Environmental programs are routinely judged by one number. Reforestation is judged by how green the satellite image looks, and decarbonization by tonnes of carbon dioxide equivalent. Those numbers can improve while the system they stand for gets worse. On China's Loess Plateau, vegetation greenness rose by 35.66% between 2000 and 2020. A composite satellite ecological index rose by 43.45%. Yet satellite gravimetry, which measures water by its gravitational pull, showed terrestrial water storage falling by about 0.6 mm per month, with the steepest losses in the most heavily reforested catchments (Vegetation Change and Eco-Environmental Quality Evaluation in the Loess Plateau). The quantity that went wrong was one nobody's headline metric watched.
We propose to treat this blind spot as a missing-data problem and to solve it with retrieval. We would build an external memory of places and periods where many indicators were measured together: greenness, surface temperature, wetness, biomass, water storage, water quality and habitat quality. A new site would then be assessed by looking up its closest analogues in that memory and reading off how the unmonitored indicators behaved there. The model that judges similarity stays frozen. The memory grows as monitoring data accumulates, much as nearest-neighbour language models improve by enlarging their text store without retraining (Generalization through Memorization: Nearest Neighbor Language Models). The central question is whether hidden-indicator estimates, and the detection of "green but drying" trade-offs, improve predictably as the memory grows. If they do, a cheap single-metric observation stream could be audited against expensive multi-metric ground truth gathered elsewhere.
Core Claim
Consider a site where only the cheap, satellite-visible signals are available. We claim that a frozen spatial encoder combined with a local regression over retrieved co-measured neighbours can estimate that site's unmonitored indicators, chiefly the trend in terrestrial water storage. Accuracy should improve monotonically with the number of sites in the memory. At equal data, adding sites to the memory should match retraining a parametric model, and it should beat retraining when new regions or new indicator columns arrive after the model is fixed.
The mechanism comes from Generalization Properties of Retrieval-based Models. That paper shows that Local Empirical Risk Minimization works when data are "locally regular", meaning a simple model is adequate inside a small neighbourhood even if the global relationship is complex. Local Empirical Risk Minimization fits a small model only on the $k$ retrieved neighbours of each query. The paper also identifies a neighbourhood size that is neither too small nor too large: too small and the estimate is noisy, too large and the local simplicity breaks down. Ecological coupling is a natural instance. Greenness and water storage are related one way in semi-arid afforested loess and another way in humid forests, but the relationship is plausibly close to linear within a climate-and-terrain neighbourhood. Our claim is therefore about scaling and locality. It is not a claim that retrieval is more accurate than any regressor on a fixed, in-distribution split. The skeptic's section addresses this directly.
Relation to Prior Work
Prior retrieval in Earth observation stores the wrong kind of memory for this question. GeoEvolver keeps a bank of successful tool scripts and failure diagnoses for geospatial agents and reaches 76.56% on the Earth-Agent benchmark. Its memory holds procedures, not biophysical measurements, so it can tell whether a pipeline ran but not whether the assessment it produced hides a trade-off. RS-RAG retrieves Wikipedia text and aggregated scalar attributes for 14,820 landmarks to support descriptive question answering, and it reports that simple tasks degrade when more than one item is retrieved. We retrieve dense multi-indicator records for continuous regression and study retrieval depth explicitly. ChangeTitans uses an internal recurrent memory to carry context across image patches for building change detection. It has no external store and makes no continuous ecological estimates.
Two papers constrain our design. The REO line shows that general vision-language models give negative $R^2$ on continuous above-ground biomass (Towards Unified Vision Language Models for Forest Ecological Analysis). A dedicated regression head lifts this only to $R^2 = 0.36$ (REO-VLM). We therefore avoid token-generating models entirely and regress numbers directly.
Two papers supply the method. Spatial contrastive learning for ecology provides a frozen ResNet-18 encoder trained on spatial smoothness, which serves as our similarity function. The theory of Basu et al. supplies the local fit and the neighbourhood-size trade-off. The motivation comes from the environmental literature: the Loess Plateau result above, carbon-optimal recycling networks that raise toxicity and acidification (Life Cycle Assessment of Industrial Symbiosis Model in Tianjin), and SustainGym's own admission that its benchmarks omit water and non-carbon pollutants.
Research Program
The memory is a table of site-period records. A site is a catchment or grid cell, and a period is a multi-year window. Each record's key is the frozen encoder's embedding of the site's satellite observables over the window: Sentinel-2 multispectral greenness, wetness and surface temperature, plus ALOS-2 PALSAR-2 radar where available, together with the trajectory of the headline metric itself. Each record's value is a vector of co-measured indicators, some of which may be missing. Water storage trend comes from GRACE-CSR. Biomass comes from REO-Instruct labels. River quality indicators come from the 109 Class-A Japanese watersheds compiled in the river environment ranking study. InVEST habitat quality comes from the Jiangsu land-use series (Spatiotemporal Dynamics and Prediction of Habitat Quality in Jiangsu).
For a query site, the system proceeds in three steps. First, it retrieves the $k$ nearest records that have the target indicator observed. Second, it fits a ridge regression from observables to that indicator on those neighbours only. Third, it reports a point estimate together with an interval derived from the spread of the neighbours' leave-one-out residuals. A trade-off flag is the estimated probability that the hidden indicator moves opposite to the headline metric, for example greenness rising while water storage falls. Missing values cost nothing, because each indicator is fitted only on neighbours that observed it. A new indicator column becomes usable the moment its records are added.
The analysis has three parts. The first traces scaling curves of error against memory size, holding the encoder fixed. The second sweeps neighbourhood size to locate the optimum predicted by Basu et al. in each biome. The third studies the resolution mismatch that threatens the whole approach: GRACE resolves only very large footprints, so water storage labels must be defined at basin scale and the observables aggregated to match. We will report how much skill survives at that scale rather than hide it.
Baselines
The naive reference point is the composite remote-sensing index itself. It assumes that if the Remote Sensing Ecological Index improved, nothing hidden got worse. On the Loess Plateau that index rose 43.45% while water storage declined, so its implied trade-off detection is, by construction, blind to this case. Every method must beat this zero-skill flag on the "green but drying" sites.
The strongest parametric competitor is a gradient-boosted or multilayer-perceptron multi-output regressor. It gets the same frozen embeddings, the same training records and the same hyperparameter search budget as the retrieval model. A second variant fine-tunes the encoder end to end at matched compute. For biomass we also report against REO-VLM, using its stated $R^2 = 0.36$ on multispectral above-ground biomass from REO-VLM. We have no reported result on Lacuna for estimating water storage trend from optical imagery, so that comparison is internal. Beating the parametric regressor means matching it within its seed-to-seed spread on random splits. It also means exceeding it on held-out regions after those regions' neighbours enter the memory but before any retraining. Finally, it means coming within that same spread of the parametric model when the parametric model is retrained on the enlarged data.
Two ablations isolate the key ingredient. The first replaces local ridge regression with a single global ridge fit over the whole memory, which removes the locality and tests whether local fitting earns its place. The second replaces embedding-space neighbours with purely geographic neighbours, which is essentially spatial interpolation. It tests whether retrieval finds ecological analogues beyond mere proximity.
Why not just…?
Why not just train a multi-output regressor on all the data? This is the real threat. On a fixed, in-distribution split, a well-tuned gradient-boosted model may match or beat retrieval, and we expect it might. The proposal stands or falls on regimes where the data change after training: new regions, new years, and new indicator columns. In those regimes the parametric model must be retrained while the memory only grows. Our insertion-versus-retraining experiment settles the question. If retraining is cheap and always wins, the honest conclusion is that memory buys only operational convenience.
Why not just interpolate the hidden indicator from nearby measured sites, as in kriging? Geographic interpolation fails exactly where audits matter. Heavily reforested catchments can sit next to unaltered ones and behave differently, as the Loess Plateau Area D result suggests. The geographic-neighbour ablation tests this directly. If geography does as well as embeddings, the encoder adds nothing and we will say so.
Why not just add water storage to the composite index? Doing so requires measuring water storage at the site, and those measurements are coarse, sparse or absent. The point is to estimate hidden dimensions where they are not measured. A richer index does not help at unmonitored sites.
Why not prompt a vision-language model with retrieved records, in the style of RS-RAG? Generating numbers as tokens gives negative $R^2$ on continuous biomass (Towards Unified Vision Language Models for Forest Ecological Analysis). RS-RAG also degraded beyond a single retrieved item on simple tasks. We nonetheless include it on the biomass task as a sanity check.
Proposed Experiments
All experiments use spatial block cross-validation that holds out whole regions. The headline test holds out the Loess Plateau, including its reforested Area D catchments. For that region we ask whether a memory built from the rest of the world predicts the sign of the water storage trend and flags the greenness-water trade-off. The metrics are sign accuracy and area under the ROC curve for the trade-off flag, $R^2$ for the continuous water storage trend, and interval coverage. Biomass on REO-Instruct and river quality on the Japanese watersheds are secondary targets that test transfer across indicator types.
The decisive scaling check fixes the encoder and grows the memory by a factor of ten, repeated across several random insertion orders. We expect error on the hidden indicator to fall monotonically. We also expect the optimal neighbourhood size to grow with memory size, since denser memories permit tighter neighbourhoods at the same variance. If error plateaus immediately, or if the global-ridge ablation matches local fitting, the local-regularity premise fails for ecological coupling and the proposal should be abandoned rather than tuned.
References
- Vegetation Change and Eco-Environmental Quality Evaluation in the Loess Plateau of China from 2000 to 2020
- Generalization through Memorization: Nearest Neighbor Language Models
- Generalization Properties of Retrieval-based Models
- Experience-Driven Multi-Agent Systems Are Training-free Context-aware Earth Observers
- Remote Sensing Retrieval-Augmented Generation: Bridging remote sensing imagery and comprehensive knowledge with a multimodal dataset and retrieval-augmented generation model
- Towards Remote Sensing Change Detection with Neural Memory
- Towards Unified Vision Language Models for Forest Ecological Analysis in Earth Observation
- REO-VLM: Transforming VLM to Meet Regression Challenges in Earth Observation
- Accelerating Ecological Sciences from Above: Spatial Contrastive Learning for Remote Sensing
- Life Cycle Assessment of Industrial Symbiosis Model in Tianjin, China: Case Study of Recycling System of Sewage Sludge and Construction Waste
- SustainGym: Reinforcement Learning Environments for Sustainable Energy Systems
- A RIVER ENVIRONMENT RELATIVE COMPARISON MODEL BASED ON THE RANKING USING MULTIPLE WATER QUALITY INDICATORS
- Spatiotemporal Dynamics and Prediction of Habitat Quality Based on Land Use and Cover Change in Jiangsu, China