{"id":"b01e8d04-4987-4d6f-8881-adf582658daa","arxiv_id":"2505.01638","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SAM-TIFF trains an RGB-only model to segment fire and estimate per-pixel temperatures, reaching about 68% of fire pixels within 25°C of radiometric ground truth on FLAME 3.","lead":"This paper trains an AI to estimate wildfire flame temperatures from ordinary color photos alone, using thermal imagery only during training. The approach could let cheaper, lighter drones monitor fires without expensive thermal cameras, though current accuracy is still far below thermal sensors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Segmentation and temperature metrics are both scored against the same SAM/TOPSIS pseudo-labels used for training, so the reported numbers may measure fit to auto-masks rather than true fire regions.","rationale":"The reader's weakest assumption is exactly the same load-bearing point: the pseudo-labels are both the training target and the evaluation reference. My independent read of Sections V and VII confirms this is the least secure link in the chain. The radiometric TIFF temperature values are trustworthy, but the region over which temperature is scored and the segmentation labels are not independently validated. I also note that the 80-20 split is not stratified by burn site, so a leave-one-burn-out test is needed for the generalization claim; however, the pseudo-label circularity is more fundamental because it can inflate even the in-distribution numbers. The proposed test with independent masks would settle whether the concern actually lands. Since the reader already conditioned acceptance on exactly this kind of check, the verdict stays conditional.","tokens_in":11374,"tokens_out":5413,"duration_ms":57836,"concrete_test":"Select ~100 test frames stratified across all six FLAME 3 burn locations. Have an annotator (or an independent protocol using a conservative radiometric threshold plus manual correction) draw fire-region polygons without seeing the SAM/TOPSIS masks. Recompute fire-class IoU (Table III row for SFAFMA-50 -> DeepLabV3+) and the ±25/±50 °C temperature accuracy (Table IV) using only pixels inside these independent masks; also compute IoU between independent masks and SAM masks. If fire IoU falls substantially below 47.3% or temperature accuracy falls below the reported 68%/84%, the pseudo-label circularity is confirmed and the headline claims need revision. If the independent masks largely coincide with SAM masks, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V builds fire-region pseudo-labels by prompting SAM with Canny/Otsu thermal-derived points and selecting masks via TOPSIS using thermal IoU, Otsu IoU, temperature means, SSIM, and SAM confidence. Section VII then uses those same pseudo-labels as the evaluation reference: Table III IoU is computed against the SAM masks, and Table IV restricts temperature evaluation to pixels where the SAM mask equals 1. In addition, the supervision itself uses the same mask: Eq. (2) masks L_Flame-L1 to F={i|SAM_i=1}. Consequently, the student is never required to detect fire boundaries independently before temperature is scored; if a test image contains fire pixels the mask missed, those pixels are excluded from both the segmentation and temperature metrics. A model that fits SAM-positive pixels well can report ~68% within ±25°C and ~84% within ±50°C while performing arbitrarily poorly on the true fire region. The manual review described in Section V only pruned low-quality masks; it does not supply independent annotations for scoring. The circularity therefore affects the central numerical claims, not just a side experiment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAM-TIFF, a teacher-student distillation framework for joint wildfire segmentation and per-pixel temperature regression from RGB-only input. A multimodal RGB-Thermal teacher network is pretrained on FLAME 2, then fine-tuned on FLAME 3 paired with automatically generated SAM/TOPSIS pseudo-masks; an RGB-only student (U-Net or DeepLabV3+) is trained to imitate the teacher's segmentation outputs and regress temperature against radiometric TIFF values via a region-masked L1 loss. On an 80-20 random split of 1,512 FLAME 3 images, the best model achieves 71.44% mIoU, 47.28% fire-class IoU, and about 68-69% of fire-region pixels within ±25°C and 84-85% within ±50°C. The paper claims to be the first to perform per-pixel wildfire temperature inference from RGB data alone and claims strong generalization on FLAME 3.","tokens_in":11570,"tokens_out":4541,"duration_ms":45530,"significance":"If the results were credible, this would be a valuable contribution: it targets a practical deployment problem (thermal-sensor-free wildfire monitoring) and makes use of the new radiometric TIFF ground truth in FLAME 3, an interesting data source that the community is only beginning to exploit. The teacher-student design, the SAM-based pseudo-labeling pipeline, and the comparison across multiple teacher architectures are useful elements. However, the significance hinges entirely on whether the reported temperature and segmentation numbers are meaningful. As detailed in the major comments, the current evaluation is compromised by shared pseudo-labels between training and testing, missing baseline comparisons, and unsupported generalization claims. These issues are load-bearing, so the contribution, while promising, is not yet established.","major_comments":[{"comment":"The evaluation is circular with respect to the pseudo-labels. Section V constructs fire-region masks using SAM prompted by thermal-derived Canny/Otsu points and selects masks via TOPSIS criteria that include IoU against a thresholded thermal JPG, Otsu IoU, and temperature patch means. Section VII then uses these same pseudo-labels as the evaluation reference: Table III computes IoU against the SAM masks, and Table IV restricts temperature evaluation to pixels where the SAM mask equals 1. In addition, Eq. (2) masks the temperature training loss to F={i|SAM_i=1}. Consequently, the student is never required to detect fire boundaries independently before temperature accuracy is scored, and any fire pixels missed by the auto-mask are excluded from both segmentation and temperature metrics. A model that fits SAM-positive pixels well can report ~68% within ±25°C and ~84% within ±50°C while performing arbitrarily poorly on the true fire region. The manual review in Section V only pruned low-quality masks; it does not provide independent annotations for scoring. This circularity affects the central numerical claims of the paper, not a side experiment. The authors should evaluate on independent ground-truth fire boundaries (e.g., manually annotated masks, or at least full-image temperature metrics) and report results over all pixels or over a held-out annotation set.","section":"Section V and Section VII, Eq. (2), Tables III and IV"},{"comment":"The paper claims that distillation from a multimodal teacher enables RGB-only temperature inference, but it never compares against the natural baseline: an RGB-only student trained directly on the RGB input with the same masked L1 temperature loss (and, ideally, the same segmentation objective) without teacher supervision. Without such a non-distilled baseline, it is impossible to tell whether the reported temperature accuracy is due to the distillation, to the pseudo-label masks, or simply to the RGB-to-temperature regression being learnable directly. The absence of this ablation is a significant omission for a paper whose central claim is about the benefit of multimodal distillation.","section":"Section VII, Tables III and IV"},{"comment":"The claim of 'strong generalization on the recent FLAME 3 dataset' is unsupported. The 80-20 random split described in Section VII likely mixes burn locations in both training and testing, so the test set may contain frames from the same burns as the training set. No cross-location experiment is reported (e.g., training on Shoetank and Sycan2A/2D and testing on Willamette Valley, or leave-one-burn-out cross-validation). The paper's own discussion in Section VII notes that some burns (e.g., Willamette Valley) are more consistent and provide better results, which suggests a strong location effect. A meaningful generalization claim requires either a cross-location evaluation or an explicit discussion of why a random split is appropriate. As is, the results may reflect overfitting to burn-specific conditions rather than generalizable RGB-only perception.","section":"Abstract and Section VII"},{"comment":"All results are reported without error bars or multiple training runs, which is problematic given the small dataset (1,512 images, 80-20 split yields roughly 300 test images). The absence of repeated seeds makes it impossible to assess the variance of the reported metrics. Additionally, the comparison across teachers is inconsistent: Section VII states that EAEF-50 converged faster and is reported at epoch 150, while all other models are reported at epoch 270. This introduces a confound—differences in Table III and Table IV across teacher rows may be due to training duration rather than teacher quality. The authors should report mean and standard deviation over at least several seeds and use a consistent training protocol (or justify the early stopping for EAEF-50 with a validation metric).","section":"Section VII (experimental setup) and Tables III and IV"},{"comment":"Several free hyperparameters are not reported, making the method non-reproducible. Specifically, λDice is set to 0.5, but the values of λstudent_dice and λFlame-L1 in Eq. (3) are not given; the temperature clipping range (stated as 0–500°C in Section II.C) is not tied to where it is applied in the pipeline; the Canny high threshold appears only in the Figure 2 caption as 200°C; the error margin ε is given as 25°C in the caption but not in the text; and the TOPSIS criterion weights are described qualitatively (thermal IoU 'given greater weight') but never specified quantitatively. A hyperparameter table and a precise description of the mask-generation thresholds are necessary for others to reproduce or build on this work.","section":"Equations (1)-(3) and Section V"}],"minor_comments":[{"comment":"There are typos: 'an approach top unify' should be 'an approach to unify', and 'low quanity' should be 'low quantity'.","section":"Section II.C"},{"comment":"The notation for Canny thresholds is confusing. The caption says 'Low Threshold = τ and High Threshold = 200 degrees Celsius', while the text defines τ as the Otsu threshold computed per image. Clarify whether the Canny low threshold is indeed the Otsu threshold or a separate parameter, and whether τ is the same variable in both places.","section":"Figure 2 caption and Section V.A"},{"comment":"The description of the FLAME 2 split is ambiguous: 'The full dataset train-validation-test split was 80-10-10 and then 10000 training, 1000 validation, and 1000 testing images were randomly sampled from this.' Clarify whether the 80-10-10 split is performed first and then samples are drawn, or whether the samples are drawn from the full dataset and then split. Also, the same seed being used 'for each network training' does not guarantee the same split if the sampling occurs before network training; this should be described precisely.","section":"Section VI"},{"comment":"Percentage values are reported inconsistently: Table II lists mIoU as 0.9563 but the text reports '96.69% mIoU', and Table III lists IoU values as 0.9515 etc. while the text reports them as percentages. Choose one format (decimal or percentage) and use it consistently throughout the tables and text.","section":"Tables II and III"},{"comment":"The temperature accuracy metric is reported only for fire-region pixels (pixels where the SAM mask equals 1). This ignores both false positives (predicted fires on non-fire pixels) and false negatives (missed fires), which are exactly the pixels where an RGB-only model could fail dangerously in practice. Reporting full-image temperature errors or per-class breakdowns would make the practical implications clearer.","section":"Section VII (temperature evaluation)"},{"comment":"The paper notes that the high FLAME 2 numbers likely reflect the networks learning the Otsu thresholding and thermal colormap rather than true fire segmentation. This is an important caveat, but it is only mentioned in one sentence and is not carried into the discussion of the final FLAME 3 results. Please discuss how this pretraining bias might propagate through the distillation and affect the final temperature predictions.","section":"Section VI, paragraph on FLAME 2 pretraining"}],"recommendation":"major_revision","confidential_remarks":"The paper tackles a worthwhile problem and uses a novel dataset element (radiometric TIFFs), but the experimental validation is not yet at the level required for the claims made. The most serious problem is the circular use of the SAM/TOPSIS pseudo-masks for both training and evaluation, which I have flagged as a major comment. If the authors cannot obtain independent fire-region annotations (even a modest manually labeled subset), they should substantially temper the claims and report results on the full image or on a clearly held-out manual annotation set. The missing non-distilled baseline and lack of cross-location evaluation are also essential to add. The paper is within scope for the journal, and the topic is timely, so a major revision is appropriate; I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick read on arXiv:2505.01638. The genuinely new thing here is the problem setup: per-pixel wildfire temperature regression from RGB-only input, using FLAME 3's radiometric TIFFs as regression targets, with a multimodal teacher distilling to an RGB student. To my knowledge that combination is new, and the authors get credit for attempting it and for reporting their numbers transparently — including the admission that their FLAME 2 pretraining probably learned Otsu's thresholding, and that the ±25°C/±50°C accuracy is far from the ±2°C of a good IR sensor.\n\nThe soft spots are real, and the biggest one is the evaluation loop. The same SAM/TOPSIS masks that supervise the student (Eq. 2 restricts the temperature loss to SAM=1 pixels) are also used as the reference for both IoU in Table III and for selecting the pixels scored in Table IV. So the segmentation numbers are fit to auto-masks, not to any independent notion of fire boundary, and the temperature numbers only tell you how well the model does on pixels the pipeline already decided are fire. That is not a completely circular regression—the TIFF temperatures themselves are independent—but it does mean the headline accuracies could shrink on scenes where the auto-masks miss real fire. The paper even mentions manual pruning of masks, but that doesn't supply independent test labels.\n\nOther gaps are more standard: no direct RGB-only baseline trained on TIFFs without distillation, no error bars or multiple seeds, and no held-out burn-location test despite the claim of 'strong generalization.' The authors do discuss some of these limitations themselves, which is to their credit.\n\nBottom line: worth a serious referee, but only with major revision. The contribution is the task and pipeline, not yet a proven accuracy level. If you're working on wildfire perception or on distillation evaluation pitfalls, this is a useful case study; I wouldn't cite the numbers as-is.","headline":"First RGB-only wildfire temperature regression task, but the evaluation uses the same auto-masks for both training and scoring, so the headline numbers are not yet trustworthy.","tokens_in":12120,"tokens_out":2006,"would_cite":false,"duration_ms":20745,"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":"RGB-only drone images can predict wildfire pixel temperatures after multimodal distillation.","keywords":["wildfire temperature prediction","RGB-only inference","knowledge distillation","radiometric thermal TIFF","SAM-guided pseudo-labels","TOPSIS mask selection","FLAME 3 dataset","UAV wildfire monitoring"],"falsifier":"Have human annotators draw fire boundaries on a held-out subset of FLAME 3, then recompute the student's ±25°C and ±50°C temperature accuracy only inside the human masks and compare with the same accuracy computed inside the SAM pseudo-label masks; a large gap would show the reported numbers are artifacts of the auto-generated labels.","tokens_in":11138,"feed_emoji":"🔥","tokens_out":7591,"duration_ms":67360,"temperature":0.7,"pith_summary":"The paper claims that a drone can estimate per-pixel wildfire temperatures from ordinary RGB images alone, as long as a multimodal teacher network has first learned from paired RGB-thermal data and radiometric TIFF ground truth. The authors build SAM-TIFF, a teacher-student distillation framework in which an RGB-thermal fusion network supervises an RGB-only U-Net or DeepLabV3+ student that simultaneously segments fire regions and regresses temperature. On the FLAME 3 test split, the best student reaches about 68% of fire-region pixels within ±25°C and 84% within ±50°C, with a segmentation mIoU of 71.4%. If correct, the result removes the need for thermal cameras at inference time, which would lower the cost, weight, and power budget of wildfire monitoring UAVs while keeping temperature information available.","feed_headline":"RGB-only drones can estimate fire temperature after distillation","feed_subtitle":"A distilled RGB student keeps thermal accuracy without a thermal camera: 68% of fire pixels within 25°C, cutting UAV cost and weight.","key_machinery":"The load-bearing mechanism is teacher-student modality distillation with a fire-region-masked temperature loss. The teacher is an RGB-thermal fusion network, such as SFAFMA with a ResNet-50 backbone, pretrained on FLAME 2 and fine-tuned on FLAME 3; the student is a U-Net or DeepLabV3+ that sees RGB input only. Distillation transfers the teacher's segmentation predictions to the student, while the region-masked L1 loss, $L_{\\text{Flame-L1}}$, supervises temperature regression only on pixels flagged as fire by SAM pseudo-labels, forcing the student to learn an RGB-to-temperature mapping. The pseudo-labels themselves come from SAM point prompts guided by Canny edge detection and Otsu thresholding, with TOPSIS selecting among SAM's candidate masks using thermal criteria such as IoU with thresholded thermal JPGs, temperature patch means, and SSIM.","core_discovery":"On the paper's own terms, the central discovery is that thermal knowledge can be transferred into an RGB-only model well enough to recover continuous temperature fields, not just fire/no-fire labels. SAM-TIFF trains a multimodal teacher on paired RGB-Thermal frames and FLAME 3's calibrated radiometric TIFFs, then distills its segmentation predictions into a student that sees only RGB. The student's temperature head uses sigmoid scaling to constrain outputs to 0–500°C and is supervised by a region-masked L1 loss, $L_{\\text{Flame-L1}}$, that computes absolute error only on pixels flagged as fire by SAM-generated pseudo-labels. The authors report that across teacher-student variants, fire-region temperature accuracy clusters around 68% within ±25°C and 83–85% within ±50°C, with the SFAFMA-50 teacher plus DeepLabV3+ student giving the best balance (71.4% mIoU, 47.3% fire IoU). They position this as the first per-pixel wildfire temperature regression from RGB UAV data, enabled by the first use of radiometric TIFFs as temperature ground truth.","pith_inferences":["The reported accuracy is measured only inside the SAM pseudo-label fire regions; if those masks are conservative or leaky, the same model evaluated on human-drawn fire boundaries could score noticeably lower, an evaluation the paper does not run.","At 68% within ±25°C, the student is far coarser than the ±2°C precision of high-end drone IR sensors, so RGB-only inference is better read as a coarse hotspot detector than a thermal-camera replacement; closing that gap is the natural next target.","A stress test worth running is deployment on smoke-obscured or nighttime RGB frames, where color cues degrade; since the student never sees thermal input, the distillation can only help to the extent the teacher's thermal channel reveals what RGB hides.","The 0–500°C clipping and the small saturated-pixel population mean predictions above 450–500°C are untrustworthy; an extension would be a separate saturation-aware head or uncertainty estimate for extreme temperatures."],"forward_implications":["Deployed fire-monitoring UAVs can drop thermal sensors entirely while still outputting per-pixel temperature maps, reducing payload, cost, and power draw.","RGB-only temperature regression makes existing RGB drone fleets and even consumer cameras potential wildfire temperature sensors via a software update.","The SAM-guided pseudo-label pipeline turns unlabeled thermal datasets into segmentation supervision, removing the manual annotation bottleneck for fire masks.","Radiometric TIFFs become a reusable ground-truth modality for training temperature-aware models beyond classification, opening the door to hotspot and heat-gradient analysis.","Because the teacher can be any RGB-thermal fusion network, the distillation recipe is architecture-agnostic; better teachers or fusion blocks should raise the student ceiling."],"supporting_citations":[{"why":"FLAME 3 dataset with paired RGB-thermal frames and radiometric TIFFs; supplies the temperature ground truth and evaluation benchmark.","marker":"[4]"},{"why":"FLAME 2 dataset used to pretrain the teacher networks with Otsu-generated segmentation masks, giving them wildfire domain knowledge before distillation.","marker":"[12]"},{"why":"Segment Anything Model; generates candidate fire masks from auto-located point prompts for pseudo-label supervision.","marker":"[20]"},{"why":"TOPSIS multi-criteria decision method; selects the best SAM candidate mask using thermal IoUs, temperature means, confidence, and SSIM.","marker":"[21]"},{"why":"Otsu's thresholding; provides the classical segmentation baseline and the point-prompt filtering used to build pseudo-labels.","marker":"[22]"},{"why":"SFAFMA; the RGB-thermal fusion network whose ResNet-50 variant serves as the strongest teacher in the reported results.","marker":"[17]"},{"why":"DeepLabV3+; the student architecture that achieves the best segmentation mIoU when paired with the SFAFMA-50 teacher.","marker":"[13]"},{"why":"U-Net; used both as a student architecture and as a teacher backbone in the studied combinations.","marker":"[24]"},{"why":"Pesonen et al.; prior work distilling smoke segmentation from a larger teacher with SAM pseudo-labels, the direct predecessor this paper extends to temperature regression.","marker":"[19]"}],"fun_headline_variants":["RGB-only wildfire temperature from distilled thermal knowledge","Seeing heat in color: RGB-only wildfire temp via distillation","Distilled thermal vision gives drones fire temperature in RGB","Thermal-free wildfire temperature prediction from RGB alone via distillation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation treats SAM-generated pseudo-labels as the ground-truth fire region for both training and temperature scoring; if those masks do not match true fire boundaries, the reported accuracy and IoU numbers measure the model against its own mask generator, not against real fire regions.","fun_headline_variants_meta":{"raw":{"variants":["RGB-only wildfire temperature from distilled thermal knowledge","Seeing heat in color: RGB-only wildfire temp via distillation","Distilled thermal vision gives drones fire temperature in RGB","Thermal-free wildfire temperature prediction from RGB alone via distillation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000684,"raw_usage":{"total_tokens":3115,"prompt_tokens":969,"completion_tokens":2146,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":2095}},"tokens_in":585,"tokens_out":2146,"duration_ms":15904,"temperature":1.0,"reasoning_tokens":2095,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:13:46.032282+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have human annotators draw fire boundaries on a held-out subset of FLAME 3, then recompute the student's ±25°C and ±50°C temperature accuracy only inside the human masks and compare with the same accuracy computed inside the SAM pseudo-label masks; a large gap would show the reported numbers are artifacts of the auto-generated labels.","supporting_citations":[{"cited_title":"Flame 3 - radiometric thermal uav imagery for wildfire management,","cited_arxiv_id":null,"evidence_quote":"FLAME 3 dataset with paired RGB-thermal frames and radiometric TIFFs; supplies the temperature ground truth and evaluation benchmark."},{"cited_title":"Hwang and K","cited_arxiv_id":null,"evidence_quote":"TOPSIS multi-criteria decision method; selects the best SAM candidate mask using thermal IoUs, temperature means, confidence, and SSIM."},{"cited_title":"Sfaf-ma: Spatial feature aggregation and fusion with modality adaptation for rgb-thermal semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"SFAFMA; the RGB-thermal fusion network whose ResNet-50 variant serves as the strongest teacher in the reported results."},{"cited_title":"Encoder-decoder with atrous separable convolution for semantic image segmentation,","cited_arxiv_id":null,"evidence_quote":"DeepLabV3+; the student architecture that achieves the best segmentation mIoU when paired with the SFAFMA-50 teacher."}],"review_version":1}