{"id":"2ed052eb-71ea-440f-be82-997a297670b9","arxiv_id":"2509.06367","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 0.231M-parameter CNN achieves 90.0% accuracy on potato drought stress detection, and removing 5% of low-gradient-score training samples improves accuracy from 88.6% to 90.0%.","lead":"This paper introduces a lightweight hybrid CNN for detecting drought stress in potato crops from aerial images, with 0.231M trainable parameters and 90% accuracy on a public dataset. It also proposes a gradient-norm-based \"unlearning\" step that removes 5% of the least influential training samples before retraining.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3's accuracy comparison is not shown to use a common test split; the 'competitive accuracy' claim is unverified.","rationale":"The reader's weakest_assumption targets the missing random-removal baseline for the unlearning mechanism. That is a real concern, but it affects a secondary contribution (the causal benefit of gradient-guided unlearning), not the central claim of competitive accuracy and parameter efficiency. The proposed model still achieves 90.0% with 0.231M parameters even if random removal gives the same result; the headline architecture claim would stand. In contrast, the comparison to baselines in Table 3 is the sole evidence that 90.0% is 'competitive' with SOTA. If those baselines were not evaluated on the same test split, the entire 'maintaining competitive accuracy' assertion is unsupported, and the paper's primary value proposition collapses. The reader did mention the non-independent baselines in the rationale, so there is partial agreement, but the weakest_assumption field emphasized the unlearning issue. I chose the comparison-fairness concern as more load-bearing because it directly determines whether the strongest_claim is justified. The concrete test—re-evaluating all models on a common split—would settle this. I recommend keeping the CONDITIONAL verdict, as the paper is an engineering contribution that could be validated with additional experiments, but the current evidence is insufficient to accept the accuracy-comparison claim as stated.","tokens_in":11222,"tokens_out":13171,"duration_ms":143742,"concrete_test":"Retrain MobileNet-based pipeline [13], DenseNet121-based pipeline [13], and ViT-TL [26] on the exact training split (300 images, augmented to 1,500) and evaluate on the exact test split (60 images, 1,135 windows) described in Section 2.1, using identical preprocessing and window extraction. If the proposed model's 90.0% accuracy is not within a few points of the baselines (or if it drops below them under the common split), the 'competitive accuracy' claim fails. Alternatively, inspect the method sections of [13] and [26] to verify they used the identical test split; if they did not, Table 3 is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a 90.0% accuracy that is 'competitive' with existing pipelines rests on Table 3, which lists accuracies for MobileNet/DenseNet121 from [13] and ViT-TL from [26] alongside the proposed model. The paper never states that these baselines were retrained and evaluated on the exact same test set used for the proposed model (60 images, 1,135 windows). Because [13] and [26] are prior publications by the same group, they may have used a different image split, window extraction, or preprocessing. If the test sets differ, the 1.6-point gap to ViT-TL and the 0.7-point gap to DenseNet121 are not interpretable, and the 'competitive accuracy' claim has no evidentiary basis. The parameter-count reduction (0.231M vs 3.5–14M) is a separate, verifiable architectural fact, but the paper's headline value—that a tiny model is near state-of-the-art—depends directly on a fair accuracy comparison. No code or detailed protocol is provided to confirm a shared evaluation, so this condition is the least secure in the central argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MRD-LiNet, a lightweight hybrid CNN combining MobileNetV2-style bottleneck residual blocks, a DenseNet-like dense block, and a transition layer, with 0.231M trainable parameters. It also introduces a machine-unlearning step that computes per-sample influence scores from gradients, removes the 5% least influential training samples, and retrains the model. The method is evaluated on a potato drought-stress aerial dataset with 60 test images and 1,135 annotated windows. The authors report accuracies of 88.1% (no augmentation), 88.6% (augmentation), and 90.0% (augmentation plus unlearning), and compare these against three prior pipelines (MobileNet, DenseNet121, ViT-TL), claiming a 15–60-fold parameter reduction and competitive accuracy.","tokens_in":11530,"tokens_out":5078,"duration_ms":56830,"significance":"If the empirical claims hold, the parameter-efficiency result (0.231M parameters at 90.0% accuracy) is practically useful for UAV/edge deployment, and the use of gradient-guided unlearning to improve generalization is an interesting direction. The paper does not, however, ship machine-checked proofs, reproducible code, or a random-removal control; the main empirical contributions are supported only by single-run comparisons. The comparison against prior work is also not demonstrated to be protocol-matched. The central ideas are therefore worthy of consideration, but the evidence as presented is insufficient for acceptance.","major_comments":[{"comment":"The only evidence for the unlearning benefit is the 1.4-point gap between Aug (88.6%) and Aug+MU (90.0%) on 1,135 test windows, i.e., roughly 16 correctly classified samples. There is no random-removal control, no repeated runs, and no error bars. Without a random-5%-removal baseline and multiple seeds, the improvement cannot be attributed to gradient-guided selection rather than data reduction or chance. Please add the control and report mean±std.","section":"§4.1, Scenario (iii)"},{"comment":"The influence score is defined inconsistently. The displayed formula defines it as the norm of the gradient of the loss with respect to the model prediction, while §3.1.1 and Algorithm 2 compute the L2 norm of the gradient of the prediction with respect to the input image, flattened and concatenated across layers. These are different quantities. The manuscript must state exactly which gradient was computed and how the score was aggregated; otherwise the unlearning mechanism is not reproducible.","section":"§3.1 vs §3.1.1"},{"comment":"The accuracy comparison is not shown to be protocol-matched. Baselines [13] and [26] are previous papers by the same group, and the text does not establish that they used the same 60-image/1,135-window test split, same preprocessing, same augmentation, or same evaluation script. As reported, the 0.7–1.6 point gaps to DenseNet121/ViT-TL are not interpretable. Additionally, §4.2 claims the proposed model has the lowest stressed-class false-negative rate (13.5%), but Table 3 shows ViT-TL has stressed recall 0.901, i.e., 9.9% false negatives. This claim is factually wrong as stated.","section":"Table 3 and §4.2"},{"comment":"The architecture specification is internally inconsistent: §3.0.1 states the initial convolution uses 32 filters, while Algorithm 1 sets Conv2D(36,(3,3)). The exact filter count affects the parameter total, which is a headline result. Please reconcile the text and algorithm, and state the precise parameter-count calculation. A public code release would also resolve this ambiguity.","section":"Algorithm 1 vs §3.0.1"}],"minor_comments":[{"comment":"The loss function is called 'binary crossentropy' in §3.0.6 and 'categorical cross-entropy' in §4. For a single-sigmoid binary output, binary crossentropy is the correct name; please fix.","section":"§3.0.6 / §4"},{"comment":"The table formatting merges healthy-class precision, recall, and F1 into a single cell (e.g., '0.760.96 * 0.85'), making it hard to read. Please separate the columns and explain the asterisk.","section":"Table 2"},{"comment":"Typo: 'calulated' should be 'calculated'.","section":"§3.1"},{"comment":"The removal fraction (5%) is presented as fixed without sensitivity analysis. A short sweep (e.g., 1%, 5%, 10%) would strengthen the claim that this choice is not arbitrary.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's contribution is closer to selective data pruning than to the certified-unlearning literature; the framing as 'machine unlearning' may need toning down or positioning against SISA/certified-removal methods. The heavy reliance on the authors' own prior works for comparison should be made explicit and justified. With the missing controls (random-removal baseline, multi-seed statistics, protocol-matched comparison) the empirical claims would be credible for an applied journal; without them, the central contributions are unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a credible engineering paper with a real efficiency claim — a 0.231M-parameter CNN hitting 90.0% on potato drought stress — but the 'competitive accuracy' headline is not backed by a fair comparison, and the unlearning gain is not backed by a control. The architecture itself is clearly specified and the parameter count is the story. Worth a referee, but the claims need tightening.\n\nWhat's new: the specific hybrid arrangement (MobileNetV2 bottlenecks + DenseNet block + skip connections) and the application of gradient-norm influence scoring to prune low-influence training samples for generalization. Neither is conceptually groundbreaking, but the combination is concrete and reproducible from Algorithm 1. The evaluation design — three scenarios (no aug, aug, aug+unlearning) on a public dataset — is clean and the learning curves and confusion matrices are honestly presented.\n\nSoft spots, in order of severity. First, Table 3 compares accuracies to the authors' own prior papers ([13], [26]) without any statement that those baselines were retrained on the same test split (60 images, 1,135 windows) with the same preprocessing. If the splits differ, the 1.6-point gap to ViT-TL and the 0.7-point gap to DenseNet121 are meaningless. The 15–60x parameter reduction is a separate architectural fact, but 'competitive accuracy' is the load-bearing claim and it's unverified. Second, the machine unlearning improvement (90.0 vs 88.6) has no random-removal baseline. Removing the 5% lowest-influence samples could simply be data reduction; the paper doesn't show it's better than removing a random 5%. Third, minor but real: the loss function is called binary crossentropy in Section 3.0.6 and categorical cross-entropy in Section 4; the initial conv uses 32 filters in the text and 36 in Algorithm 1. No error bars, no repeated runs, no code. These are fixable.\n\nMy take: the efficiency claim is likely solid and useful for edge deployment in precision agriculture. The unlearning story needs a control before it can be believed. The comparison table needs a common protocol or an explicit statement that numbers are quoted from prior publications.\n\nWho this is for: applied ML folks in agriculture who care about model size; not someone looking for methodological novelty. Yes, it deserves a serious referee — the architecture is concrete, the parameter count is checkable, and the unlearning idea is testable. I'd ask for the random-removal control and a clarified comparison before acceptance.","headline":"A credible lightweight-CNN paper whose efficiency claim is real but whose accuracy comparison and unlearning gain are both under-supported.","tokens_in":11969,"tokens_out":2497,"would_cite":false,"duration_ms":24140,"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 a 0.231M-parameter hybrid CNN detects potato drought stress at 90.0% accuracy, and that gradient-guided removal of 5% low-influence training images improves accuracy and reduces missed stressed plants.","keywords":["drought stress identification","lightweight CNN","machine unlearning","gradient influence score","precision agriculture","potato crop","aerial imagery","parameter efficiency"],"falsifier":"Repeat the Aug+MU protocol with a random 5% of training images removed, replicated over several seeds. If random removal gives the same ~90.0% accuracy and similar stressed-class false-negative counts, the gradient-guided selection is not responsible for the gain. A sharper check: remove 2.5%, 7.5%, and 10% and see whether the gain tracks the influence ranking rather than the removal fraction.","tokens_in":11126,"feed_emoji":"🌾","tokens_out":9965,"duration_ms":87338,"temperature":0.7,"pith_summary":"This paper claims that accurate drought-stress identification does not need large, parameter-heavy models. A hybrid CNN combining bottleneck residual blocks, a dense feature-reuse block, and a transition layer reaches 90.0% test accuracy on an aerial potato dataset with only 0.231 million trainable parameters, within 1.6 points of a 14-million-parameter vision transformer baseline. It further claims that scoring each training image by the L2 norm of gradients of the model's prediction with respect to the input, then removing the lowest-scoring 5% and retraining, improves accuracy from 88.6% to 90.0% and lowers false negatives for stressed plants. If true, the practical payoff is real-time drought monitoring on drones and edge devices.","feed_headline":"Compact CNN with 0.23M params detects drought at 90% accuracy","feed_subtitle":"Runs on edge devices while matching models 15 to 60 times larger, and unlearning raises accuracy by 1.4 points.","key_machinery":"Two mechanisms carry the argument. The architecture combines bottleneck residual blocks (1x1 expansion, 3x3 depthwise convolution, 1x1 projection, and an optional skip connection), a four-layer dense block that concatenates every layer's output with all preceding feature maps, and a transition layer that compresses channels with a 1x1 convolution and 2x2 average pooling; global average pooling feeds a 128-unit dense layer and a sigmoid output. The unlearning mechanism is a gradient-norm influence score: after training, each image's score is the L2 norm of the flattened gradient of the model's prediction with respect to that input image; the lowest-scoring 5% of the training set is discarded","core_discovery":"The central claim is twofold. First, a custom lightweight hybrid CNN—bottleneck residual blocks in the style of MobileNetV2, a four-layer dense block in the style of DenseNet, a channel-compressing transition layer, and global average pooling—can classify healthy versus drought-stressed potato patches at 90.0% accuracy with only 0.231M trainable parameters. This sits close to the 91.6% of a 14M-parameter transfer-learned vision transformer and above the 88.7% MobileNet pipeline, with a 15-60x parameter reduction. Second, the paper defines an influence score for each image as the L2 norm of the concatenated gradients of the model's prediction with respect to the input image, removes the 5% of","pith_inferences":["Extension: the unlearning gain is not yet isolated from plain data reduction. The paper lacks a random-removal baseline, so the 1.4-point accuracy gain could be produced by dropping any 5% of the training set.","Extension: the influence score measures sensitivity of the prediction to input pixels, not the sample's contribution to the loss; what the paper calls 'forgetting' may amount to pruning outliers or easy redundant images rather than privacy-style unlearning.","Extension: because the test set comes from a single potato field and season, the 90.0% figure should be checked on other fields, crops, growth stages, and sensor conditions before treating it as a general drought-detection result."],"forward_implications":["Drought-stress screening can run on UAVs and edge devices: 0.231M parameters means a small memory footprint and fast inference, with accuracy within 1.6 points of the best heavy baseline.","Parameter efficiency does not have to come at the cost of unbalanced errors; the reported stressed-class recall and F1 improve with augmentation plus unlearning, reducing missed stressed plants.","Gradient-guided unlearning can be used as a data-cleaning step in agricultural datasets, letting a deployed model adapt by forgetting low-influence or noisy samples.","The architecture recipe—combining bottleneck residuals, a dense block, and a transition layer—is a viable template for resource-limited image classification beyond drought detection."],"supporting_citations":[{"why":"Supplies the potato aerial imagery dataset with healthy and stressed annotated windows used in all training and testing.","marker":"[12]"},{"why":"Provides the MobileNet- and DenseNet-based comparison pipelines and the lightweight explainable CNN that this work builds on.","marker":"[13]"},{"why":"Provides the vision transformer transfer-learning baseline whose 91.6% accuracy and 14M parameters frame the efficiency comparison.","marker":"[26]"},{"why":"Cited as the source of ResNet-style residual connections used in the hybrid blocks.","marker":"[31]"},{"why":"Cited as the source of the DenseNet dense block with feature concatenation.","marker":"[32]"},{"why":"Cited as the source of MobileNet-style depthwise separable convolutions used in the bottleneck blocks.","marker":"[33]"},{"why":"Introduces the sharded retraining-based unlearning approach that motivates retraining on a reduced dataset.","marker":"[36]"},{"why":"Formalizes the notion of removing a deleted sample's influence, the conceptual basis for targeted data removal.","marker":"[37]"},{"why":"Provides gradient-based influence estimation used to score training images and decide which to remove.","marker":"[39]"},{"why":"Publishes the multispectral potato plant image collection that the training patches are drawn from.","marker":"[42]"}],"fun_headline_variants":["0.23M-param CNN detects potato drought at 90% accuracy","Unlearning boosts tiny CNN to 90% drought detection","15x smaller CNN matches big models on drought stress","Tiny CNN + unlearning: drought detection at 90% accuracy","Edge-ready CNN spots drought with 0.23M params"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that deleting the 5% of training images with the lowest gradient-norm influence scores is what improves generalization; the paper does not compare with random deletion, so the measured gain could come from sample reduction or chance rather than from the influence ranking.","fun_headline_variants_meta":{"raw":{"variants":["0.23M-param CNN detects potato drought at 90% accuracy","Unlearning boosts tiny CNN to 90% drought detection","15x smaller CNN matches big models on drought stress","Tiny CNN + unlearning: drought detection at 90% accuracy","Edge-ready CNN spots drought with 0.23M params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000297,"raw_usage":{"total_tokens":1585,"prompt_tokens":797,"completion_tokens":788,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":700}},"tokens_in":541,"tokens_out":788,"duration_ms":8096,"temperature":1.0,"reasoning_tokens":700,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:42:09.850495+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the Aug+MU protocol with a random 5% of training images removed, replicated over several seeds. If random removal gives the same ~90.0% accuracy and similar stressed-class false-negative counts, the gradient-guided selection is not responsible for the gain. A sharper check: remove 2.5%, 7.5%, and 10% and see whether the gain tracks the influence ranking rather than the removal fraction.","supporting_citations":[{"cited_title":"Potatocropstressidentificationinaerialimages using deep learning-based object detection.Agronomy Journal, 113:3991–4002, 2021","cited_arxiv_id":null,"evidence_quote":"Supplies the potato aerial imagery dataset with healthy and stressed annotated windows used in all training and testing."},{"cited_title":"Explainable light-weight deep learning pipeline for improved drought stress identification","cited_arxiv_id":null,"evidence_quote":"Provides the MobileNet- and DenseNet-based comparison pipelines and the lightweight explainable CNN that this work builds on."},{"cited_title":"An explainable vision transformer with transfer learning based efficient drought stress identification.Plant Molecular Biology, 115(4):98, 2025","cited_arxiv_id":null,"evidence_quote":"Provides the vision transformer transfer-learning baseline whose 91.6% accuracy and 14M parameters frame the efficiency comparison."},{"cited_title":"Resnet50","cited_arxiv_id":null,"evidence_quote":"Cited as the source of ResNet-style residual connections used in the hybrid blocks."},{"cited_title":"Machine unlearning","cited_arxiv_id":null,"evidence_quote":"Introduces the sharded retraining-based unlearning approach that motivates retraining on a reduced dataset."},{"cited_title":"Towards making systems forget with machine unlearning","cited_arxiv_id":null,"evidence_quote":"Formalizes the notion of removing a deleted sample's influence, the conceptual basis for targeted data removal."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Publishes the multispectral potato plant image collection that the training patches are drawn from."}],"review_version":1}