Land-Centric Modeling Methods

• Land-Centric Models are a paradigm that explicitly represents land systems—like parcels, patches, and biogeographical domains—to capture spatial structure and heterogeneity.
• They use techniques such as rotation-aware training, domain generalization, and local mapping to address challenges from continental, seasonal, and sensor-induced variations.
• Applications include parcel-based label denoising, generative landscape simulation, and process-based land-surface modeling, merging machine learning with physical science insights.

Searching arXiv for recent and foundational papers on "Land-Centric Model" and related land-cover / land-surface modeling usage. A land-centric model is a modeling paradigm in which the primary unit of analysis is not an abstract feature space alone, but the land system itself: land parcels, land-cover patches, continent-by-season domains, static geophysical attributes, ecological parameter fields, or physically interpretable land-surface embeddings. In current arXiv usage, the term spans geospatial land-cover mapping, parcel-based label denoising, local teacher–student mapping, autoregressive landscape generation, foundation-model embeddings tied to environmental variables, differentiable terrestrial biosphere modeling, and kilometer-scale land-surface simulation (Hu et al., 2020, Pekhale et al., 2024, Tadesse et al., 2024, Krapu et al., 2024, Rahman, 10 Feb 2026, Fang et al., 2024, Wang et al., 19 Jan 2025). Across these variants, the common premise is that land organization, land heterogeneity, and land-surface process structure must be represented explicitly rather than treated as incidental context.

1. Conceptual scope and main formulations

The phrase does not denote a single canonical architecture. Instead, it denotes a family of approaches that center land-relevant structure at the level most appropriate to the problem. In land-cover mapping, that structure may be continent, season, orientation distribution, and patch texture; in label denoising, it may be parcel-like regions inferred from imagery; in land-surface science, it may be ecological parameters conditioned on climate, soils, and demography; in foundation models, it may be embeddings whose dimensions map onto precipitation, temperature, vegetation, hydrology, and terrain (Hu et al., 2020, Pekhale et al., 2024, Fang et al., 2024, Rahman, 10 Feb 2026).

Usage	Central land object	Representative paper
Geospatial classification	Continent×season domain and landscape orientation	(Hu et al., 2020)
Parcel-based denoising	SAM-derived land parcels	(Pekhale et al., 2024)
Local LULC mapping	Geography-specific teacher–student training	(Tadesse et al., 2024)
Generative landscape modeling	Spatial grammar of patches and corridors	(Krapu et al., 2024)
Interpretable embeddings	64-D land-surface representation	(Rahman, 10 Feb 2026)
Hybrid land-surface science	Environment-conditioned ecological parameters	(Fang et al., 2024)

This breadth has methodological consequences. A land-centric model is usually evaluated not only by average predictive skill but also by how well it preserves domain transfer, patch geometry, physical plausibility, spatial structure, or environmental interpretability. That emphasis distinguishes it from purely pixel-centric or globally uniform approaches.

2. Geospatial generalization and the land-cover mapping problem

In geospatial remote sensing, the land-centric problem is sharply exposed by model transfer. "Model Generalization in Deep Learning Applications for Land Cover Mapping" shows that when deep models are trained on data from specific continents or seasons, performance varies substantially on out-of-sample continents or seasons, and that rotation biases in geospatial datasets constitute a concrete failure mode (Hu et al., 2020). The central implication is that high in-domain performance does not guarantee reliable land-cover prediction when landscapes recur with different dominant orientations, textures, acquisition geometries, or seasonal states.

The relevant domain shift is multi-axis. The paper’s framing highlights orientation bias, but the associated difficulty also includes spectral-signature shifts across seasons, continent-specific differences in landscape heterogeneity and scale, and acquisition effects such as solar/view geometry and atmospheric state (Hu et al., 2020). This suggests that a land-centric model for land-cover mapping must be designed around cross-continent and cross-season transfer rather than around a single pooled benchmark.

Methodologically, this has favored rotation-aware and domain-aware learning. The data summary associated with the paper identifies standard patch classifiers and segmentation networks such as ResNet, EfficientNet, U-Net, DeepLabv3+, and HRNet, but emphasizes that vanilla CNNs are translation-equivariant rather than rotation-equivariant; therefore group-equivariant CNNs, steerable CNNs, harmonic networks, and domain generalization methods such as IRM, GroupDRO, CORAL, and MMD-regularized training are natural countermeasures (Hu et al., 2020). The recommended empirical protocol is likewise land-centric: split data by continent and season, enforce strict spatial separation, and report per-domain and per-class metrics rather than only pooled averages.

A related extension appears in "Geographical Context Matters: Bridging Fine and Coarse Spatial Information to Enhance Continental Land Cover Mapping" (Ghassemi et al., 16 Apr 2025). Although the detailed manuscript was unavailable in the supplied record, the abstract states that BRIDGE-LC jointly uses fine-grained latitude/longitude and coarse-grained biogeographical region information during training, then requires only fine-grained information at inference. This places geospatial land context, rather than imagery alone, at the center of continental-scale classification (Ghassemi et al., 16 Apr 2025).

3. Parcel-centric and local land-use/land-cover workflows

A second major strand defines land-centricity at the parcel or local geography level. "SAModified: A Foundation Model-Based Zero-Shot Approach for Refining Noisy Land-Use Land-Cover Maps" defines a land-centric model as one centered on land parcels or regions rather than individual pixels (Pekhale et al., 2024). The method is a two-stage zero-shot pipeline: Segment Anything Model (SAM) first delineates segments interpreted as candidate land parcels, after which local label statistics within each segment relabel uncertain pixels. If $P_i$ is the set of pixels in segment $s_i$, the parcelwise label distribution is

$$p(l \mid s_i) = \frac{1}{|P_i|}\sum_{j=1}^{m_i}\mathbf{1}[l_j=l],$$

and uncertain pixels labeled as “mosaic of uses” are reassigned by majority vote:

$y'_j = \arg\max_l p(l \mid s_i).$

Using Harmonized Landsat and Sentinel-2 imagery and MapBiomas labels over Brazil, the study samples 112,092 Areas of Interest, each about $235.93\ \text{km}^2$, and reports an approximately $5\%$ improvement in downstream UNet segmentation performance when trained on denoised labels (Pekhale et al., 2024).

The same land-centric intuition motivates local rather than global training. "Local vs. Global: Local Land-Use and Land-Cover Models Deliver Higher Quality Maps" defines a land-centric model as a local, data-centric approach trained and validated on data representative of a specific geography and its land systems (Tadesse et al., 2024). The DATS framework uses a high-resolution teacher trained on Maxar RGB imagery at $0.331\ \text{m/pixel}$ over 51.55% of Murang’a County, Kenya, and a Sentinel-2 student at $10\ \text{m/pixel}$ trained on public imagery plus teacher pseudo-labels. The teacher uses 11,382 expert-annotated polygons, 14,577 Building polygons and 6,910 Road polygons from OpenStreetMap, plus hard negatives, for a total of 51,618 polygons covering 5.69% of pixels in the Maxar footprint; the student uses 1,892,036 weakly labeled polygons (Tadesse et al., 2024).

The empirical outcome is explicitly comparative: local models improve macro F1 by 0.14 and IoU by 0.21 relative to the best global baseline, while existing global maps in Murang’a have a maximum inter-map agreement of only 0.30 (Tadesse et al., 2024). The paper also reports that DATS better captures built-up linear features such as roads and settlements following road corridors. This indicates that, in smallholder mosaics and other heterogeneous landscapes, locality itself can function as a key inductive bias.

4. Generative, embedding-based, and foundation-model interpretations

A third strand treats land-centricity as representation learning over landscape structure. "Deep autoregressive modeling for land use land cover" formulates LULC modeling as a conditional generative problem analogous to image inpainting, with a modified PixelCNN of approximately 19 million parameters learning

$$p(\mathbf{Y}\mid \mathbf{X}) = \prod_{i=1}^{N} p(y_i \mid y_{<i}, \mathbf{X}),$$

where $\mathbf{X}$ contains land-centric covariates such as elevation, slope, flow accumulation, distance to roads, and distance to water (Krapu et al., 2024). The objective is to reproduce not only categorical accuracy but also the “spatial grammar” of landscapes: roads, water bodies, adjacency frequencies, patch counts, and corridor structure. The study finds that the model captures richer spatial correlation patterns than a benchmark spatial statistical model, but produces underdispersed predictive distributions; temperature scaling and logit noise injection partially ameliorate the problem (Krapu et al., 2024).

Foundation models extend this idea beyond autoregression. "Physically Interpretable AlphaEarth Foundation Model Embeddings Enable LLM-Based Land Surface Intelligence" uses 12.1 million annual samples across CONUS from 2017–2023 to analyze Google AlphaEarth’s 64-dimensional embeddings against 26 environmental variables spanning climate, vegetation, hydrology, temperature, terrain, and urban indicators (Rahman, 10 Feb 2026). The study reports that 12 of 26 variables exceed $s_i$0 under Random Forest reconstruction, with temperature and elevation approaching $s_i$1, and that a multi-task Transformer raises 14 variables above $s_i$2; spatial block cross-validation yields a mean $s_i$3, and inter-year interpretability profiles have mean correlation $s_i$4 (Rahman, 10 Feb 2026). This is an explicitly land-centric embedding because individual dimensions are linked to physical properties such as precipitation, daytime land-surface temperature, vegetation structure, hydrological cycling, and tree cover.

"StefaLand: An Efficient Geoscience Foundation Model That Improves Dynamic Land-Surface Predictions" adopts a different foundation-model formulation centered on landscape attributes rather than Earth observation pixels (Kraabel et al., 22 Sep 2025). It fuses dynamic forcings and static attributes into a masked autoencoder backbone with a learnable static token, then injects frozen pretrained representations into downstream sequence models through residual adapters. The underlying Transformer has roughly 12 million parameters and is pretrained in approximately 720 GPU hours, a compute profile the paper contrasts with much larger EO vision foundation models (Kraabel et al., 22 Sep 2025). On CAMELS streamflow, StefaLand-resConn improves RMSE from 1.402 to 1.111 and NSE from 0.636 to 0.717 in PUB, and from 1.609 to 1.344 and 0.554 to 0.635 in PUR; on Europe-holdout soil moisture, RMSE improves from 0.112 to 0.090 and correlation from 0.510 to 0.638 (Kraabel et al., 22 Sep 2025). The model’s land-centric claim is therefore not only architectural but also operational: its tokenization scheme privileges terrain, soils, geology, vegetation, and meteorological forcings over imagery.

At a global reconstruction scale, "AI4Land: Scalable Deep Learning for Global High-Resolution Land Use Reconstruction" defines a land-centric, GPU-accelerated U-Net framework that fuses coarse LUH2 scenario inputs, static geophysical features, and HILDA+ supervision to reconstruct annual global LULC at 1 km from 1850 to 2100 (Mozaffari et al., 10 Jun 2026). Under a 60% masking configuration, it reports global accuracy of 94.67% and $s_i$5; for 2014 global inference, accuracy is 94.88% and mIoU is 0.8569, with class IoU values including Forest 0.889, Water 0.991, and Urban 0.463 (Mozaffari et al., 10 Jun 2026). This suggests that land-centricity can also be formulated as physically informed coarse-to-fine land-surface emulation for Earth system applications.

5. Land-surface science, ecohydrology, and process-based formulations

In land-surface science, the term shifts from mapping to process representation. "Differentiable Land Model Reveals Global Environmental Controls on Ecological Parameters" presents DifferLand, a hybrid physics–ML system coupling a spatialization neural network to a differentiable DALEC terrestrial biosphere model (Fang et al., 2024). The network takes PFT fractions, climate, forest demography, and soils, uses three hidden layers of 32 ReLU neurons, and outputs 40 quantities: 31 DALEC ecological parameters, 2 phenology parameters, and 7 initial states (Fang et al., 2024). A principal result is that Plant Functional Types explain less than half of the parameter-explainable spatial variation: PFTs contribute only 24–36%, while local climate and forest demography add comparable or larger predictive power. PCA of 13 robust parameters identifies three near-orthogonal gradients explaining 83.3% of spatial variability: growing season length, leaf economics, and agricultural intensity (Fang et al., 2024). On held-out pixels, the model attains $s_i$6 of 0.90 for LAI, 0.81 for SIF, 0.86 for VOD, 0.66 for ET, and 0.88 for biomass; at coarse scales, NBE reaches 0.79 and EWT 0.58 (Fang et al., 2024).

"Ecohydrological land reanalysis" defines ECHLA as a land-centric reanalysis because it sequentially assimilates microwave brightness temperatures into a land surface model that explicitly simulates vegetation biomass dynamics (Sawada et al., 2021). CLVDAS with EcoHydro-SiB uses a 20-layer soil column down to 1.95 m, daily outputs on a $s_i$7 grid, and a Particle Filter with $s_i$8 particles to assimilate AMSR-E/2 6.925 and 10.25 GHz H/V night observations over 2003–2010 and 2013–2019 (Sawada et al., 2021). For assimilated brightness temperatures, the reanalysis reports global MAE below 2 K, ubRMSE below 6 K, and correlation above 0.6; data assimilation improves the seasonal cycle of vegetation and shallow soil moisture skill, especially in the top 0–0.15 m (Sawada et al., 2021).

A fully process-based, high-resolution formulation appears in "Kilometer-Scale E3SM Land Model Simulation over North America" (Wang et al., 19 Jan 2025). The km-scale ELM simulates 21,624,900 land grid cells over approximately 21.5 million km² of North America at 1 km resolution, with up to 100,800 CPU cores across 2,400 nodes (Wang et al., 19 Jan 2025). Strong scaling raises land SYPD from 1.26 to 11.60 as execution time drops from 939.447 s to 102.042 s, with parallel efficiency of approximately 87% at 50,400 cores and 58% at 100,800 cores (Wang et al., 19 Jan 2025). Here the model is land-centric because it resolves surface heterogeneity, land-unit structure, soil columns, and PFT composition at scales at which topography, soils, snow, vegetation, and hydrologic partitioning directly shape land–atmosphere exchange.

A related process-based formulation is the updated CABLE land surface model, which couples a CMIP6-ready land-use and land-cover change module to the POP woody demography module and implements an optimization-based approach to plant coordination of photosynthesis (Haverd et al., 2017). In this sense, land-centricity denotes explicit accounting of land-use transitions, secondary forest age structure, disturbance, and land-atmosphere carbon, water, and energy exchange.

6. Limitations, disputes, and emerging directions

Despite their diversity, land-centric models share recurrent limitations. In geospatial classification, rotation bias is only one axis of domain shift; phenology, sensor differences, atmospheric effects, and landscape heterogeneity remain major sources of degradation (Hu et al., 2020). Parcel-based methods depend on the quality of region delineation: SAModified notes failure modes in homogeneous areas, small or fragmented parcels, and mixed-use edges, where majority voting may overwrite minority classes or reduce rare-class fidelity (Pekhale et al., 2024). Local mapping frameworks improve accuracy but can depend on expensive high-resolution imagery and remain vulnerable to class ambiguity, cloud contamination, and single-date supervision (Tadesse et al., 2024).

Representation-learning approaches have their own constraints. The PixelCNN study finds predictive underdispersion and imperfect calibration in generated landscapes (Krapu et al., 2024). AlphaEarth embedding analysis is restricted to CONUS, and urban indicators and fine-scale geomorphology are reconstructed more weakly than temperature, terrain, or vegetation (Rahman, 10 Feb 2026). StefaLand reports its strongest evidence in settings with dense observations and notes limits from a relatively narrow pretraining corpus and from concept drift across soil datasets (Kraabel et al., 22 Sep 2025). AI4Land identifies class imbalance and domain shift as continuing difficulties, with Urban the weakest reported class in the 2014 global inference summary (Mozaffari et al., 10 Jun 2026).

Process-based land-centric models also face structural uncertainty. DifferLand treats environmental predictors as temporally invariant over 2010–2023 and leaves nutrient cycling, detailed hydraulics, and multi-decadal pools only partially constrained (Fang et al., 2024). ECHLA excludes snow-affected pixels in its present release and inherits sensitivity to radio-frequency interference and observation-operator parameterization (Sawada et al., 2021). Kilometer-scale ELM demonstrates continental feasibility but does not yet provide the comprehensive model–data validation that would establish physical skill at comparable resolution (Wang et al., 19 Jan 2025).

Taken together, these works indicate that land-centric modeling is best understood as a unifying research orientation: it prioritizes spatial organization, land heterogeneity, and land-process structure over generic feature learning. A plausible implication is that future progress will come from combining the strengths already visible in the literature: continent×season-aware evaluation, parcel-level spatial coherence, local data curation, physically interpretable embeddings, hybrid physics–ML parameterization, and scalable high-resolution simulation (Hu et al., 2020, Pekhale et al., 2024, Rahman, 10 Feb 2026, Fang et al., 2024, Wang et al., 19 Jan 2025).