SPyCer: Semi-Supervised Physics-Guided Contextual Attention for Near-Surface Air Temperature Estimation from Satellite Imagery

Sofiane Bouaziz; Adel Hafiane; Raphael Canals; Rachid Nedjai

cs.CVMar 5, 2026

SPyCer: Semi-Supervised Physics-Guided Contextual Attention for Near-Surface Air Temperature Estimation from Satellite Imagery

Authors: Sofiane Bouaziz, Adel Hafiane, Raphael Canals, Rachid Nedjai

Organizations: INSA CVL, Université d’Orléans, PRISME UR 4229, Bourges, 18022, Centre Val de Loire, France · Université d’Orléans, INSA CVL, PRISME UR 4229, Orléans, 45067, Centre Val de Loire, France · Université d’Orléans, CEDETE, UR 1210, Orléans, 45067, Centre Val de Loire, France

Abstract

Modern Earth observation relies on satellites to capture detailed surface properties. Yet, many phenomena that affect humans and ecosystems unfold in the atmosphere close to the surface. Near-ground sensors provide accurate measurements of certain environmental characteristics, such as near-surface air temperature (NSAT). However, they remain sparse and unevenly distributed, limiting their ability to provide continuous spatial measurements. To bridge this gap, we introduce SPyCer, a semi-supervised physics-guided network that can leverage pixel information and physical modeling to guide the learning process through meaningful physical properties. It is designed for continuous estimation of NSAT by proxy using satellite imagery. SPyCer frames NSAT prediction as a pixel-wise vision problem, where each near-ground sensor is projected onto satellite image coordinates and positioned at the center of a local image patch. The corresponding sensor pixel is supervised using both observed NSAT and physics-based constraints, while surrounding pixels contribute through physics-guided regularization derived from the surface energy balance and advection-diffusion-reaction partial differential equations. To capture the physical influence of neighboring pixels, SPyCer employs a multi-head attention guided by land cover characteristics and modulated with Gaussian distance weighting. Experiments on real-world datasets demonstrate that SPyCer produces spatially coherent and physically consistent NSAT estimates, outperforming existing baselines in terms of accuracy, generalization, and alignment with underlying physical processes.

Figures & tables

Figure 1: Illustration of the sensing gap between satellite observations and near-ground conditions. Satellites capture surface properties, while NSAT, measured 2 meters above the ground, drives human comfort and environmental processes. Near-ground sensors provide accurate but sparse measurements, leaving most areas unsampled. SPyCer leverages the energy exchange between land surface and near-surface atmosphere to estimate continuous, physically consistent NSAT from sparse sensors and satellite imagery.

Figure 2: Overview of SPyCer. For each near-ground sensor, a square patch of satellite-derived LST and auxiliary variables is extracted, representing the local spatial neighborhood. SPyCer predicts the NSAT at the central pixel while leveraging contextual information from neighboring pixels through a physics-informed semi-supervised strategy. Learnable spatial contextual weights quantify the physical relevance of each neighbor, which guide the total loss defined in Eq. 9 .

Figure 3: Representation of the SEB. Net radiation, primarily from solar and atmospheric inputs, is partitioned into three flux components: sensible heat flux ( $H$ ) transferring energy to the air, latent heat flux ( $LE$ ) driving evapotranspiration, and ground heat flux ( $G$ ) conducted into the soil.

Method	Metric	April	May	June	July	August	September	Average
LR	RMSE	3.66 $\pm$ 0.67	4.09 $\pm$ 0.47	3.56 $\pm$ 0.62	3.89 $\pm$ 0.64	3.61 $\pm$ 0.69	3.91 $\pm$ 0.66	3.77 $\pm$ 0.54
LR	MAE	2.95 $\pm$ 0.56	3.39 $\pm$ 0.42	2.99 $\pm$ 0.59	3.26 $\pm$ 0.55	3.01 $\pm$ 0.60	3.30 $\pm$ 0.66	3.11 $\pm$ 0.49
RF	RMSE	2.91 $\pm$ 0.60	2.92 $\pm$ 0.51	3.27 $\pm$ 0.67	3.52 $\pm$ 0.64	3.81 $\pm$ 0.80	3.70 $\pm$ 0.75	3.33 $\pm$ 0.54
RF	MAE	2.36 $\pm$ 0.55	2.41 $\pm$ 0.49	2.67 $\pm$ 0.59	2.96 $\pm$ 0.60	3.18 $\pm$ 0.73	3.17 $\pm$ 0.71	2.73 $\pm$ 0.49
GB	RMSE	2.93 $\pm$ 0.65	2.99 $\pm$ 0.68	3.19 $\pm$ 0.72	3.45 $\pm$ 0.72	3.69 $\pm$ 0.90	3.65 $\pm$ 0.77	3.29 $\pm$ 0.64
GB	MAE	2.38 $\pm$ 0.60	2.47 $\pm$ 0.58	2.62 $\pm$ 0.63	2.90 $\pm$ 0.65	3.08 $\pm$ 0.77	3.14 $\pm$ 0.73	2.70 $\pm$ 0.56

Table 1: Quantitative comparison of SPyCer against baseline methods (LR [ 72 , 42 , 65 ] , RF [ 49 , 76 , 27 ] , GB [ 62 , 38 ] , MLP [ 53 ] ). Metrics (RMSE and MAE) are reported per month over 100 Monte Carlo runs. Values indicate mean ± standard deviation. The blue, yellow, and orange colors represent the top three results in order.

Figure 4: Qualitative comparison of NSAT estimates on 08 August 2025. Columns show IDW, GB, MLP, and SPyCer predictions, alongside the high-resolution satellite image for context. Rows correspond to (a) the full region, (b) a semi-urban corridor with a river, and (c) a major industrial zone. SPyCer accurately captures large-scale temperature patterns while resolving fine-scale variability, including cold regions, river features, and industrial hotspots, outperforming GB and MLP in spatial fidelity and local detail.

Figure 5: Temporal evolution of NSAT estimates from SPyCer, MLP, and GB compared to ground-truth measurements at two randomly selected near-ground sensors from April to October 2025. SPyCer closely follows both the amplitude and timing of temperature variations, whereas MLP and GB overestimate during warm periods and fail to capture fine-scale fluctuations.

Figure 6: Spatial attention maps learned by SPyCer for nine randomly selected sensors. For each near-ground sensor, the left panel shows the local RGB satellite context (yellow dot marks the sensor location), and the right panel shows the corresponding $7\times 7$ neighborhood attention map weights, where warmer colors indicate higher influence on the central NSAT prediction. SPyCer consistently emphasizes neighboring pixels with similar land-cover and thermal characteristics.

Configuration	RMSE	MAE
Config. 1	2.80 $\pm$ 0.28	2.26 $\pm$ 0.23
Config. 2	2.57 $\pm$ 0.27	2.08 $\pm$ 0.23
SPyCer	2.27 $\pm$ 0.07	1.83 $\pm$ 0.07

Table 2: Quantitative comparison of different configurations for NSAT estimation. Metrics (RMSE and MAE) are reported as mean ± standard deviation.

Remote Sensing

Remote sensing uses satellite, aerial, and ground-based observations to measure and map Earth and other planetary surfaces.

See all 528 papers

Sparse-Observation Atmospheric Thermal Forecasting with Physics-Informed Neural Networks for Climate-Aware Digital Twins
Sep 23, 2026Tannaz Goodarzvand Chegini, Elyas Shivanian, Behzad Karimi +1Weather Forecasting Time Series Forecasting
Department of Mathematical Sciences / Montana State University, Bozeman, Montana, US · Department of Mathematics / Imam Khomeini International University, Qazvin, Iran · Department of Electrical and Computer Engineering, Montana State University, Bozeman, Montana, US +1
Partial recovery of meter-scale surface weather
Feb 26, 2026Jonathan Giezendanner, Qidong Yang, Ruizhe Huang +7Weather Forecasting
Massachusetts Institute of Technology · Shell Information Technology · Allen Institute for AI +1
Spatially Aware Deep Learning for Microclimate Prediction from High-Resolution Geospatial Imagery
Feb 27, 2026Idan Sulami, Alon Itzkovitch, Michael R. Kearney +2Spatial Context Modeling Climate Modeling
Tel Aviv University, Faculty of Life Sciences, School of Zoology, Tel Aviv, Israel · School of BioSciences, The University of Melbourne, Melbourne, Victoria, Australia · Tel Aviv University, Center for Artificial Intelligence & Data Science, Tel Aviv, Israel

Sep 23, 2026J/K move · Enter open · S save

Sparse-Observation Atmospheric Thermal Forecasting with Physics-Informed Neural Networks for Climate-Aware Digital Twins

Tannaz Goodarzvand Chegini, Elyas Shivanian, Behzad Karimi, Faraz Dadgostari

Department of Mathematical Sciences / Montana State University, Bozeman, Montana, US · Department of Mathematics / Imam Khomeini International University, Qazvin, Iran · Department of Electrical and Computer Engineering, Montana State University, Bozeman, Montana, US+1

Weather Forecasting Time Series Forecasting Physics-Informed ML

Open paper AskPDF arXiv

Short-horizon forecasts of atmospheric temperature are needed to support climate-aware digital-twin systems, but such forecasts must be produced where thermal observations are incomplete. This study evaluates a physics-informed neural network for potential-temperature forecasting, constrained by a pressure-coordinate thermodynamic advection-source equation and a diabatic-source closure fit from the preceding 12-hour period and frozen before future-time training. Using hourly ERA5 reanalysis at three pressure levels, the model is evaluated as a conditional hindcast at lead times of one, two and three hours against persistence, local-trend, and two matched neural-network baselines, one of which receives the same future meteorological forcing as the PINN, helping distinguish the physical constraint from access to future forcing. In an Oklahoma development case, mean RMSE improvement over the strongest baseline grew from 8.1% at one hour to 23.8% at three hours; under an observation-density sweep down to 5% of candidate locations, this 3-hour advantage remained 14.6--16.9%, with no evidence that lower density improves performance. Under a fixed protocol transferred to an Alabama heat event with three virtual-observation layouts, three-hour improvement ranged 19.7-24.4% with consistent origin-level wins. A parallel Montana stress test, in which fixed pressure levels intersected complex terrain, produced a three-hour degradation of roughly 17.5%, identifying a terrain-related applicability limit of the formulation. Together, these results indicate that the physics constraint's benefit grows with forecast horizon, persists under severe observation sparsity, and transfers across regions, but is bounded by the validity of a fixed vertical-coordinate representation over complex terrain, evidence relevant to physics-constrained components of climate-aware forecasting and digital-twin systems.

Abstract

Remote Sensing

Explore similar work

Sparse-Observation Atmospheric Thermal Forecasting with Physics-Informed Neural Networks for Climate-Aware Digital Twins

Partial recovery of meter-scale surface weather

Spatially Aware Deep Learning for Microclimate Prediction from High-Resolution Geospatial Imagery