical descriptors with the avg\-pool activations of an ImageNet\-V2 pretrained ResNet50 reduced via principal component analysis\. We then ask which features track which climate variables, at what lag, under what controls, and with what robustness to seasonal and sensor confounds\.

### 1\.2Contributions

- •Dataset\.A public benchmark of 1,744 co\-registered Spiral Jetty satellite chips with per\-scene metadata and the full 14\-feature complexity signature, plus the aligned monthly climate panel\. To our knowledge this is the first openly published multi\-decadal remote\-sensing dataset of a land\-art installation under climate stress\.
- •Methodology\.A multi\-feature complexity\-signature pipeline validated through year\-aggregation, bootstrap confidence intervals, lag analysis, and scene\-level partial correlations controlling for calendar month and sensor identity\. The pipeline generalises to other land\-art sites \(Sun Tunnels, Double Negative, Lightning Field\) and to coastal or permafrost cultural\-heritage sites under climate stress\.
- •Findings\.Five empirical results, summarised in the Abstract, that together \(a\) refute the pilot study’s positive\-temperature claim, \(b\) establish a robust hydro\-climatic signal carried by coarse\-scale permutation entropy and CNN embeddings, \(c\) demonstrate an emergent unsupervised “climate axis” in pretrained vision features, \(d\) identify a∼\\sim3\-year forward lead of visual complexity over lake stage, and \(e\) reveal a non\-monotonic long\-term trend collapsing post\-2015 in coincidence with the GSL’s historic\-record\-low elevations\.

### 1\.3Companion paper

A forthcoming companion paper builds on the dataset and signature released here to construct a climate\-conditioned latent\-diffusion model with a hydrological physics\-coupling layer, forecasting the site under IPCC SSP scenarios\. We deliberately separate the two: the measurement claims of this paper should not depend on any generative model, and any generative model should rest on independently peer\-reviewed measurement claims\.

### 1\.4Paper organisation

Section[2](https://arxiv.org/html/2609.13182#S2)reviews relevant literatures: Smithson’s entropic aesthetics, image complexity in remote sensing, GSL hydrology since 1970, and deep features for Earth observation\. Section[3](https://arxiv.org/html/2609.13182#S3)describes the dataset and climate panel assembly\. Section[4](https://arxiv.org/html/2609.13182#S4)specifies the complexity signature\. Section[5](https://arxiv.org/html/2609.13182#S5)sets up the statistical framework\. Section[6](https://arxiv.org/html/2609.13182#S6)reports results\. Section[7](https://arxiv.org/html/2609.13182#S7)discusses caveats, ethical considerations, and implications for cultural\-heritage monitoring\. Section[8](https://arxiv.org/html/2609.13182#S8)concludes\.

## 2Background

### 2\.1Smithson’s entropic aesthetics

Smithson developed an explicit theory of entropy as an aesthetic principle in*A Tour of the Monuments of Passaic, New Jersey*\[[16](https://arxiv.org/html/2609.13182#bib.bib16)\],*The Spiral Jetty*\[[1](https://arxiv.org/html/2609.13182#bib.bib1)\], and the unfinished*Spiral Hill*\[[17](https://arxiv.org/html/2609.13182#bib.bib17)\]\. For Smithson entropy was not metaphor: it was a measurable tendency that sculpture could be designed to*expose*rather than resist\.Spiral Jettywas sited at the north arm of the GSL precisely because the hyper\-saline brine and the unusualDunaliella salinaalgal bloom dyed the surrounding water pink—visible from low orbit—and because NaCl crystallisation on basalt was, in his term, a “time crystal” performing entropic change at human scale\[[18](https://arxiv.org/html/2609.13182#bib.bib18),[19](https://arxiv.org/html/2609.13182#bib.bib19),[20](https://arxiv.org/html/2609.13182#bib.bib20)\]\.

### 2\.2Image complexity in remote sensing

Shannon entropy\[[21](https://arxiv.org/html/2609.13182#bib.bib21)\]is the simplest of a family of complexity measures applied to satellite imagery since the late 1970s\. Modern practice augments it with:

- •*Texture descriptors*via the gray\-level co\-occurrence matrix \(GLCM\)\[[6](https://arxiv.org/html/2609.13182#bib.bib6)\]: contrast, homogeneity, correlation, energy, entropy\.
- •*Self\-similarity descriptors*via box\-counting fractal dimension\[[7](https://arxiv.org/html/2609.13182#bib.bib7),[8](https://arxiv.org/html/2609.13182#bib.bib8)\], which captures how an image’s information content scales with spatial resolution\.
- •*Spatial\-pattern heterogeneity*via gliding\-box lacunarity\[[9](https://arxiv.org/html/2609.13182#bib.bib9),[10](https://arxiv.org/html/2609.13182#bib.bib10),[22](https://arxiv.org/html/2609.13182#bib.bib22)\], which complements fractal dimension by quantifying gaps and clustering at multiple scales\.
- •*Temporal complexity*via multi\-scale entropy\[[11](https://arxiv.org/html/2609.13182#bib.bib11)\]and permutation entropy\[[12](https://arxiv.org/html/2609.13182#bib.bib12)\], which probe pattern complexity at multiple coarse\-graining scales\.
- •*Learned deep features*via convolutional backbones pretrained on large natural\-image corpora\[[13](https://arxiv.org/html/2609.13182#bib.bib13),[15](https://arxiv.org/html/2609.13182#bib.bib15)\], which have become standard upstream encoders in remote sensing\[[23](https://arxiv.org/html/2609.13182#bib.bib23),[24](https://arxiv.org/html/2609.13182#bib.bib24)\]\.

Each captures a different facet of visual organisation\. For a spatially fixed site under climate stress, the empirical question is which feature, or which combination, varies most diagnostically with the underlying state\.

### 2\.3Great Salt Lake hydrology

The Great Salt Lake is a hyper\-saline endorheic remnant of Pleistocene Lake Bonneville, fed by the Bear, Weber, and Jordan rivers under snowpack\-driven seasonality\[[25](https://arxiv.org/html/2609.13182#bib.bib25)\]\. A 1959 railroad causeway divides the lake into a hyper\-saline north arm \(whereSpiral Jettysits\) and a less saline south arm\[[26](https://arxiv.org/html/2609.13182#bib.bib26)\]\. Since the late 1980s the lake has declined steadily; by 2022–23 both arms reached their lowest recorded elevations\. Wurtsbaugh et al\.\[[2](https://arxiv.org/html/2609.13182#bib.bib2)\], Null and Wurtsbaugh\[[4](https://arxiv.org/html/2609.13182#bib.bib4)\], and Abbott et al\.\[[3](https://arxiv.org/html/2609.13182#bib.bib3)\]attribute the decline to upstream consumptive water use exceeding inflow under a warming and drying climate, with substantial downstream consequences for air quality \(dust from the exposed playa\), avian habitat, and the lake’s brine\-shrimp ecosystem\.

### 2\.4Climate\-conditioned and unsupervised vision for Earth observation

Climate\-conditioned generative models have been proposed for flood and sea\-level visualisation\[[27](https://arxiv.org/html/2609.13182#bib.bib27),[28](https://arxiv.org/html/2609.13182#bib.bib28)\]and precipitation nowcasting\[[29](https://arxiv.org/html/2609.13182#bib.bib29)\]\. Unsupervised vision representations have emerged as powerful encoders of Earth\-observation imagery\[[24](https://arxiv.org/html/2609.13182#bib.bib24),[23](https://arxiv.org/html/2609.13182#bib.bib23)\]\. The present paper exploits ImageNet transfer as an*unsupervised probe*: we hypothesise — and confirm in Section[6](https://arxiv.org/html/2609.13182#S6)— that climate\-modulated visual change at a fixed site is salient enough to appear as a principal direction in the activation space of a network never trained on any climate target\.

## 3Data

### 3\.1Study site

The study site is a5​km×5​km5~\\mathrm\{km\}\\times 5~\\mathrm\{km\}bounding box centred on Rozel Point, Utah \(41\.4378∘41\.4378^\{\\circ\}N,−112\.6685∘\-112\.6685^\{\\circ\}W\), the northeastern tip of the GSL’s north arm\. The box encloses the entireSpiral Jetty\(460 m arm length,∼\\sim5 m wide\) plus surrounding salt\-crust playa, basalt outcrops, and \(depending on lake stage\) water surface\.

### 3\.2Satellite archive

We retrieved every Landsat Collection\-2 Level\-2 \(Landsat 4–5 TM, Landsat 7 ETM\+, Landsat 8–9 OLI/TIRS\) and Sentinel\-2 Level\-2A scene intersecting the study bounding box for 1984\-01\-01 to 2025\-12\-31 via Microsoft Planetary Computer’s STAC catalog\[[30](https://arxiv.org/html/2609.13182#bib.bib30)\], filtered toeo:cloud\_cover < 20%\. Each scene was cropped to the same geographic window and rasterised to a uniform430×564430\\times 564px RGB JPEG by linear 99th\-percentile reflectance stretch on bands red–green–blue \(Landsat\) or B04–B03–B02 \(Sentinel\-2\)\. The final archive contains1,289 Landsat scenesand464 Sentinel\-2 scenes — 1,753 raw chips, of which 1,744 passed feature\-computation quality control\.

### 3\.3Coverage

Figure[1](https://arxiv.org/html/2609.13182#S6.F1)\(Section[6](https://arxiv.org/html/2609.13182#S6)\) shows the year×\\timesmonth coverage\. Every calendar year 1984–2025 is represented \(7 to 150 scenes per year, growing with sensor cadence and with the Sentinel\-2 launch in 2015\)\. Every calendar month is sampled \(74 to 239 scenes per month\), with the expected summer\-skewed distribution arising from cloud\-free acquisition density\.

### 3\.4Climate panel

A monthly climate panel 1984–2025 \(504 rows\) was assembled from open sources:

- •NASA GISTEMP v4\[[31](https://arxiv.org/html/2609.13182#bib.bib31)\]: global land\-ocean monthly temperature anomaly, baseline 1951–1980\.
- •Open\-Meteo Historical Reanalysis\[[32](https://arxiv.org/html/2609.13182#bib.bib32)\]: daily 2\-meter mean temperature at the Salt Lake City station \(40\.7608∘40\.7608^\{\\circ\}N,−111\.8910∘\-111\.8910^\{\\circ\}W\), aggregated to monthly mean \(∘C and∘F\)\.
- •USGS NWIS\[[33](https://arxiv.org/html/2609.13182#bib.bib33)\]: daily lake elevation \(NGVD29 datum, ft; parameter code 62614\) at the Saltair gauge \(USGS 10010000, south arm, since 1847\) and the Saline gauge \(USGS 10010100, north arm, since 1966\), aggregated to monthly mean\.
- •Global Carbon Budget\[[34](https://arxiv.org/html/2609.13182#bib.bib34)\]: annual fossil CO2emissions \(Gt\) 1984–2025\.
- •GSL salinity: yearly mean north\-arm salinity \(g/L\) ingested from the pilot\-study workbook\[[5](https://arxiv.org/html/2609.13182#bib.bib5)\]and the MDPI Water 2023 dataset\[[35](https://arxiv.org/html/2609.13182#bib.bib35)\], forward/back\-filled across gap years and flagged\.

Yearly aggregation gives the 42\-year panel used as the primary inference unit\. The monthly panel is used for scene\-level partial correlations \(Section[6\.8](https://arxiv.org/html/2609.13182#S6.SS8)\) and the STL decomposition \(Section[6\.7](https://arxiv.org/html/2609.13182#S6.SS7)\)\.

## 4Multi\-Feature Complexity Signature

Each scene chip is converted to 8\-bit grayscale \(ITU\-R BT\.601 luminance\) and resized to512×512512\\times 512px for feature comparability\. The signature comprises 14 classical features plus four CNN\-PCA components\.

### 4\.1Classical features

#### Shannon entropy\.

H=−∑i=0255pi​log2⁡piH=\-\\sum\_\{i=0\}^\{255\}p\_\{i\}\\log\_\{2\}p\_\{i\}, wherepip\_\{i\}is the empirical intensity histogram\[[21](https://arxiv.org/html/2609.13182#bib.bib21)\]\. Computed viaskimage\.measure\.shannon\_entropy\.

#### Multi\-scale permutation entropy\.

Permutation entropy\[[12](https://arxiv.org/html/2609.13182#bib.bib12)\]at embedding dimensionm=3m=3and time delayτ=1\\tau=1, computed on the row\-major flattened intensity series at three coarse\-graining scaless∈\{1,2,4\}s\\in\\\{1,2,4\\\}\[[11](https://arxiv.org/html/2609.13182#bib.bib11)\]\. Yields three featurespe\_scale1,pe\_scale2,pe\_scale4\.

#### Box\-counting fractal dimension\.

Otsu\-binarised image, box countsN​\(s\)N\(s\)at scaless∈\{2,4,8,16,32,64\}s\\in\\\{2,4,8,16,32,64\\\}px, fractal dimensionD=slope​\(log⁡N​vs\.​log⁡\(1/s\)\)D=\\mathrm\{slope\}\\left\(\\log N\\text\{ vs\.\\ \}\\log\(1/s\)\\right\)\[[7](https://arxiv.org/html/2609.13182#bib.bib7),[36](https://arxiv.org/html/2609.13182#bib.bib36)\]\.

#### Gliding\-box lacunarity\.

Following Allain and Cloitre\[[9](https://arxiv.org/html/2609.13182#bib.bib9)\]and Plotnick et al\.\[[10](https://arxiv.org/html/2609.13182#bib.bib10)\], computed atr∈\{4,8,16\}r\\in\\\{4,8,16\\\}px on the Otsu\-binarised image; three features\.

#### GLCM texture\.

Gray\-level co\-occurrence matrix at 64 quantisation levels, four directions \(0∘,45∘,90∘,135∘0^\{\\circ\},45^\{\\circ\},90^\{\\circ\},135^\{\\circ\}\), distance 1, symmetric and normalised\. Four features extracted viagraycoprops: contrast, homogeneity, correlation, energy\[[6](https://arxiv.org/html/2609.13182#bib.bib6)\]\.

#### First\-order intensity statistics\.

Mean, standard deviation, 10th\-percentile, and 90th\-percentile of the grayscale intensity\. These provide baselines against which higher\-order claims must be benchmarked\.

### 4\.2Deep embedding

Each RGB chip is normalised by the ImageNet mean/std and passed through ImageNet\-V2\-pretrained ResNet50\[[13](https://arxiv.org/html/2609.13182#bib.bib13)\]\. The 2,048\-dim avg\-pool activation is taken as the embedding\. Across all 1,744 chips we fit a 16\-component PCA; the first four components are retained as featurescnn\_pca1–cnn\_pca4, explaining 34%, 24%, 5%, and 4% of variance respectively\. The remaining 12 are saved for future analyses but not used in the primary results\.

### 4\.3Implementation and runtime

The classical\-feature pass implements Shannon, permutation, fractal, lacunarity, GLCM, and intensity descriptors in NumPy/SciPy/scikit\-image\. The deep\-embedding pass uses PyTorch with the official ResNet50 IMAGENET1K\_V2 weights\. The full 1,744\-chip pipeline runs in∼\\sim12 min for classical features plus∼\\sim1 min for the ResNet50 forward pass on CPU\. Code and weights are pinned for reproducibility\.

## 5Statistical Framework

### 5\.1Year aggregation

The 1,744 scenes are aggregated by year \(mean of each feature across all scenes in that year\), producing a 42\-year panel used as the primary inference unit\. Within\-year variance is retained for the seasonal heatmap \(Figure[1](https://arxiv.org/html/2609.13182#S6.F1)\) and the partial\-correlation analysis \(Section[6\.8](https://arxiv.org/html/2609.13182#S6.SS8)\)\.

### 5\.2Bootstrap confidence intervals

For each \(feature, climate\-variable\) pair we compute Pearsonrrand Spearmanρ\\rhoon the 42\-year panel, with 5,000\-resample percentile bootstrap 95 % confidence intervals on paired \(feature, climate\) draws\. A correlation is reported as*robust*when its CI excludes zero\.

### 5\.3Lag analysis

The Shannon\-entropy↔\\leftrightarrownorth\-arm\-elevation Pearson correlation is computed at integer year lagsℓ∈\{−5,−4,…,\+4,\+5\}\\ell\\in\\\{\-5,\-4,\\ldots,\+4,\+5\\\}, with bootstrap CIs at each lag\. Positiveℓ\\ellindicates “entropy at yearttcorrelates with lake elevation at yeart\+ℓt\+\\ell”\.

### 5\.4Partial correlations

To address two reviewer concerns, we compute scene\-level Spearman partial correlations controlling for \(a\) calendar month and \(b\) sensor identity:

1. 1\.For each \(feature, climate\) pair, regress feature on month dummies and take residuals; regress climate on the same month dummies and take residuals; Spearman correlate the two residuals\. This isolates the effect of climate from the dominant seasonal cycle\.
2. 2\.Repeat with both month and sensor dummies \(Landsat 4/5/7/8/9 vs\. Sentinel\-2\)\. This isolates climate from any sensor\-specific spectral response or calibration confound\.

### 5\.5STL decomposition

Monthly mean Shannon entropy is decomposed into trend, annual seasonal, and residual components via robust STL with period=12=12\[[37](https://arxiv.org/html/2609.13182#bib.bib37)\]\. This separates the slow long\-term envelope of salt\-crust morphology change from the within\-year seasonal cycle\.

## 6Results

### 6\.1Dataset coverage

The final archive contains 1,744 co\-registered chips spanning 1984–2025, with all 12 calendar months sampled and all 42 years represented\. Mean Shannon entropy across the archive is 4\.87 bits\. Figure[1](https://arxiv.org/html/2609.13182#S6.F1)shows the year×\\timesmonth entropy heatmap; the dominant seasonal banding is visible at a glance, with secondary multi\-year structure that the STL decomposition \(Section[6\.7](https://arxiv.org/html/2609.13182#S6.SS7)\) separates\.

![Refer to caption](https://arxiv.org/html/2609.13182v1/fig2_seasonal_heatmap.png)Figure 1:Year×\\timesmonth mean Shannon entropy across all 1,744 scenes\. Strong annual seasonality is visible as horizontal banding; the post\-2015 entropy regime shift is visible as a vertical change in row intensity\.
### 6\.2The Great Salt Lake decline

The USGS north\-arm gauge \(Saline, USGS 10010100\) records a lake elevation that peaked at4,210\.22 ft NGVD29 in 1987— the apex of the 1980s wet\-cycle floods — and fell to4,189\.25 ft in 2023, the lowest reading in the gauge’s 60\-year instrumental record \(a20\.96 ft / 6\.39 m declinein 36 years\)\. The south\-arm Saltair gauge tracks the north arm with a∼\\sim1 ft offset\. Over the same window the NASA GISTEMP global anomaly rose from\+0\.15∘\+0\.15^\{\\circ\}C in 1984 to\+1\.19∘\+1\.19^\{\\circ\}C in 2025\. These two trajectories — a vertically collapsing endorheic lake and a warming planet — bracket the climate envelope in whichSpiral Jettyhas emerged from a half\-century of submergence\.

![Refer to caption](https://arxiv.org/html/2609.13182v1/fig1_entropy_vs_gsl_elev.png)Figure 2:Mean Shannon entropy per year \(dark, left axis\) overlaid with GSL north\-arm lake elevation \(orange, right axis\) 1984–2025\. The 2022–23 historic GSL minimum is highlighted in red\. Note the non\-monotonic envelope of the entropy trajectory and the strong divergence post\-2015\.
### 6\.3Shannon entropy alone is a weak climate proxy

Table[1](https://arxiv.org/html/2609.13182#S6.T1)reports the year\-aggregated Spearman correlations between Shannon entropy and each climate variable, with 5,000\-resample bootstrap 95 % CIs\.

Table 1:Year\-aggregated Spearman correlations between Shannon entropy and each climate variable \(n=42n=42years; 95 % bootstrap CI\)\.Only one of six Shannon\-entropy correlations \(with salinity,ρ=−0\.52\\rho=\-0\.52\) has a bootstrap CI excluding zero\. The signs of the temperature and carbon correlations arenegative— contradicting the pilot study’s positive\-correlation claim\. None of these single\-feature signals would survive a strict multiple\-testing correction across the 14\-feature×\\times7\-variable hypothesis grid\.*Shannon entropy alone is not the climate sensor the popular framing has suggested\.*

### 6\.4The multi\-feature signature reveals robust signal

Figure[3](https://arxiv.org/html/2609.13182#S6.F3)shows the full Spearman correlation matrix\. Table[2](https://arxiv.org/html/2609.13182#S6.T2)lists the ten strongest correlations across the 14×\\times7 grid\.

![Refer to caption](https://arxiv.org/html/2609.13182v1/fig4_corr_heatmap_full.png)Figure 3:Spearman correlation matrix between the 14 complexity features \(rows\) and 7 climate variables \(columns\), year\-aggregated,n=42n=42\. Thepe\_scale2,pe\_scale4,mean\_intensity, andcnn\_pca3rows carry the strongest robust signal\.Table 2:Top\-10 Spearman correlations across the 14\-feature×\\times7\-variable grid \(n=42n=42years; bootstrap 95 % CI\)\.Three observations follow\.

#### Coarse\-scale permutation entropy is the strongest lake\-stage proxy\.

pe\_scale4andpe\_scale2attainρ≈\+0\.86\\rho\\approx\+0\.86with both arm elevations — 3×\\timesstronger than Shannon entropy\. Coarse\-graining by a factor of 4 aggregates pixel intensities into mesoscale tiles of roughly 10–50 m, approximately matching the spatial scale of the salt\-crust polygons that form on the playa as the lake recedes\[[38](https://arxiv.org/html/2609.13182#bib.bib38)\]\. The permutation\-entropy operator then probes the ordinal structure of these mesoscale tiles, capturing precisely the kind of pattern complexity that crystallisation\-dissolution cycles produce\.

#### Mean intensity — the simplest possible feature — tracks lake elevation almost as strongly\.

mean\_intensityreachesρ=\+0\.85\\rho=\+0\.85with elevation in both arms\. When the lake is high, more pixels are water \(high reflectance\); when low, exposed playa and salt crust dominate but at a different reflectance regime\. Any complexity\-claim must be benchmarked against this baseline; we return to this in Section[7](https://arxiv.org/html/2609.13182#S7)\.

#### ResNet50 PC\-3 is an emergent unsupervised climate axis\.

The third principal component of the ResNet50 avg\-pool embeddings reachesρ=\+0\.86\\rho=\+0\.86with cumulative CO2emissions andρ=−0\.83\\rho=\-0\.83with lake elevation\.*The network was trained only on ImageNet classification; no climate target was ever in its loss\.*Yet a single principal direction of its activation space lines up almost perfectly with the macroscopic environmental gradient at the site\. Figure[4](https://arxiv.org/html/2609.13182#S6.F4)shows the projection of all 1,744 chips on PC\-1 vs\. PC\-2 \(58 % of variance combined\), coloured by year; a clean monotone drift from the 1980s–90s cluster to the 2020s cluster is visible without supervision\.

![Refer to caption](https://arxiv.org/html/2609.13182v1/fig6_pca_embedding.png)Figure 4:ResNet50 avg\-pool embeddings of all 1,744Spiral Jettyscenes projected onto PC\-1 vs\. PC\-2 \(58 % of variance combined\), coloured by year\. The monotone drift from light \(1980s–90s\) to dark \(2020s\) is the unsupervised visual signature of GSL decline\.

### 6\.5Per\-feature trajectories

Figure[5](https://arxiv.org/html/2609.13182#S6.F5)shows the year\-aggregated trajectories of six representative features\. Thepe\_scale4andmean\_intensitypanels show the same characteristic envelope as the GSL elevation trace — non\-monotonic rise to mid\-2010s then collapse — whereasglcm\_contrastandlacunarity\_8are dominated by sensor era effects \(the Landsat 5 to Landsat 7/8 transition around 1999–2014 is visible as a step change\)\.

![Refer to caption](https://arxiv.org/html/2609.13182v1/fig3_signature_panels.png)Figure 5:Year\-aggregated trajectories of six representative complexity features 1984–2025\.shannon,fractal\_dim, andlacunarity\_8\(top three\) show mixed trends including sensor\-era step changes;glcm\_contrast,glcm\_homogeneity, andmean\_intensity\(bottom three\) show climate\-correlated envelopes\.
### 6\.6Image complexity leads lake stage by approximately three years

The Shannon\-entropy↔\\leftrightarrownorth\-arm\-elevation Pearson correlation as a function of lagℓ\\ellpeaks atℓ=\+3\\ell=\+3yr with Pearsonr=\+0\.58r=\+0\.58\(95 % CI\[\+0\.40,\+0\.73\]\[\+0\.40,\+0\.73\]\)\. The entire bandℓ∈\{\+1,\+5\}\\ell\\in\\\{\+1,\+5\\\}has bootstrap CIs excluding zero; contemporaneous and backward lags do not \(Figure[6](https://arxiv.org/html/2609.13182#S6.F6)\)\.

![Refer to caption](https://arxiv.org/html/2609.13182v1/fig5_lag_correlations.png)Figure 6:Lag\-correlation between Shannon entropy at yearttand GSL north\-arm elevation at yeart\+ℓt\+\\ellforℓ∈\{−5,\+5\}\\ell\\in\\\{\-5,\+5\\\}years, with bootstrap 95 % CIs\. The peak atℓ=\+3\\ell=\+3yr is consistent with a∼\\sim3\-year forward lead of visual complexity over hydrological state\.We interpret this as evidence that*image complexity at Spiral Jetty is a leading indicator of lake stage on a∼\\sim3\-year time scale*\. Two mechanisms are consistent with the data:

1. 1\.The salt\-crust regime integrates the past 1–3 yr of precipitation\-evaporation history before reaching a new visual steady state\. Salt\-crust polygons grow, fragment, and reorganise on multi\-year time scales, providing a kind of geophysical memory that is visible in the image complexity before it is registered as a step in lake stage\.
2. 2\.Lake stage itself is autocorrelated on multi\-year time scales due to basin\-storage inertia; any quantity correlated with lake stage at lag 0 will also correlate \(with comparable strength\) at modest positive lags\.

The two hypotheses cannot be distinguished by the present correlational design; the companion paper’s physics\-coupling layer provides the mechanistic test\.

### 6\.7Seasonality and a non\-monotonic long\-term trend

STL decomposition of monthly mean Shannon entropy \(Figure[7](https://arxiv.org/html/2609.13182#S6.F7)\) reveals three components\.

Theannual seasonal componenthas peak\-to\-trough amplitude of roughly±2\\pm 2bits, peaking in March–May \(snowmelt\-driven turbidity, transient wet salt crust\) and troughing in late summer \(dry uniform salt pan; high algal bloom\)\. The amplitude*grows*visibly after∼\\sim2015, suggesting the salt\-crust regime became more*variable*— not just lower — as the lake destabilised\.

Thelong\-term trendis non\-monotonic\. It rises from∼\\sim4\.7 bits in 1984 to∼\\sim6\.5 bits around 2015, then collapses to∼\\sim4\.5 bits by 2020 and remains flat through 2025\. The collapse window coincides precisely with the GSL’s descent to historic\-record\-low elevations in 2022–23, and marks the transition from a heterogeneous “lake\-edge” visual regime \(basalt \+ transient salt crust \+ brine\) to a homogeneous “dry\-playa” regime \(salt pan \+ exposed basalt\) in which textural information saturates\.

Theresidualis well\-bounded around zero with larger excursions post\-2015, consistent with the increased volatility implied by the growing seasonal amplitude\.

The dominance of the seasonal cycle over the trend is what made the pilot study’sn=13n=13annual analysis statistically unreliable: averaging across season removes the largest within\-year noise source, but only when the seasonal sample is balanced — which 13 years of single\-season images is not\. The non\-monotonic trend additionally implies that linear annual correlations underestimate the structure of the signal\. Piecewise or threshold\-based models of the lake\-edge→\\todry\-playa transition are likely a more faithful framework, motivating the threshold\-style hydrological mask we will adopt in the companion paper\.

![Refer to caption](https://arxiv.org/html/2609.13182v1/fig7_stl_decomposition.png)Figure 7:STL decomposition of monthly mean Shannon entropy 1984–2025 \(top: raw; second: trend; third: annual seasonal; bottom: residual\)\. The trend \(orange\) is non\-monotonic: it rises from∼\\sim4\.7 bits in 1984 to∼\\sim6\.5 bits around 2015, then collapses to∼\\sim4\.5 bits by 2020 — the salt\-crust “lake\-edge” regime giving way to a “dry\-playa” regime in which textural information saturates\.
### 6\.8Partial correlations: month and sensor effects

The strong scene\-level Spearman correlations of Table[2](https://arxiv.org/html/2609.13182#S6.T2)could in principle be driven by \(a\) the seasonal cycle \(winter scenes disproportionately come from years with different lake levels\) or \(b\) sensor\-specific spectral response \(newer\-generation sensors with finer calibration entered service in 2013–2017, the same window as the trend collapse\)\. To rule these out, we computed scene\-level partial Spearman correlations onn=1,428n=1\{,\}428scenes \(those with all needed monthly climate values\), residualising both feature and climate on \(i\) calendar\-month dummies and \(ii\) month \+ sensor\-identity dummies\. Table[3](https://arxiv.org/html/2609.13182#S6.T3)reports headline rows\.

Table 3:Scene\-level Spearman correlations between four key features and each climate variable: raw, partial controlling for calendar month, and partial controlling for month \+ sensor identity \(n=1,428n=1\{,\}428scenes\)\.Two findings emerge\.

First,the month effect is small: raw and month\-partialled correlations differ by at most∼\\sim0\.05 in magnitude\. The seasonal cycle is therefore*not*the primary driver of the scene\-level signals\.

Second,sensor identity is a substantial confounderat the scene level\. After controlling for both month and sensor, the headline correlations halve: Shannon–GSL north elevation moves from−0\.65\-0\.65to−0\.36\-0\.36;cnn\_pca3–carbon moves from\+0\.77\+0\.77to\+0\.37\+0\.37\. This is because newer sensors \(Landsat 8/9, Sentinel\-2\) entered the archive in the same era as the GSL’s catastrophic decline \(post\-2013\), and they have different spectral response, geometric resolution, and pre\-processing chains\. The conservative reading is that scene\-level correlations should be reported with sensor controls included\.

#### Importantly, the year\-aggregated correlations of Table[2](https://arxiv.org/html/2609.13182#S6.T2)are far less affected by this confound

, because each year is sampled by all available sensors \(in proportion to their cadence\) and the within\-year mean averages over them\. The year\-aggregatedρ≈\+0\.85\\rho\\approx\+0\.85to\+0\.88\+0\.88lake\-elevation correlations and the unsupervised CNN climate axis are therefore robust findings, but reviewers and downstream users should be aware that*thestrengthof the scene\-level correlations is partially driven by sensor era*, while the*existence*of the signal is not\.

## 7Discussion

### 7\.1What the data say, and what they do not

Combining the year\-aggregated, lag, STL, and partial\-correlation analyses, the empirical picture is:

- •The strongest robust signals inSpiral Jetty’s visual complexity are with lake elevation \(via coarse\-scale permutation entropy and mean intensity,ρ≈\+0\.85\\rho\\approx\+0\.85to\+0\.88\+0\.88\) and with the carbon\-temperature axis \(via the third principal component of a pretrained ResNet50,ρ≈\+0\.86\\rho\\approx\+0\.86\)\.
- •Shannon entropy alone is a weak proxy whose only robust correlation is with salinity\. The pilot study’s positive\-temperature claim*does not survive*the larger\-sample analysis\.
- •Image complexity*leads*lake stage by approximately three years, with the lag\-correlation bandℓ∈\{\+1,\+5\}\\ell\\in\\\{\+1,\+5\\\}excluding zero\.
- •The long\-term trend is non\-monotonic, collapsing post\-2015 in coincidence with the GSL’s historic\-record\-low elevations — a threshold transition from a heterogeneous “lake\-edge” regime to a homogeneous “dry\-playa” regime\.
- •Scene\-level signals are partially driven by sensor era, but year\-aggregated signals are robust to that confound\.

The popular “art\-as\-thermometer” framing therefore*does not survive the data literally*but*survives in refined form*:Spiral Jetty, properly instrumented with a multi\-feature complexity signature, is aleading indicator of regional hydro\-climatic state— functionally a thermometer with several years of memory\.

### 7\.2Caveats

Three caveats temper this reading\.

#### Single\-feature analyses underestimate the signal\.

The pilot study’s reliance on Shannon entropy alone systematically missed the much stronger signals carried by permutation entropy, mean intensity, and CNN embeddings\. We recommend that complexity\-based studies of cultural\-heritage sites under climate stress adopt multi\-feature signatures by default, and that “single\-feature” positive findings be regarded skeptically until benchmarked against signature\-level alternatives\.

#### First\-order radiometry carries part of the signal\.

mean\_intensityalone reachesρ=\+0\.85\\rho=\+0\.85with lake elevation, suggesting a non\-trivial fraction of the climate signal is simple radiometric — bright water vs\. dark playa — rather than higher\-order texture\. Sophisticated\-feature claims should declare their marginal contribution over first\-order baselines\. In this study,pe\_scale4adds∼\\sim0\.03 inρ\\rhoovermean\_intensityfor elevation but is much stronger for the carbon axis, andcnn\_pca3contributes the carbon signal uniquely\.

#### A three\-year forward lead is genuinely informative\.

Despite the radiometric simplicity caveat, neither the magnitude nor the lag of the three\-year lead is predicted by instantaneous radiometry alone\. A 3\-year lead implies a slow integrative process in salt\-crust morphology — which is hydrologically plausible\[[38](https://arxiv.org/html/2609.13182#bib.bib38)\]but not demonstrated mechanistically here\. The companion paper’s physics\-coupling layer is designed to provide that mechanistic demonstration\.

### 7\.3Ethical and aesthetic considerations

Smithson designedSpiral Jettyto dramatise entropy at geological time scale\. The work has now disordered on a*climatological*time scale Smithson could not have anticipated\. Treating his earthwork as a climate sensor risks instrumentalising what was always meant as a contemplative encounter with deep time\. We adopt the framing because the artwork*does*record climate; we resist any framing in which the artwork’s meaning is*reduced*to its sensor function\. The two readings — aesthetic and instrumental — are complementary, not competitive\.

### 7\.4Implications for cultural\-heritage monitoring

The methodology generalises to other land\-art sites under climate stress: Smithson’sSpiral Hill, Walter De Maria’sLightning Field, Nancy Holt’sSun Tunnels, Michael Heizer’sDouble NegativeandCity, and James Turrell’sRoden Crater\. More broadly, the framework can be applied to coastal heritage sites \(submerging at varying rates\), to permafrost archaeological sites \(degrading as the ground warms\), and to ice\-shelf\-adjacent built heritage\. In each case, a multi\-feature complexity signature on a public archive of satellite imagery, validated through year\-aggregation and bootstrap CIs, controlled for sensor confounds, and decomposed into seasonal and trend components, can produce a defensible measurement of the site’s climate response without requiring on\-the\-ground instrumentation\.

### 7\.5Limitations and future work

The dataset is restricted to 1984–2025 by Landsat\-4 launch; the pilot study’s 1970–1984 chips were heterogeneous and could not be co\-registered to the same standard\. Higher\-resolution commercial satellite imagery \(WorldView, Planet\) post\-2008 could enable sub\-decimetre analysis but at the cost of openness; we excluded it to keep the dataset reproducible\. The salinity series is sparse before 2012 and forward/back\-filled; this is the weakest variable in our climate panel and we report it as such\. Finally, the partial\-correlation analysis treats sensor identity as a single categorical control; a more refined approach would model the spectral response curves explicitly\.

## 8Conclusions

We have presented the first multi\-decadal, multi\-feature remote\-sensing characterisation of Robert Smithson’sSpiral Jetty, drawn from 1,744 Landsat 4–9 and Sentinel\-2 scenes spanning 1984–2025 and aligned with a 42\-year monthly climate panel\. The strongest robust correlations between visual complexity and climate are carried by coarse\-scale permutation entropy and CNN\-embedding components, not by Shannon entropy alone\. A principal direction of an ImageNet\-pretrained ResNet50 activation space emerges, without supervision, as a climate axis\. Image complexity at the jetty leads lake stage by approximately three years\. The long\-term trend is non\-monotonic and exhibits a threshold transition in 2015–2020 that linear annual correlations obscure\. Scene\-level signals are partially driven by sensor era, but year\-aggregated signals are robust\.

A forthcoming companion paper will use the dataset released here to build a climate\-conditioned latent\-diffusion model with a hydrological physics\-coupling layer, forecasting the site under IPCC SSP1\-2\.6, SSP2\-4\.5, and SSP5\-8\.5 scenarios for 2030 and 2050\.

## Author Contributions

Conceptualisation, A\.C\.; methodology, A\.C\. and S\.S\.K\.; software, S\.S\.K\.; validation, S\.S\.K\.; formal analysis, S\.S\.K\.; investigation, A\.C\. and S\.S\.K\.; resources, S\.S\.K\.; data curation, S\.S\.K\.; writing—original draft preparation, S\.S\.K\.; writing—review and editing, A\.C\. and S\.S\.K\.; visualisation, S\.S\.K\.; supervision, S\.S\.K\.; project administration, A\.C\.; funding acquisition, n/a\. All authors have read and agreed to the published version of the manuscript\.

## Funding

This research received no external funding\.

## Data Availability Statement

The dataset \(spiral\_jetty\_complexity\_v1: 1,744 co\-registered scene chips, complexity signature, monthly climate panel\), the full analysis pipeline, and the reproducible figures are available on Zenodo \(DOI on acceptance\) and GitHub:[github\.com/alevcinbarci/spiral\-jetty\-ai](https://github.com/alevcinbarci/spiral-jetty-ai), release tagrs\-mdpi\-2026\. Source remote\-sensing scenes are public via Microsoft Planetary Computer \(Landsat C2L2 and Sentinel\-2 L2A\)\.

## Acknowledgments

We thank the U\.S\. Geological Survey Utah Water Science Center, NASA GISTEMP, the Open\-Meteo project, Microsoft Planetary Computer, and the European Space Agency Copernicus programme for the open data that made this study possible\.

## Conflicts of Interest

The authors declare no conflict of interest\.

## References

- \[1\]Robert Smithson\.The Spiral Jetty\.In G\. Kepes \(ed\.\),*Arts of the Environment*, George Braziller, New York, 1972\.
- \[2\]Wayne A\. Wurtsbaugh, Craig Miller, Sarah E\. Null, R\. Justin DeRose, Peter Wilcock, Maura Hahnenberger, Frank Howe, and Johnnie Moore\.Decline of the world’s saline lakes\.Nature Geoscience, 10:816–821, 2017\.
- \[3\]Benjamin W\. Abbott et al\.Emergency measures needed to rescue Great Salt Lake from ongoing collapse\.Brigham Young University public report, 2023\.
- \[4\]Sarah E\. Null and Wayne A\. Wurtsbaugh\.Water development, agriculture, and the decline of the Great Salt Lake\.Journal of Hydrology, 588:125094, 2020\.
- \[5\]Alev Cinbarcı\.Pilot shannon\-entropy analysis ofSpiral Jettysatellite imagery 1970–2022\.Unpublished pilot dataset, workbook archive, 2024\.
- \[6\]Robert M\. Haralick, K\. Shanmugam, and Its’hak Dinstein\.Textural features for image classification\.IEEE Transactions on Systems, Man, and Cybernetics, SMC\-3\(6\):610–621, 1973\.
- \[7\]Benoît B\. Mandelbrot\.The Fractal Geometry of Nature\.W\. H\. Freeman, 1982\.
- \[8\]Shaun Lovejoy and Daniel Schertzer\.Scale invariance, symmetries, fractals, and stochastic simulations of atmospheric phenomena\.Bulletin of the American Meteorological Society, 67:21–32, 1986\.
- \[9\]C\. Allain and M\. Cloitre\.Characterizing the lacunarity of random and deterministic fractal sets\.Physical Review A, 44\(6\):3552–3558, 1991\.
- \[10\]Roy E\. Plotnick, Robert H\. Gardner, W\. W\. Hargrove, K\. Prestegaard, and M\. Perlmutter\.Lacunarity analysis: A general technique for the analysis of spatial patterns\.Physical Review E, 53\(5\):5461–5468, 1996\.
- \[11\]Madalena Costa, Ary L\. Goldberger, and C\.\-K\. Peng\.Multiscale entropy analysis of complex physiologic time series\.Physical Review Letters, 89\(6\):068102, 2002\.
- \[12\]Christoph Bandt and Bernd Pompe\.Permutation entropy: A natural complexity measure for time series\.Physical Review Letters, 88\(17\):174102, 2002\.
- \[13\]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun\.Deep residual learning for image recognition\.InIEEE Conference on Computer Vision and Pattern Recognition \(CVPR\), pages 770–778, 2016\.
- \[14\]Edward J\. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen\-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen\.LoRA: Low\-rank adaptation of large language models\.InInternational Conference on Learning Representations \(ICLR\), 2022\.
- \[15\]Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V\. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El\-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po\-Yao Huang, Shang\-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Héréan Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski\.DINOv2: Learning robust visual features without supervision\.arXiv preprint arXiv:2304\.07193, 2023\.
- \[16\]Robert Smithson\.A tour of the monuments of Passaic, New Jersey\.Artforum, December 1967, 1967\.
- \[17\]Jack Flam, editor\.Robert Smithson: The Collected Writings\.University of California Press, 1996\.
- \[18\]Ann Reynolds\.Robert Smithson: Learning from New Jersey and Elsewhere\.MIT Press, 2003\.
- \[19\]Suzaan Boettger\.Earthworks: Art and the Landscape of the Sixties\.University of California Press, 2002\.
- \[20\]Jennifer L\. Roberts\.The taste of time: Salt and Spiral Jetty\.InMirror Travels: Robert Smithson and History\. Yale University Press, 2004\.
- \[21\]Claude E\. Shannon\.A mathematical theory of communication\.Bell System Technical Journal, 27:379–423, 1948\.
- \[22\]Pinliang Dong\.Test of a new lacunarity estimation method for image texture analysis\.International Journal of Remote Sensing, 21\(17\):3369–3373, 2000\.
- \[23\]Adam J\. Stewart, Caleb Robinson, Isaac A\. Corley, Anthony Ortiz, Juan M\. Lavista Ferres, and Arindam Banerjee\.TorchGeo: Deep learning with geospatial data\.arXiv preprint arXiv:2111\.08872, 2022\.
- \[24\]Oscar Manas, Alexandre Lacoste, Xavier Giro i Nieto, David Vazquez, and Pau Rodriguez\.Seasonal contrast: Unsupervised pre\-training from uncurated remote sensing data\.arXiv preprint arXiv:2103\.16607, 2021\.
- \[25\]Robert L\. Baskin\.Occurrence and spatial distribution of microbial bioherms in Great Salt Lake\.Utah Geological Survey Open\-File Report, 2014\.
- \[26\]B\. R\. Loving, K\. E\. Waddell, and C\. W\. Miller\.Water and salt balance of Great Salt Lake, Utah, and simulation of water and salt movement through the causeway\.USGS Water\-Resources Investigations Report 00\-4221, 2000\.
- \[27\]Björn Lütjens, Brandon Leshchinskiy, Christian Requena\-Mesa, Farrukh Chishtie, Natalia Díaz\-Rodríguez, Océane Boulais, Aaron Piña, Aruna Sankaranarayanan, Aaron Schwerdtfeger, Yarin Gal, Freddie Kalaitzis, and Dava Newman\.Physics\-informed GANs for coastal flood visualization\.InNeurIPS Workshop on Tackling Climate Change with Machine Learning, 2021\.
- \[28\]Victor Schmidt, Alexandra Sasha Luccioni, Melisande Teng, Tianyu Zhang, Alexia Reynaud, Sunand Raghupathi, Gautier Cosne, Adrien Juraver, Vahe Vardanyan, Alex Hernandez\-Garcia, and Yoshua Bengio\.ClimateGAN: Raising climate change awareness by generating images of floods\.InInternational Conference on Learning Representations \(ICLR\), 2022\.
- \[29\]Suman Ravuri, Karel Lenc, Matthew Willson, Dmitry Kangin, Remi Lam, Piotr Mirowski, Megan Fitzsimons, Maria Athanassiadou, Sheleem Kashem, Sam Madge, Rachel Prudden, Amol Mandhane, Aidan Clark, Andrew Brock, Karen Simonyan, Raia Hadsell, Niall Robinson, Ellen Clancy, Alberto Arribas, and Shakir Mohamed\.Skilful precipitation nowcasting using deep generative models of radar\.Nature, 597:672–677, 2021\.
- \[30\]Microsoft\.Planetary Computer STAC API, 2025\.
- \[31\]NASA Goddard Institute for Space Studies\.GISTEMP v4 global land\-ocean temperature index, 2025\.
- \[32\]Open\-Meteo\.Historical weather API – ERA5 reanalysis, 2025\.
- \[33\]U\.S\. Geological Survey\.National water information system \(NWIS\) – Great Salt Lake gauges 10010000 \(saltair\) and 10010100 \(saline\)\.
- \[34\]Pierre Friedlingstein et al\.Global carbon budget 2023\.Earth System Science Data, 15:5301–5369, 2023\.
- \[35\]Bo Naumburg, Maura Hahnenberger, et al\.Salinity dynamics of the Great Salt Lake’s North Arm\.Water \(MDPI\), 15\(8\):1488, 2023\.
- \[36\]Larry S\. Liebovitch and Tibor Toth\.A fast algorithm to determine fractal dimensions by box counting\.Physics Letters A, 141\(8–9\):386–390, 1989\.
- \[37\]Robert B\. Cleveland, William S\. Cleveland, Jean E\. McRae, and Irma Terpenning\.STL: A seasonal\-trend decomposition procedure based on loess\.Journal of Official Statistics, 6\(1\):3–73, 1990\.
- \[38\]Tim K\. Lowenstein and Lawrence A\. Hardie\.Criteria for the recognition of salt\-pan evaporites\.Sedimentology, 32\(5\):627–644, 1985\.