{"id":"2c28690b-5d64-4b84-93fc-d6db14d568c8","arxiv_id":"2507.09872","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An additive model of annual temperature cycle, linear ERA5 amplification, and a U-Net residual reconstructs all-weather daily land surface temperature at 30 m with 2-3 K accuracy.","lead":"A deep learning framework blends a seasonal temperature curve, a scaled copy of coarse weather-model temperatures, and a neural network to produce daily 30-meter land surface temperatures from sparse satellite images. It reports errors around 2-3 degrees Kelvin against ground stations and held-out data, covering days when clouds hide the surface.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Time-invariant residual is the soft spot: M_resid(X) in Eq. (10) has no time or cloud-state input, so all day-to-day variation rests on the per-pixel ERA5 scaling in Eq. (8); the all-weather claim needs a test of time-dependent residuals.","rationale":"The reader's weakest assumption correctly identifies the time-invariant residual as the central risk. Eqs. (6), (8), and (10) show that the only time-dependent data-driven term is the linear ERA5 amplification, and Eq. (11) trains on observed clear-sky pixels. The paper's description of M_resid as a 'spatiotemporal correction' is stronger than what the architecture delivers, since the CNN receives no time input. This is an internal gap between the model formulation and the all-weather claim, not a disagreement with external consensus. The validation evidence is otherwise credible: in situ comparisons across four datasets, hold-out 20% for GOES-16, and ablations (ATC, ATC+ERA5, naive CNN) all support the value of the ERA5 linear term and the static correction on observed days. However, the cloudy-day component of the claim is not directly tested because the metrics are not split by cloud state. A time-dependent residual test would settle whether the static correction is sufficient. Since the reader already conditioned acceptance on clarifying exactly this issue, the verdict remains CONDITIONAL, and no further adjustment is needed.","tokens_in":12438,"tokens_out":3931,"duration_ms":46362,"concrete_test":"Add the date (or, better, cloud fraction from GOES/ERA5) as an additional input channel to the U-Net, making M_resid depend on time and cloud state, and retrain under the same protocol. Then compare, on the PSU and BON in situ sets, the cloudy-day MAE/RMSE of the time-invariant model versus the time-dependent model, where cloudy days are defined by the satellite cloud mask. If the time-dependent residual gives no significant improvement on cloudy days, the static-residual concern is not load-bearing; if it improves cloudy-day MAE by more than about 0.5 K, the original all-weather claim is unsupported without that term. Also report the same clear/cloudy split for ATC+ERA5 versus the proposed model to identify where the CNN contribution actually occurs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that a time-invariant CNN residual, M_resid(X) in Eqs. (6) and (10), can serve as the 'spatiotemporal correction' underpinning the all-weather reconstruction claim. The U-Net input is a single annual-mean reflectance image, with no time argument and no cloud-state or cloud-fraction input. Consequently, all day-to-day variability beyond the annual cosine is carried by the pixelwise linear term w_ij * Tc(i,j,t) in Eq. (8). Because the loss in Eq. (11) is evaluated only on observed clear-sky pixels, both w_ij and the static residual are fitted under clear-sky conditions and cannot adapt to cloudy conditions by construction. The paper's all-weather claim therefore presumes that the relationship between fine-scale LST and coarse ERA5 temperature is stationary across clear/cloudy states and seasons. The in situ results do not isolate this assumption: they aggregate all days, and on cloudy days the reconstruction is essentially ERA5's temporal shape plus a static spatial correction. This is a real soft spot because 'all-weather 30 m reconstruction' is exactly what the static residual cannot, as formulated, guarantee: the model has no mechanism to learn a different residual under clouds.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-guided deep learning framework for reconstructing daily, all-weather land surface temperature at high spatial resolution (30 m for Landsat, 2 km for GOES-16). The model combines three additive components: a pixel-wise annual temperature cycle (ATC), a pixel-wise linear scaling of ERA5 reanalysis temperature to capture daily fluctuations, and a U-Net-based CNN that maps surface reflectance to a spatiotemporal residual field. Training uses an L1 loss evaluated only on observed (cloud-free) pixels. The method is evaluated on four datasets (two Landsat sites and two GOES-16 regions) with hold-out validation for GOES-16 and in situ validation against SURFRAD stations for all datasets, reporting 2–3 K accuracy and better agreement with in situ data than cloud-free satellite observations. The central claim is that this is the first end-to-end, single deep learning model for seamless all-weather spatiotemporal temperature reconstruction.","tokens_in":12740,"tokens_out":4338,"duration_ms":47990,"significance":"If the all-weather claim holds, the framework addresses a real and important gap: the trade-off between spatial and temporal resolution in satellite thermal remote sensing, with direct applications in climate, health, and urban studies. The architectural decomposition into physically interpretable components (ATC, ERA5 linear term, residual CNN) is a sensible way to inject physical structure while retaining end-to-end trainability, and the paper is generally clearly written. The evaluation is substantial: four datasets, two sensors, hold-out and in situ validation, and quantitative comparison with two existing methods. A notable strength is that the proposed method achieves lower RMSE against in situ measurements than the actual cloud-free satellite observations, an interesting and non-obvious result. However, the central all-weather contribution hinges on the assumption that a time-invariant, reflectance-only residual can correct cloudy-day temperatures, an assumption that is not directly tested; the headline 'first end-to-end' novelty claim also needs more support.","major_comments":[{"comment":"The residual term M_resid in Eq. (10) is a function only of the surface reflectance tensor X and has no time or cloud-cover input; the U-Net in Fig. 1 receives a single annual-mean reflectance image. All day-to-day variability beyond the annual cosine in Eq. (7) is therefore carried by the pixel-wise linear term w_ij * Tc(i,j,t) in Eq. (8). Because the loss in Eq. (11) is evaluated only at observed clear-sky pixels, both w_ij and the static residual are fitted under clear-sky conditions and cannot, by construction, adapt to cloudy conditions. The in situ evaluations in Tables 1 and 2 and Fig. 4 aggregate all days, so they do not isolate a cloudy-day test; on cloudy days the reconstruction is essentially the ERA5 temporal shape plus a static spatial correction. The all-weather claim therefore presumes that the relationship between fine-scale LST and coarse ERA5 temperature is stationary across cloud states and seasons, which is not demonstrated. I recommend reporting in situ error stratified by cloud cover, or adding a time/cloud-state input to the residual branch, or explicitly limiting the all-weather claim to the demonstrated clear-sky-corrected ERA5 extrapolation.","section":"§3.3, Eq. (10), Eq. (8), Eq. (11), Fig. 1"},{"comment":"The comparison with prior methods is not apples-to-apples: the proposed method is evaluated on 362/359 days at PSU/BON, whereas the 4-step ATC competitor is evaluated on 48/20 days and GEC-SEB on 229/222 days. The reported RMSE improvements therefore conflate method performance with test-set composition. To support the claim of state-of-the-art reconstruction, the authors should report metrics on the intersection of days available to all methods, or clearly state that the competitors cannot produce daily all-weather outputs and compare only on the day sets each method supports. Without this, the accuracy advantage is not yet established.","section":"Tables 1 and 2"},{"comment":"The conclusion claims 'achieving, for the first time, end-to-end temperature reconstruction based on a single deep learning model.' This is overstated as written: the model is a hybrid of an analytically parameterized ATC, a linear ERA5 scaling, and a CNN residual, with the ATC parameters obtained via optimization rather than through end-to-end backpropagation as part of a single differentiable model. No details are given on how gradients flow into the ATC branch. Please clarify what 'end-to-end' means in this architecture and temper the novelty claim, or provide evidence that prior two-step approaches cannot be trained jointly in the same manner.","section":"Conclusion, §1"}],"minor_comments":[{"comment":"The text 'As we are optimizing the F1 loss (MAE)' should read 'L1 loss'; F1 is a classification metric and this is a typo.","section":"§4.3, Eq. (11)"},{"comment":"The years of the two GOES-16 datasets are inconsistent: the first dataset is stated as 2022, while the second dataset is not given a year; Fig. 4 shows N=364, which suggests 2023. Please clarify the years for both GOES-16 datasets.","section":"§4.1"},{"comment":"The competitor rows list RMSE and bias but not MAE; since MAE is the primary metric used elsewhere, please report MAE for the competing methods as well, or explain its omission.","section":"Tables 1 and 2"},{"comment":"The 'Valid Cloud-free Observations' plots are compared with the reconstruction, but the sample sizes differ (e.g., N=145 vs N=364 at PSU); a fairer comparison would restrict both to the same days, or report the comparison over the common subset.","section":"Fig. 4"},{"comment":"The prior work of Liu et al. [22] is described as achieving 1.48–2.11 K accuracy for daily 30 m reconstruction, while the proposed method achieves 2.75–3.43 K MAE on the Landsat sites; the discussion should directly address how the proposed method improves on that baseline beyond the architectural differences.","section":"§2.2, reference [22]"}],"recommendation":"major_revision","confidential_remarks":"The paper's strongest contribution is the clean decomposition of a physics-guided reconstruction into interpretable components, but the all-weather claim needs a direct test of cloudy-day behavior. The time-invariant residual is the main soft spot: without a time or cloud-state input, the model has no mechanism to learn different corrections under clouds. The comparison with prior methods on different day sets is a second issue that should be addressed before publication. The 'first end-to-end' novelty claim may also draw criticism from readers familiar with two-step approaches; I suggest the authors soften it or provide a more precise definition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, well-executed integration of known pieces into a genuinely new end-to-end trainable model for daily 30 m LST reconstruction. The main weakness is that the all-weather claim outruns the architecture: the CNN residual is time-invariant, so cloudy-day fields are essentially ERA5's temporal shape plus a static spatial correction.\n\nWhat's new: the three-part additive architecture (ATC + pixelwise linear ERA5 scaling + U-Net residual) jointly trained is not present in the cited two-step methods. That's a real contribution. The evaluation is solid: four datasets, two sensors, hold-out and in situ validation. The ablation (ATC, ATC+ERA5, proposed) shows the residual term reduces bias. The reported 2-3 K errors are consistent with sensor noise, so the value is coverage and consistency rather than a step-change in accuracy. The paper is clearly written and the method is reproducible in principle.\n\nSoft spots: Eq. (10) defines M_resid as a function only of surface reflectance X, with no time or cloud-state input. All day-to-day variability beyond the annual cosine is carried by the per-pixel ERA5 scaling. Because the loss is computed only on clear-sky observed pixels, both the scaling and the static residual are fitted under clear skies. Applying that to cloudy days presumes stationarity across cloud states; the in situ evaluation aggregates all days and doesn't isolate this. This doesn't sink the paper, but the \"all-weather\" claim needs supporting evidence or a time-dependent residual. Also, the comparison with prior methods uses different day sets (48/229 vs 362 days); the full-year ATC baselines are fairer, but the headline numbers for 4-step ATC and GEC-SEB are not apples-to-apples. Minor: no uncertainty estimates despite a related paper on deep ensembles.\n\nBottom line: worth a serious referee. The core idea is sound, the execution is careful, and the issues are addressable with additional experiments. I'd engage with it.","headline":"Credible integration of known components into an end-to-end model, but the all-weather claim rests on a time-invariant residual that the paper never tests under cloudy conditions.","tokens_in":13247,"tokens_out":3565,"would_cite":true,"duration_ms":39450,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a single physics-guided convolutional network can reconstruct daily, all-weather land surface temperature at 30 m and 2 km with 2–3 K accuracy.","keywords":["land surface temperature","temperature reconstruction","physics-guided deep learning","annual temperature cycle","ERA5 reanalysis","convolutional neural network","all-weather LST","GOES-16"],"falsifier":"Choose a cloud-prone area with several in situ stations, train on clear-sky Landsat or GOES-16 pairs, and compare reconstructed cloudy-day temperatures against stations on days when the coarse ERA5 field places a temperature gradient opposite to the fine-scale pattern; if the model's error jumps on those days while clear-sky error stays low, the time-invariant reflectance residual is the failing component.","tokens_in":12228,"feed_emoji":"🌡️","tokens_out":5705,"duration_ms":65402,"temperature":0.7,"pith_summary":"The paper tries to establish that one end-to-end deep learning model, not a hand-crafted multi-stage pipeline, can produce seamless daily land surface temperature at very high spatial resolution. The model combines three additive components: a pixel-wise annual temperature cycle, a linear scaling of coarse ERA5 reanalysis temperature to capture daily fluctuations, and a U-Net-style convolutional network that learns residual spatial structure from surface reflectance. This design keeps the physically meaningful temporal behavior explicit while letting the neural network only fill in what the physics terms miss. If true, it closes the gap between high-resolution but cloud-gapped satellite observations and coarse but continuous reanalysis data, which would benefit climate-health studies, urban heat analysis, and ecological monitoring.","feed_headline":"One physics-guided CNN delivers daily 30 m land temperature","feed_subtitle":"Fusing satellite snapshots and ERA5 reanalysis, the model reaches 2–3 K accuracy and even beats cloud-free observations.","key_machinery":"The load-bearing mechanism is the additive decomposition of Eq. (6): reconstructed temperature equals a pixel-wise annual temperature cycle (a cosine with learnable mean, amplitude, and phase) plus a per-pixel linear scaling of coarse ERA5 temperature plus a residual field learned by a U-Net-style CNN from surface reflectance. The ATC and ERA5 branches supply physically plausible seasonal and daily variability, so the CNN only learns spatial patterns that those branches cannot express. This is what enables end-to-end training with heavily missing observations: the masked L1 loss penalizes only observed pixels, preventing the network from overfitting to sparse clear-sky dates.","core_discovery":"The central discovery is that temperature reconstruction can be reformulated as a single additive model embedded inside a CNN, rather than separate temporal and spatial modeling steps. The reconstruction is the sum of a learned ATC cosine per pixel, a per-pixel linear amplification of resampled ERA5 temperature, and a residual surface produced by a U-Net from surface reflectance. Because the loss is masked to observed pixels only, the whole model trains end-to-end without manual thresholds. On the four datasets, the method reaches hold-out MAE near 1.8 K and RMSE near 2.5–2.7 K for GOES-16, in situ RMSE of 3.44 K and 4.42 K for 30 m Landsat reconstructions at two sites, and, on GOES-16, better agreement with in situ measurements than the cloud-free satellite observations themselves.","pith_inferences":["A global deployment would need to prove that the reflectance-only residual branch generalizes beyond the two mid-latitude U.S. sites tested; snow, desert, irrigation, and complex terrain may produce time-varying reflectance-to-temperature relationships that the current architecture cannot express.","A directly testable extension is to make the CNN residual time-dependent by adding an auxiliary surface-state band such as snow cover or vegetation index, which would address the stationary-residual assumption without changing the physics branches.","The same additive architecture could be re-targeted to hourly 30 m temperature by replacing the annual cosine with a diurnal cycle, using GOES-16's hourly observations as training signal and Landsat for fine spatial detail.","The result that reconstructions beat cloud-free observations against in situ data suggests that evaluations of all-weather LST products should be reported separately for cloudy and clear days, since the error mechanisms differ."],"forward_implications":["Daily 30 m land surface temperature can be reconstructed for every day of the year from only tens of clear-sky Landsat scenes, with in situ RMSE of 3.44 K at the PSU site and 4.42 K at the BON site.","Hourly all-weather LST at 2 km can be reconstructed with hold-out MAE near 1.8 K, roughly half the error of the ATC-only model on the same test split.","Cloudy-day reconstructions can agree better with ground-station measurements than direct cloud-free satellite observations, because the model estimates temperature where the sensor sees clouds.","End-to-end single-model training replaces hand-crafted two-stage approaches, removing manual thresholding and the separate temporal/spatial modeling steps used by earlier methods."],"supporting_citations":[{"why":"Supplies the ATC model and the 4-step ATC baseline compared on Landsat data.","marker":"[38]"},{"why":"The nearest prior daily 30 m reconstruction approach, using enhanced ATC plus Gaussian processes, which the paper's end-to-end framework replaces.","marker":"[22]"},{"why":"The 2025 GEC-SEB competitor that provides cloudy-day reconstruction results at 1,000 m for comparison.","marker":"[6]"},{"why":"The RTM two-stage reanalysis-thermal merging method that the paper cites as the representative hand-crafted pipeline.","marker":"[36]"},{"why":"Establishes the Landsat surface temperature product's stated uncertainty of about 60% of observations within 2 K, used to contextualize reconstruction accuracy.","marker":"[15]"},{"why":"Describes GOES-16 ABI spectral bands and the sensor's roughly 2.5 K LST accuracy assumptions used in the GOES experiments.","marker":"[28]"},{"why":"Provides the emissivity estimation method used to compute in situ LST values from SURFRAD measurements.","marker":"[4]"},{"why":"Documents the roughly 66% average cloud cover that motivates all-weather reconstruction rather than clear-sky-only modeling.","marker":"[14]"}],"fun_headline_variants":["CNN fuses satellites and reanalysis for 30 m daily temperatures","Physics-guided CNN beats cloud-free satellite temperature data","Single CNN model reconstructs fine-scale temperature from coarse inputs","30 m temperature from CNN that beats satellite observations","Additive physics in CNN yields high-res daily land temperatures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the residual correction network, which takes only a surface reflectance image and no time information, can correct the ERA5-scaled field equally well on cloudy days and across all seasons; if fine-scale temperature patterns under clouds are not a stationary function of reflectance, the cloudy-day fields will simply inherit ERA5's coarse spatial shape.","fun_headline_variants_meta":{"raw":{"variants":["CNN fuses satellites and reanalysis for 30 m daily temperatures","Physics-guided CNN beats cloud-free satellite temperature data","Single CNN model reconstructs fine-scale temperature from coarse inputs","30 m temperature from CNN that beats satellite observations","Additive physics in CNN yields high-res daily land temperatures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000609,"raw_usage":{"total_tokens":2822,"prompt_tokens":916,"completion_tokens":1906,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":1827}},"tokens_in":532,"tokens_out":1906,"duration_ms":15396,"temperature":1.0,"reasoning_tokens":1827,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:45:25.048659+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a cloud-prone area with several in situ stations, train on clear-sky Landsat or GOES-16 pairs, and compare reconstructed cloudy-day temperatures against stations on days when the coarse ERA5 field places a temperature gradient opposite to the fine-scale pattern; if the model's error jumps on those days while clear-sky error stays low, the time-invariant reflectance residual is the failing component.","supporting_citations":[{"cited_title":"Reconstruction of land surface temperature under cloudy conditions from landsat 8 data using annual temperature cycle model","cited_arxiv_id":null,"evidence_quote":"Supplies the ATC model and the 4-step ATC baseline compared on Landsat data."},{"cited_title":"Daily land surface temperature reconstruction in landsat cross-track areas us- ing deep ensemble learning with uncertainty quantification","cited_arxiv_id":null,"evidence_quote":"The nearest prior daily 30 m reconstruction approach, using enhanced ATC plus Gaussian processes, which the paper's end-to-end framework replaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The 2025 GEC-SEB competitor that provides cloudy-day reconstruction results at 1,000 m for comparison."},{"cited_title":"A practical reanalysis data and thermal infrared re- mote sensing data merging (rtm) method for reconstruction of a 1-km all-weather land surface temperature","cited_arxiv_id":null,"evidence_quote":"The RTM two-stage reanalysis-thermal merging method that the paper cites as the representative hand-crafted pipeline."},{"cited_title":"Uncertainty estimation method and landsat 7 global validation for the landsat surface temperature product","cited_arxiv_id":null,"evidence_quote":"Establishes the Landsat surface temperature product's stated uncertainty of about 60% of observations within 2 K, used to contextualize reconstruction accuracy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes GOES-16 ABI spectral bands and the sensor's roughly 2.5 K LST accuracy assumptions used in the GOES experiments."},{"cited_title":"Estimating the optimal broadband emissivity spectral range for calculating surface longwave net radiation","cited_arxiv_id":null,"evidence_quote":"Provides the emissivity estimation method used to compute in situ LST values from SURFRAD measurements."},{"cited_title":"Spatial and temporal dis- tribution of clouds observed by modis onboard the terra and aqua satellites","cited_arxiv_id":null,"evidence_quote":"Documents the roughly 66% average cloud cover that motivates all-weather reconstruction rather than clear-sky-only modeling."}],"review_version":1}