{"id":"bf7e50ba-0b14-4dfb-b627-aa88134c92d6","arxiv_id":"2509.02656","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A low-cost reflectance sensor using commodity color and IR detectors plus random forest reports 89.58% accuracy at distinguishing symptomatic HLB-infected from healthy citrus leaves, but leaf-level validation is missing.","lead":"This paper builds a low-cost LED and Arduino based optical sensor that classifies citrus leaves as healthy or infected with Huanglongbing (citrus greening), reporting 89.58% accuracy for the infrared band with a random forest model. The validation is structurally weak because measurements from the same leaf appear in both training and test sets, so independent confirmation is needed before the device can be considered a screening tool.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 89.58% accuracy likely inflated by leaf-identity leakage: 10 spot measurements per leaf (Sec 2.1) are split at measurement level (Sec 2.3), so the same leaf occurs in train and test. Leaf-level cross-validation is needed before the central claim holds.","rationale":"The reader's weakest assumption identifies exactly the load-bearing flaw: the 80:20 split is performed on measurements, not on leaves, so the test set is not independent. This is the single most important issue because it directly invalidates the 89.58% accuracy as an estimate of generalization to new leaves. The paper's own methods (Sec 2.1 and 2.3) make the leakage mechanism explicit: 24 leaves × 10 measurements, stratified split on the full dataset. A leaf-level cross-validation is the minimal check that would settle whether the result survives. The additional problem of unverified healthy-leaf labels, noted in the reader's rationale, further weakens the central claim but is secondary to the independence violation. Because the central accuracy claim is not supported, the reader's REJECT verdict stands; no adjustment is needed.","tokens_in":10037,"tokens_out":6024,"duration_ms":72251,"concrete_test":"Release the raw 240×4 dataset and re-run the Random Forest evaluation using leave-one-leaf-out cross-validation: train on all measurements from 23 leaves, test on all 10 measurements of the held-out leaf, repeat for each of the 24 leaves, and report aggregate plus per-leaf accuracy. If the aggregate accuracy falls substantially below 89.58% (for example toward chance or below 70%), the reported test accuracy is inflated by within-leaf leakage. Also report whether healthy leaves were qPCR-negative; if any 'healthy' leaf tests positive, the class labels themselves are invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is that a Random Forest on IR reflectance reaches 89.58% accuracy, supporting a low-cost HLB screening tool. For that claim to transfer to new leaves, test measurements must be independent of training leaves. The protocol violates this: 24 leaves (7 healthy, 17 infected) yield 240 measurements at 10 per leaf (Sec 2.1), and the 80:20 stratified split is described on 'the full dataset' (Sec 2.3), not on leaves. Thus the same leaf can contribute measurements to both training and test partitions. Reflectance spot measurements from one leaf share leaf thickness, surface texture, sensor contact, and lighting conditions; a 500-tree Random Forest can exploit these leaf-specific artifacts rather than disease biology. With only 48 test measurements from a small, imbalanced leaf set, the reported accuracy is not an estimate of performance on new leaves. Compounding this, the 'healthy' leaves were not qPCR-confirmed negative; the tree/branches were HLB-positive (Sec 2.1), so the healthy label itself is unverified. No code, data, or confidence intervals are provided, so the result cannot be independently checked. These issues directly undermine the paper's headline accuracy and its applicability as a screening tool.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a low-cost LED-based reflectance sensor for distinguishing citrus greening (HLB)-infected from healthy citrus leaves. The sensor uses white and IR LEDs with a color sensor (TCS34725) and a custom IR photodiode module, and reflectance measurements in blue, green, red, and IR bands are classified with k-NN, decision tree, and random forest. The authors report best performance for Random Forest in the IR band: 89.58% accuracy and 93.75% precision. They claim the device costs under ₹5000 and could serve as a hand-held early-screening tool. The evaluation uses 24 leaves (7 healthy, 17 infected) with 10 spot measurements per leaf, split 80:20 at the measurement level.","tokens_in":10305,"tokens_out":4449,"duration_ms":53220,"significance":"If the result held under proper validation, a ~$60/₹5000 hand-held reflectance sensor with ~90% accuracy would be a genuinely useful screening tool for smallholder citrus growers, and the physiological rationale (starch accumulation visible in IR, chlorophyll loss in green) is plausible. The paper also provides a useful comparison with expensive research-grade spectrometers. However, the current evaluation has fundamental flaws: the train/test split is at the measurement level, so the same leaf appears in both training and test sets; the 'healthy' leaves were not confirmed negative by qPCR; and all infected leaves were visibly symptomatic. These issues mean the headline accuracy cannot be interpreted as performance on new leaves or as evidence for 'early' detection. The engineering contribution is promising, but the central claim as stated is not supported by the reported experimental design.","major_comments":[{"comment":"The train/test split is at the measurement level, not the leaf level. Ten reflectance measurements are taken from each of 24 leaves (Section 2.1), and the 80:20 stratified split is applied to 'the full dataset' (Section 2.3). Thus the same leaf contributes measurements to both training and test partitions. Reflectance spot measurements from one leaf share leaf thickness, surface texture, sensor contact, and lighting conditions; a 500-tree random forest can exploit these leaf-specific artifacts rather than disease biology. The reported 89.58% accuracy (Table 1) is therefore not an estimate of performance on new leaves. Leaf-grouped cross-validation (e.g., leave-one-leaf-out or grouped k-fold) must be reported, along with per-leaf predictions and confidence intervals.","section":"§2.1, §2.3"},{"comment":"The 'healthy' leaf label is not validated. The branches from which the leaves were taken were qPCR-positive for HLB, and no qPCR test on the 7 healthy-appearing leaves is reported. These leaves may be asymptomatic but infected. At the same time, the 17 infected leaves all exhibited blotchy mottle and green island symptoms. The classification therefore discriminates visibly symptomatic leaves from asymptomatic leaves on the same infected tree, not healthy vs. diseased leaves. The authors must either confirm the healthy leaves as qPCR-negative (and ideally sample from disease-free trees) or explicitly reframe the claim as 'symptomatic vs. asymptomatic leaf discrimination.'","section":"§2.1"},{"comment":"Several reported performance metrics are internally inconsistent. For Random Forest IR, precision is 0.9375 and recall is 0.9231, but F1-score is listed as 0.9091; the harmonic mean of these two values is approximately 0.930. Similar inconsistencies appear for k-NN Green and Decision Tree Blue. This suggests either arithmetic errors or misreported confusion-matrix values. Please provide the confusion matrices for all models and bands and recompute the metrics, or share code so the numbers can be reproduced.","section":"§3.2, Table 1"},{"comment":"The claim of 'early detection' or 'early screening' is not supported by the data. All infected leaves already had visible HLB symptoms (blotchy mottle, green islands; Section 2.1). No asymptomatic infected leaves were tested. If the device is being proposed for early screening, the authors need to include asymptomatic infected leaves (confirmed by qPCR) or tone down the claim to 'detection of symptomatic HLB leaves.' This is a load-bearing issue for the stated contribution.","section":"§4, Abstract"},{"comment":"The evaluation is based on a single tree and only 24 leaves, with no confidence intervals, no external validation set, and no code/data availability. The ROC/PR AUC values reported in Section 3.3 inherit the same measurement-level leakage. With only ~48 test measurements (unbalanced), the uncertainty around 89.58% is large. At minimum, the authors should provide bootstrap or leaf-level confidence intervals and make the data available. Without this, the central accuracy claim cannot be independently checked.","section":"§3.3, §3.4"}],"minor_comments":[{"comment":"Typo: 'spreaded' should be 'spread.' Also 'Disease is highly destructive' should be 'The disease is highly destructive.'","section":"§1"},{"comment":"Cost is given as 'under ₹5000' in the text but '$60' in Table 2. Please clarify the currency and the basis of the estimate (component price, prototype cost, or projected manufacturing cost).","section":"§3.4, Table 2"},{"comment":"The description of hyperparameters is incomplete: k-NN uses k=3, Random Forest uses 500 trees and √p features, but no maximum-depth or pruning details are given for the decision tree. A brief hyperparameter table would improve reproducibility.","section":"§2.3"},{"comment":"The learning curves are described in §3.3, but the figure panels are not individually labeled in the text. Please ensure that panel (d) is clearly referenced and that axes and legends are legible.","section":"Figure 3"},{"comment":"No data availability or code availability statement is provided. Given the reproducibility concerns, these statements are essential and should be added.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The fundamental evaluation flaws (measurement-level split, unvalidated 'healthy' class, symptomatic-only infected leaves) are serious. They are, however, addressable in principle: leaf-level cross-validation can be rerun on the existing data, and new qPCR-confirmed healthy/asymptomatic leaves can be collected. The engineering contribution is plausible and of interest to an applied readership. I recommend major revision rather than rejection, but if the authors cannot provide a leaf-level validation and a meaningful healthy control, the manuscript should not be accepted in any form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before reading this one. The hook is real: a sub-₹5000 Arduino-based reflectance sensor (TCS34725 plus a 940 nm IR emitter/photodiode) that reportedly distinguishes symptomatic HLB leaves from \"healthy\" ones with 89.58% accuracy via Random Forest in the IR band. The catch: the evaluation protocol does not support that number. The 80:20 split is on the 240 spot measurements, not on the 24 leaves, so the same leaves appear in both training and test sets. That inflates accuracy through leaf identity leakage. No leaf-level cross-validation, no confidence intervals, no code, no data.\n\nWhat is genuinely new is the hardware. Prior HLB spectroscopy work mostly used research-grade vis-NIR spectrometers costing far more. This design is simple, well-explained, and the physiological rationale (starch accumulation in IR, chlorophyll loss in green) is reasonable. The writing is clear, and the comparison table with prior instruments is useful. If the sensor worked, it would be a meaningful triage tool for smallholders.\n\nBut the soft spots go beyond the split. The healthy leaves came from a tree whose branches were qPCR-confirmed HLB-positive; the paper never states the healthy leaves were confirmed negative. The infected leaves already had blotchy mottle and green islands, so this is symptomatic-versus-unsymptomatic discrimination, not \"early screening\" as claimed. The dataset is 24 leaves from a single tree, so even a clean split would not support orchard-level generalization. The RF hyperparameters and the choice of band are all selected on the same data, so the 89.58% is the best of four bands and three models—an optimistic selection effect stacked on top of the leakage.\n\nThese are not minor quibbles. The central claim is load-bearing, and the evaluation as reported cannot distinguish leaf identity artifacts from disease biology. That said, the flaws are fixable: leaf-level cross-validation, PCR-confirmed healthy leaves, and multi-tree/multi-season sampling would make the claim credible.\n\nThis paper is for researchers working on low-cost plant disease sensors and anyone teaching data-leakage pitfalls. With revision it could be a decent engineering report. I would not accept the current numbers as evidence, but I would send it to peer review if I were the editor—the topic is important and the hardware idea deserves a serious referee pushing for proper validation. If you read it, read Section 2.3 closely.","headline":"A useful low-cost hardware idea undercut by a measurement-level train/test split; the reported 89.58% accuracy should not be trusted until leaf-level validation is done.","tokens_in":10862,"tokens_out":2182,"would_cite":false,"duration_ms":24633,"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 reports that a hand-held LED reflectance sensor priced under ₹5,000 identifies citrus greening (HLB)-infected leaves with 89.58% accuracy and 93.75% precision in the infrared band.","keywords":["Huanglongbing","citrus greening","LED reflectance sensor","Random Forest","infrared sensing","plant disease screening","reflectance spectroscopy","machine learning"],"falsifier":"Retrain the Random Forest with all ten measurements from each leaf kept in the same training or test fold (leaf-disjoint cross-validation). If the resulting test accuracy falls well below 89.58%, the published number is inflated by reading-level leakage; alternatively, collect fresh leaves from a second tree and apply the already trained model—if accuracy collapses, the model has not generalised beyond the original tree.","tokens_in":9887,"feed_emoji":"🍊","tokens_out":7145,"duration_ms":82336,"temperature":0.7,"pith_summary":"The paper's goal is to show that citrus greening (Huanglongbing, HLB) can be screened for with a cheap, portable LED-based reflectance sensor instead of research-grade spectrometers or lab assays. The proposed device illuminates the upper surface of a leaf with white and infrared LEDs, records reflected intensities in four spectral bands (blue, green, red, infrared), and feeds the normalized readings to machine-learning classifiers. The central result is that a Random Forest model on the infrared band separates HLB-infected from healthy leaves with 89.58% accuracy and 93.75% precision, with the green band close behind. If the result holds, smallholder farmers could screen trees for a few thousand rupees and reserve PCR tests for suspected positives, which matters because HLB has no cure and spreads before symptoms appear.","feed_headline":"89% accuracy: a Rs 5,000 LED sensor detects citrus greening","feed_subtitle":"A hand-held reflectance sensor that flags infected leaves for PCR confirmation could give small farms affordable early screening.","key_machinery":"The load-bearing mechanism is reflectance spectroscopy with two illumination sources and four wavelength channels. White LEDs excite chlorophyll-sensitive bands (blue 465 nm, green 525 nm, red 615 nm) and an IR LED/photodiode pair reads around 940 nm. The paper's physiological explanation is that IR reflectance tracks starch accumulation caused by phloem blockage in infected leaves, while the green band tracks chlorophyll decline; these four intensity features, scaled to [0,1], are classified by Random Forest with 500 bootstrap-aggregated trees. This machinery translates an expensive spectroscopy measurement into four cheap intensity readings that still carry enough disease information for t","core_discovery":"The paper claims that a minimal optoelectronic circuit—an Arduino Nano driving a TCS34725 colour sensor and a custom 940 nm IR emitter–photodiode pair—captures disease-relevant reflectance changes from the adaxial leaf surface. The underlying physical claim is that HLB infection blocks phloem transport, causes abnormal starch accumulation and chlorophyll loss, and these biochemical shifts alter how leaves reflect light at the sensor's chosen wavelengths. After min-max normalization and training on reflectance measurements from 24 leaves, Random Forest achieved 89.58% test accuracy and 93.75% precision in the IR band, and 85.42% accuracy in the green band. The authors position the device as a","pith_inferences":["Because all infected leaves showed blotchy mottle and green island symptoms and came from one tree, the paper's 'early screening' claim is tested only on visibly symptomatic leaves; a genuinely early field test would target asymptomatic qPCR-positive leaves, on which this sensor has not yet been evaluated.","The reading-level split means the 89.58% figure likely overstates leaf-level generalization; a leaf-disjoint cross-validation would give a fairer estimate and is a direct next step.","If starch and chlorophyll are the underlying discriminative signals, the same four-band sensor could plausibly detect other phloem-limited or nutrient-stress conditions, but nothing in the paper demonstrates that.","Field deployment will face temperature, humidity, and leaf-age variability; a calibration set spanning orchards, seasons, and cultivars is a natural extension, not yet reported."],"forward_implications":["If the sensor works as reported, citrus orchards in low-income settings could run a first-pass screen with a device costing under ₹5,000 and send only positives for qPCR confirmation.","Infrared and green band reflectance become practical proximal markers for HLB-related starch and chlorophyll changes, measurable outside a laboratory.","The reported result suggests that four narrow wavelength channels can reproduce, at lower accuracy, what research-grade vis-NIR spectrometers do across hundreds of wavelengths.","The device could be redeployed for routine monitoring of symptom progression if starch and chlorophyll signatures shift with disease severity."],"supporting_citations":[{"why":"Supplies the qPCR diagnostic protocol used as ground truth for HLB-positive status of sampled branches.","marker":"[1]"},{"why":"Describes the real-time PCR method used to confirm HLB in the collected leaves.","marker":"[21]"},{"why":"Provides the starch accumulation mechanism in HLB-affected trees that the paper invokes to explain why the IR band works.","marker":"[30]"},{"why":"Supports chlorophyll loss as a symptom-related marker, the physiological basis for the green band's discrimination.","marker":"[31]"},{"why":"Reports 97% SVM accuracy on vis-NIR spectra for HLB detection, the research-grade baseline the low-cost device is compared against.","marker":"[8]"},{"why":"Demonstrates citrus greening identification using a VIS-NIR spectrometer, another high-cost baseline in the comparison table.","marker":"[16]"},{"why":"Describes an earlier active optical sensor and key disease-stress wavelengths; the comparison table benchmarks the proposed device against it.","marker":"[19]"},{"why":"Shows spectral feature extraction results for citrus greening detection, providing another baseline and the average 85% accuracy cited.","marker":"[20]"}],"fun_headline_variants":["LED sensor spots citrus greening with 89% accuracy for under Rs 5,000","Cheap LED sensor flags HLB-infected leaves at 89% accuracy","Hand-held LED device detects citrus greening early at low cost","89% accurate sensor for citrus greening costs just Rs 5,000","Affordable optoelectronic sensor screens citrus leaves for HLB"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The reported accuracy rests on treating the ten reflectance readings from each leaf as independent samples and splitting them randomly between training and testing; if the split is done leaf-wise instead, the model may be memorising leaf-specific patterns, and the 89.58% figure could drop sharply.","fun_headline_variants_meta":{"raw":{"variants":["LED sensor spots citrus greening with 89% accuracy for under Rs 5,000","Cheap LED sensor flags HLB-infected leaves at 89% accuracy","Hand-held LED device detects citrus greening early at low cost","89% accurate sensor for citrus greening costs just Rs 5,000","Affordable optoelectronic sensor screens citrus leaves for HLB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001279,"raw_usage":{"total_tokens":5058,"prompt_tokens":731,"completion_tokens":4327,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":4240}},"tokens_in":475,"tokens_out":4327,"duration_ms":33144,"temperature":1.0,"reasoning_tokens":4240,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:35:23.419871+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the Random Forest with all ten measurements from each leaf kept in the same training or test fold (leaf-disjoint cross-validation). If the resulting test accuracy falls well below 89.58%, the published number is inflated by reading-level leakage; alternatively, collect fresh leaves from a second tree and apply the already trained model—if accuracy collapses, the model has not generalised beyond the original tree.","supporting_citations":[{"cited_title":"Current Distribution of Huanglongbing (citrus greening disease) in India as Diagnosed by Real-Time PCR,","cited_arxiv_id":null,"evidence_quote":"Describes the real-time PCR method used to confirm HLB in the collected leaves."},{"cited_title":"Anatomical distribution of abnormally high levels of starch in HLB-affected Valencia orange trees,","cited_arxiv_id":null,"evidence_quote":"Provides the starch accumulation mechanism in HLB-affected trees that the paper invokes to explain why the IR band works."},{"cited_title":"Developing Citrus Huanglongbing (HLB) Management Strategies Based on the Severity of Symptoms in HLB-Endemic Citrus-Producing Regions,","cited_arxiv_id":null,"evidence_quote":"Supports chlorophyll loss as a symptom-related marker, the physiological basis for the green band's discrimination."},{"cited_title":"Visible-near infrared spectroscopy for detection of Huanglongbing in citrus orchards,","cited_arxiv_id":null,"evidence_quote":"Reports 97% SVM accuracy on vis-NIR spectra for HLB detection, the research-grade baseline the low-cost device is compared against."},{"cited_title":"Identification of Citrus Greening (HLB) Using a VIS-NIR Spectroscopy Technique,","cited_arxiv_id":null,"evidence_quote":"Demonstrates citrus greening identification using a VIS-NIR spectrometer, another high-cost baseline in the comparison table."},{"cited_title":"Evaluation of an active optical sensor for detection of Huanglongbing (HLB) disease,","cited_arxiv_id":null,"evidence_quote":"Describes an earlier active optical sensor and key disease-stress wavelengths; the comparison table benchmarks the proposed device against it."},{"cited_title":"Visible-near infrared spectroscopy based citrus greening detection: Evaluation of spectral feature extraction techniques,","cited_arxiv_id":null,"evidence_quote":"Shows spectral feature extraction results for citrus greening detection, providing another baseline and the average 85% accuracy cited."}],"review_version":1}