Pith. sign in

REVIEW 3 major objections 4 minor 62 references

Safety-Aware Cascaded Inference for Crop Damage Assessment with Controlled Error Trade-offs

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read The paper shows that a two-stage cascade with a recall-constrained triage gate can hold damaged-crop recall at 0.974 at τ=0.5, cutting missed cases by up to 54%, and that the protection comes from decision structure, not better features.

desk verdict A coherent, unusually self-aware safety-constrained cascade paper whose headline 54% missed-case reduction is not reproducible from its own tables, and whose plot-level split leakage risk is real but unquantified. read the letter →

arxiv 2607.25468 v2 pith:5DCNS75A submitted 2026-07-28 cs.CV

classification cs.CV
keywords cropdamageassessmentcascadedinferencesafety-constrainedlearningasymmetriccostsarchitecturalisolationsmallholderagriculturepicture-basedinsurancedistributionshift
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Picture-based crop insurance treats missing a damaged plot as far costlier than flagging a healthy one for review, but an ordinary classifier has no way to encode that asymmetry at inference time. CascadeCropNet separates triage from diagnosis: a cheap Sentinel model routes suspicious images to a specialist Expert model, with the routing threshold τ chosen to keep recall on damaged crops at or above 0.95. At τ=0.5 the cascade reports Rec-Damaged=0.974 and 13.2% less expert workload, cutting missed damage cases by up to 54% relative to a flat baseline. Under input corruption applied to the Sentinel, the Expert's diagnostics stay unchanged because it always receives clean images; the system escalates to full expert routing rather than making confident errors. The paper argues that safety constraints must be enforced by system structure, not just by the loss function.

What carries the argument

The central object is CascadeCropNet, a two-stage cascade. Stage one, SentinelNet, is a mobile backbone trained with a masked ClassBalancedFocalLoss and checkpoint selection that enforces Rec-Damaged ≥ 0.95; its routing threshold τ is a deployment-time control that navigates a Pareto frontier between expert load ρ and safety recall. Stage two, ExpertNet, only sees Sentinel-routed images, runs at higher resolution, and uses masked supervision so its damage-type head never receives healthy samples. The key identity is architectural isolation: ExpertNet's inputs are invariant to corruption applied to Sentinel inputs, so diagnostic outputs are stable while routing escalates—the system fails towa

What would settle it

Group all images by GPS and timestamp to reconstruct plot identities, then rerun the τ=0.5 comparison with no plot shared across train and test; if Rec-Damaged and the 54% missed-case reduction collapse toward the flat baseline, the headline empirical claim is an artifact of split leakage.

Watch

Extended reading notes

Core claim

The paper's central claim is that a mismatch between training-time risk encoding and deployment-time behavior under distribution shift can be mitigated by decision structure. It proposes a two-stage cascade in which a lightweight Sentinel performs binary health triage under an explicit safety constraint (maximize precision subject to Rec-Damaged ≥ 0.95), and a specialist Expert performs Drought/Weeds diagnosis only on routed samples. This yields Rec-Damaged = 0.974 at τ = 0.5, roughly 54% fewer missed damage cases than a flat baseline, and leaves Expert diagnostic metrics unchanged when corruption is applied to Sentinel inputs. The paper shows through controlled experiments that the represen

Load-bearing premise

The results rest on the assumption that the image-level train/test splits do not hide repeated photographs of the same farm plots; because plot-level IDs are unavailable, independence at the field level cannot be confirmed, and any such leakage would inflate the reported recall and missed-case reduction.

Editorial extensions

If this is right

  • At τ=0.5, the paper's operating point, an insurer processing 10,000 images would see about 1,320 fewer expert reviews at Rec-Damaged 0.974; at τ=0.6 the saving grows to 22.6% but Rec-Damaged falls to 0.929.
  • The 54% reduction in missed damage cases (roughly 148 to 68 on the test set) is the consequence that matters most for insurance: fewer legitimate claims denied.
  • Because the protective effect is architectural, the guarantee is conditional on keeping ExpertNet's preprocessing clean; if the pipeline degrades, Expert metrics collapse (F1-DGT −36% under sensor noise applied directly to Expert inputs).
  • Under aligned evaluation, the flat baseline is only 0.008 F1-macro behind, so the cascade's value is decision structure, not representation quality; researchers should report both end-to-end and decomposed metrics.
  • The threshold τ transfers across same-distribution splits with a maximum recall gap of 0.013, so validation-set calibration without retuning is plausible when deployment conditions match training conditions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: If these results transfer, the same gate-and-escalate design could be applied to other hierarchical, asymmetric-cost decisions—medical image triage, credit underwriting, fraud screening—where a cheap gate routes uncertain cases to a specialist; the paper only tests crop damage.
  • Inference: The paper's own experiments imply that training the Sentinel to be deliberately aware of corruption, or replacing the fixed threshold with entropy-based routing, could preserve the safety floor while recovering some of the efficiency that corruption currently eliminates; the paper lists these as future directions.
  • Inference: Since the stability is architectural rather than learned, a simple stress test for any deployment is to corrupt the Expert's inputs directly; if the pipeline ever lets that happen, the system loses its advantage over a flat model.
  • Inference: The representational gap of +0.008 suggests that further gains on this data will come from better routing and label structure, not from bigger backbones; a testable prediction is that backbone swaps alone will not move Rec-Damaged or F1-Expert much.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes CascadeCropNet, a two-stage cascade for crop damage assessment in picture-based agricultural insurance. A lightweight SentinelNet performs binary health triage with a training-time constraint Rec-Damaged ≥ 0.95, enforced through checkpoint selection and a deployment-time routing threshold τ; samples above τ are escalated to ExpertNet for Drought/Weeds diagnosis. The central claims are that this architecture operationalizes asymmetric error costs, that it reduces missed damage cases by up to 54% relative to a flat baseline at τ=0.5 (Rec-Damaged=0.974), that under evaluation alignment the representational gap versus the flat baseline is only +0.008 F1-macro, and that under Sentinel-input corruption the system exhibits error containment rather than intrinsic robustness. The paper includes seven hypothesis-testing experiments (H1–H4 plus oracle, end-to-end, and both-corrupted analyses), with cautious limitations sections and a clear distinction between intrinsic and routed evaluation.

Significance. If the empirical results hold, the paper would make a useful contribution: it gives a concrete mechanism for enforcing a recall floor in a deployed vision pipeline and provides a clear decomposition of where robustness gains come from (architectural isolation vs. learned invariance). The authors are unusually transparent: they report the intrinsic-versus-deployed gap, the low Rec-Healthy at τ=0.5, the unresolved F1-DGT=0.791, and the fact that the two central evaluations use different splits. The Experiment 2 alignment analysis and Experiment 7 both-corrupted stress test are valuable falsifiable checks. These strengths, however, cannot compensate for two load-bearing problems: the image-level split over repeated field photos creates a potential leakage risk that undermines all reported accuracy numbers, and the headline 54% missed-case reduction mixes validation and test values in a way that is not reproducible from the paper's own tables.

major comments (3)
  1. [Section 3, Table 1] The dataset split is at image level and the paper concedes "plot-level identifiers are not available, ... independence at the field level cannot be confirmed" (Section 3). Under the picture-based insurance protocol, the same plots are photographed repeatedly over the season, so images of the same field can appear in both training and test splits. This can let the model memorize plot-specific appearance and would inflate every reported metric, including Rec-Damaged, F1-Expert, and the missed-case comparison. The limitation is acknowledged but never quantified; there is no sensitivity analysis (e.g., grouping by village/timestamp or removing near-duplicate captures). Because all empirical conclusions in Sections 6.1–6.7 rest on this partition, this is a load-bearing data-integrity concern.
  2. [Section 6.2] The 54% missed-case reduction is not reproducible from the paper's own tables. The calculation uses a cascade Rec-Damaged of 0.974 taken from the validation sweep (Table 4) together with the test-set damaged count (≈2,616), and a baseline Rec-Damaged of 0.943 that is unsourced. On the test set, Table 5 (Clean) reports V11 Rec-Damaged=0.912 and cascade Rec-Damaged=0.969; Table 9 reports cascade test Rec-Damaged=0.978 at τ=0.5; Table 11 implies 0.978. None of these pairs yields 0.974/0.943. The abstract and conclusion present this headline without split provenance. Recompute the missed-case comparison on a single split with a sourced baseline.
  3. [Section 6.7, Experiment 1 (Table 6)] The rejection of H1 is not well-defined. V11 is a flat single-model baseline (Section 4.6), yet Table 6 reports ρ for V11 and says "routing 80.8–95.6% of samples to Expert." A flat model has no second-stage Expert; if ρ is instead the fraction of samples whose damage probability exceeds τ, then "expert load" is not an efficiency metric for V11, and the threshold-calibrated comparison is not a cascade. The conclusion that threshold calibration alone cannot replicate cascade behavior depends on this undefined construct. Please specify what the "Expert" is in Table 6 and how F1-DGT/F1-WED are computed on the "routed" subset, or replace the experiment with a well-defined two-stage baseline.
minor comments (4)
  1. [Section 4.1 vs Section 6.4, Figure 5] The Pareto frontier axes are described inconsistently. Section 4.1 says the frontier has x-axis expert load ρ and y-axis Rec-Damaged; Section 6.4 and the Figure 5 caption say the left panel shows ρ versus Rec-Healthy. Clarify which quantity defines the frontier.
  2. [Table 4] The column header "Reduction" should be defined explicitly as expert-load reduction relative to ρ=1; the values are 1−ρ but this is never stated. Also, use consistent notation for Rec-Damaged/Rec-D and Rec-Healthy/Rec-H throughout the tables and text.
  3. [Section 6.2] The sentence "Rec-Healthy at τ=0.5 is 0.424" does not state that this is the validation-set value. Since Table 5 reports test Clean Rec-Healthy=0.455 and Experiment 6 reports 0.449 on the test set, the split should be made explicit here.
  4. [Tables 5, 9, 11] Test-set Rec-Damaged at τ=0.5 is reported as 0.969 (Table 5, Clean), 0.978 (Table 9), and effectively 0.978 (Table 11, from TP counts). The discrepancies are not explained; if they arise from different evaluation subsets or rounding, state this.

Circularity Check

1 steps flagged · score 6.0 of 10

One headline robustness 'result' is definitional (Expert-metric invariance is the architectural isolation property restated); the Rec-Damaged=0.974 headline is in-sample but is backed by independent test-set tables.

  1. self definitional [Section 4.4 (Architectural Isolation); Section 6.5 (Robustness Under Distribution Shift); Table 5]
    "We define architectural isolation formally as follows: ExpertNet's input distribution is invariant to corruptions applied to Sentinel inputs, under a fixed preprocessing pipeline. ... Under all three conditions, ExpertNet's diagnostic metrics remain stable within measurement precision: F1-Expert=0.877, F1-DGT=0.805, F1-WED=0.949 across Clean, Field, and Sensor conditions without variation."

    The claimed stability is a direct restatement of the definition plus the evaluation protocol: corruption is applied only to Sentinel inputs, ExpertNet always receives clean images of the same truly-damaged sample set, and the model is fixed and deterministic. The paper itself concedes: 'This stability is not a finding about model robustness in any general sense — it is a direct consequence of the isolation property.' Presenting this designed input-invariance as a headline robustness result is a self-definitional 'prediction', not an independent empirical discovery. The non-tautological parts (rho escalation and Expert degradation under direct corruption in Experiments 3 and 7) do provide independent content, which prevents full circularity.

full rationale

The paper's central architectural claim is not wholly circular: Experiments 3 and 7 directly test the isolation mechanism (ExpertNet collapses when its own inputs are corrupted: F1-DGT -36% under Sensor), and the end-to-end test-set tables give independent evidence for similar recall at tau=0.5 (Rec-D 0.969 and 0.978 in Tables 5 and 9). The definitional component is the stability of Expert metrics under Sentinel-only corruption, which is guaranteed by the definition of architectural isolation and by the evaluation protocol, and the paper acknowledges this. The headline Rec-Damaged=0.974 is a validation-set number produced after checkpoint and threshold selection on that same validation split, so the 'satisfies Rec-Damaged >= 0.95' component is enforced by construction rather than independently predicted; however, the independent test-set tables report comparable values, making this a selection/reporting concern rather than a full circular reduction. The 54% missed-case calculation mixes a validation recall with a test-set damaged count and an unsourced baseline recall of 0.943, and the paper itself notes the threshold sweep and stress test come from different splits; I treat that as an internal-consistency problem, not a circularity. The dataset split limitation (Section 3: field-level independence cannot be confirmed) is a data-premise risk, not a circular derivation. No load-bearing self-citation or imported uniqueness theorem was found; references to prior work are external.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claims rest on domain assumptions about insurance cost asymmetry, hierarchical label meaning, clean Expert inputs, routing-shift generalization, plot-level split independence, and label reliability. The free parameters — τ, recall floors, loss/selection weights — are selected on validation or by hand without ablation. No invented physical or conceptual entities are introduced.

free parameters (7)
  • routing threshold τ = primary operating points 0.5, 0.6; sweep 0.3–0.8
    Deployment-time threshold chosen on validation set to maximize Safety Score subject to Rec-Damaged ≥ 0.95; headline recall/efficiency numbers are functions of it.
  • Rec-Damaged safety floor = 0.95
    Design constraint used in checkpoint selection (§4.2); chosen as a 'conservative operational floor' and asserted rather than derived from a cost model.
  • Rec-Healthy secondary floor = 0.40
    Second constraint in the Sentinel Safety Score (§4.2); hand-chosen minimum to avoid a trivial Healthy classifier.
  • Sentinel loss weights = 5 / 1 / 0.1
    Eq. 2 hand-weights the binary health head; no ablation is reported.
  • Expert loss weights = 1 / 3 / 0.1
    Eq. 3; damage-type head weighted 3 with FocalLoss γ=2; no sensitivity analysis.
  • Checkpoint selection weights = V11: 0.6·F1-D + 0.4·F1-G; V14: 0.7·F1-Exp + 0.3·Rec-H
    Selection criteria chosen by the authors; no ablation or stability analysis.
  • Corruption augmentation mix for ExpertNet = 50% clean / 20% blur / 30% colour jitter; RandomNoise σ=0.08 p=10%
    Training augmentation chosen by hand (§4.3); robustness results depend on this recipe.
assumptions (6)
  • domain assumption In picture-based insurance, a missed damage detection (false negative) is categorically more costly than an unnecessary review (false positive).
    Motivates the whole safety constraint; cited to [1,2,10] but not quantified, and the optimal trade-off would depend on the cost ratio.
  • domain assumption Damage type is only semantically meaningful when damage is present; healthy samples have undefined damage-type labels.
    Drove masked supervision and cascade routing (§3, §4.5); true for this label taxonomy but not a universal property.
  • ad hoc to paper At deployment, ExpertNet always receives clean, server-side preprocessed images while Sentinel inputs may be corrupted.
    The robustness stress test in §6.5 applies corruption only to Sentinel; if Expert input quality is not maintained, F1-DGT collapses −36% (Experiment 3, Table 8).
  • ad hoc to paper ExpertNet's representation quality, measured on the full truly-damaged set, transfers to the actual routed subpopulation at deployment.
    Explicitly stated in §5: 'That assumption remains empirically unverified.' End-to-end routing bias is quantified only in Experiment 6.
  • domain assumption Image-level splits with no plot-level IDs are treated as independent train/val/test samples.
    Section 3: 'independence at the field level cannot be confirmed; this may introduce residual correlation across splits due to repeated observations of the same plots over time.' If violated, metrics are optimistic.
  • domain assumption Mixed manual/automatic labels are sufficiently reliable to train and evaluate the system.
    Section 3 documents an in-house automatic labeling algorithm for some images; structured label noise can bias class boundaries and calibration, acknowledged in §3/§7.4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Safety-Aware Cascaded Inference for Crop Damage Assessment with Controlled Error Trade-offs." pith.science (2026). https://pith.science/paper/5DCNS75A

@misc{pith2026260725468,
  author       = {Pith},
  title        = {Pith review of: Safety-Aware Cascaded Inference for Crop Damage Assessment with Controlled Error Trade-offs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5DCNS75A}},
  note         = {Machine review of arXiv:2607.25468}
}
read the original abstract

In picture-based agricultural insurance for smallholder farmers, missed damage detections carry substantially higher cost than false alarms: a farmer who sustained real losses receives no payout, while unnecessary expert review is operationally costly but reversible. Standard multi-class classifiers optimize global accuracy but provide no mechanism to operationalize or control this asymmetric cost structure at inference time. We propose CascadeCropNet, a two-stage cascade architecture calibrated to satisfy a target recall constraint (Rec-Damaged >= 0.95) through threshold selection. A lightweight Sentinel model performs binary health triage; samples exceeding a calibrated damage probability threshold tau are escalated to a specialist Expert model for fine-grained diagnosis. This design provides explicit, deployment-time control over the safety-efficiency trade-off without retraining. Evaluated on the Eyes on the Ground dataset (23,804 images from Kenyan smallholder maize farms), the cascade achieves Rec-Damaged = 0.974 at tau = 0.5, reducing missed damage cases by up to 54% relative to a flat baseline. Under evaluation alignment, the representational gap reduces to +0.008 F1-macro, confirming the contribution is architectural rather than representational. Under input degradation, the system prioritizes escalation over confident misclassification, reflecting error containment through architectural isolation rather than intrinsic model robustness. These results demonstrate that cascade architectures can operationalize safety-oriented decision constraints through calibrated routing in settings where reliability matters more than aggregate accuracy. These properties depend on threshold calibration and deployment conditions and do not constitute guarantees under arbitrary distribution shift.

Figures

Figures reproduced from arXiv: 2607.25468 by the authors.

Figure 1
Figure 1. Class distribution across train, validation, and test splits. [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗
Figure 2
Figure 2. V11 flat baseline training dynamics across 20 epochs. [PITH_FULL_IMAGE:figures/full_fig_p024_2.png] view at source ↗
Figure 3
Figure 3. SentinelNet (V13) training dynamics across 30 epochs. [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: ExpertNet (V14) training dynamics across 25 epochs. [PITH_FULL_IMAGE:figures/full_fig_p025_4.png]
Figure 5
Figure 5. Figure 5: Cascade threshold sweep and Pareto frontier on the validation set ( [PITH_FULL_IMAGE:figures/full_fig_p036_5.png]
Figure 6
Figure 6. Figure 6: Normalized confusion matrices for all three tasks at [PITH_FULL_IMAGE:figures/full_fig_p040_6.png]
Figure 7
Figure 7. Figure 7: Cascade vs. V11 baseline robustness stress test on the test set ( [PITH_FULL_IMAGE:figures/full_fig_p040_7.png]
Figure 5
Figure 5. Figure 5: While thresholding can be applied to flat classifiers [ [PITH_FULL_IMAGE:figures/full_fig_p051_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 13 canonical work pages

  1. [1]

    N. D. Jensen, C. B. Barrett, A. G. Mude, Index insurance quality and basis risk: Evidence from northern Kenya, American Journal of Agricultural Economics 98 (5) (2016) 1450–1469.doi:10.1093/ajae/ aaw046.5

  2. [2]

    B. J. Barnett, O. Mahul, Weather index insurance for agriculture and rural areas in lower-income countries, American Journal of Agricultural 65 Economics 89 (5) (2007) 1241–1247.doi:10.1111/j.1467-8276.2007. 01091.x

  3. [3]

    Elkan, The foundations of cost-sensitive learning, in: Proceedings of10 the Seventeenth International Joint Conference on Artificial Intelligence, 2001, pp

    C. Elkan, The foundations of cost-sensitive learning, in: Proceedings of10 the Seventeenth International Joint Conference on Artificial Intelligence, 2001, pp. 973–978

  4. [4]

    Cannon, J

    A. Cannon, J. Howse, D. Hush, C. Scovel, Learning with the Neyman- Pearson and min-max criteria, Tech. Rep. LA-UR-02-2951, Los Alamos National Laboratory (2002).15

  5. [5]

    Scott, R

    C. Scott, R. Nowak, A Neyman-Pearson approach to statistical learning, IEEE Transactions on Information Theory 51 (11) (2005) 3806–3819. doi:10.1109/TIT.2005.856955

  6. [6]

    Geifman, R

    Y. Geifman, R. El-Yaniv, Selective classification for deep neural networks, in: Advances in Neural Information Processing Systems, Vol. 30, 2017.20

  7. [7]

    Hendrycks, T

    D. Hendrycks, T. Dietterich, Benchmarking neural network robustness to common corruptions and perturbations, in: International Conference on Learning Representations, 2019

  8. [8]

    Taori, A

    R. Taori, A. Dave, V. Shankar, N. Carlini, B. Recht, L. Schmidt, Mea- suring robustness to natural distribution shifts in image classification, in:25 Advances in Neural Information Processing Systems, Vol. 33, 2020, pp. 18583–18599

Show all 62 references
  1. [9]

    Waithaka, B

    L. Waithaka, B. Kramer, K. Hufkens, B. Kivuva, S. Mansabdar, Eyes on the Ground image data, version 1.0, accessed via Radiant MLHub. License: CC-BY-SA-4.0 (2022).doi:10.34911/rdnt.1bs2jw.30 URLhttps://doi.org/10.34911/rdnt.1bs2jw 66

  2. [10]

    Ceballos, B

    F. Ceballos, B. Kramer, M. Robles, The feasibility of picture-based insurance (PBI): Smartphone pictures for crop loss measurement and parametric insurance triggers, Development Engineering 4 (2019) 100042. doi:10.1016/j.deveng.2019.100042.35

  3. [11]

    Ramcharan, K

    A. Ramcharan, K. Baranowski, P. McCloskey, B. Ahmed, J. Legg, D. P. Hughes, Deep learning for image-based cassava disease detection, Fron- tiers in Plant Science 8 (2017) 1852.doi:10.3389/fpls.2017.01852

  4. [12]

    Picon, A

    A. Picon, A. Alvarez-Gila, M. Seitz, A. Ortiz-Barredo, J. Echazarra, A. Johannes, Deep convolutional neural networks for mobile capture40 device-based crop disease classification in the wild, Computers and Electronics in Agriculture 161 (2019) 280–290.doi:10.1016/j.compag. 2018.04.002

  5. [13]

    Ovadia, E

    Y. Ovadia, E. Fertig, J. Ren, Z. Nado, D. Sculley, S. Nowozin, J. V. Dillon, B. Lakshminarayanan, J. Snoek, Can you trust your model’s45 uncertainty? evaluating predictive uncertainty under dataset shift, in: Advances in Neural Information Processing Systems, Vol. 32, 2019

  6. [14]

    S. P. Mohanty, D. P. Hughes, M. Salathé, Using deep learning for image- based plant disease detection, Frontiers in Plant Science 7 (2016) 1419. doi:10.3389/fpls.2016.01419.50

  7. [15]

    D. P. Hughes, M. Salathé, An open access repository of images on plant health to enable the development of mobile disease diagnostics, arXiv preprint arXiv:1511.08060 (2015)

  8. [16]

    J. G. A. Barbedo, Factors influencing the use of deep learning for plant disease recognition, Biosystems Engineering 172 (2018) 84–95. doi:55 10.1016/j.biosystemseng.2018.05.013. 67

  9. [17]

    Atila, M

    U. Atila, M. Uçar, K. Akyol, E. Uçar, Plant leaf disease classification using EfficientNet deep learning model, Ecological Informatics 61 (2021) 101182.doi:10.1016/j.ecoinf.2020.101182

  10. [18]

    Y. Toda, F. Okura, How convolutional neural networks diagnose plant60 disease, Plant Phenomics 2019 (2019) 9237136.doi:10.34133/2019/ 9237136

  11. [19]

    Arsenovic, M

    M. Arsenovic, M. Karanovic, S. Sladojevic, A. Anderla, D. Stefanovic, Solvingcurrentlimitations of deeplearning basedapproaches forplantdis- ease detection, Symmetry 11 (7) (2019) 939.doi:10.3390/sym11070939.65

  12. [20]

    Kamilaris, F

    A. Kamilaris, F. X. Prenafeta-Boldú, Deep learning in agriculture: A survey, Computers and Electronics in Agriculture 147 (2018) 70–90. doi:10.1016/j.compag.2018.02.016

  13. [21]

    Coulibaly, B

    S. Coulibaly, B. Kamsu-Foguem, D. Kamissoko, D. Traore, Deep neural networks with transfer learning in millet crop images, Computers in70 Industry 108 (2019) 115–120.doi:10.1016/j.compind.2019.02.003

  14. [22]

    Sibiya, M

    M. Sibiya, M. Sumbwanyambe, A computational procedure for the recognition and classification of maize leaf diseases out of healthy leaves using convolutional neural networks, AgriEngineering 1 (1) (2019) 119– 131.doi:10.3390/agriengineering1010009.75

  15. [23]

    Brahimi, K

    M. Brahimi, K. Boukhalfa, A. Moussaoui, Deep learning for tomato dis- eases: Classification and symptoms visualization, Applied Artificial Intel- ligence 31 (4) (2017) 299–315.doi:10.1080/08839514.2017.1315516

  16. [24]

    Caruana, Multitask learning, Machine Learning 28 (1) (1997) 41–75

    R. Caruana, Multitask learning, Machine Learning 28 (1) (1997) 41–75. doi:10.1023/A:1007379606734.80 68

  17. [25]

    Kendall, Y

    A. Kendall, Y. Gal, R. Cipolla, Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 7482–7491.doi:10.1109/CVPR.2018.00781

  18. [26]

    S. Liu, E. Johns, A. J. Davison, End-to-end multi-task learning with85 attention, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 1871–1880.doi:10.1109/CVPR.2019. 00197

  19. [27]

    Ruder, An overview of multi-task learning in deep neural networks, arXiv preprint arXiv:1706.05098 (2017).90

    S. Ruder, An overview of multi-task learning in deep neural networks, arXiv preprint arXiv:1706.05098 (2017).90

  20. [28]

    Standley, A

    T. Standley, A. R. Zamir, D. Chen, L. Guibas, J. Malik, S. Savarese, Which tasks should be learned together in multi-task learning?, in: Proceedings of the International Conference on Machine Learning, 2020, pp. 9120–9132

  21. [29]

    K. P. Ferentinos, Deep learning models for plant disease detection and95 diagnosis, Computers and Electronics in Agriculture 145 (2018) 311–318. doi:10.1016/j.compag.2018.01.009

  22. [30]

    Zhang, M

    Z. Zhang, M. R. Sabuncu, Generalized cross entropy loss for training deep neural networks with noisy labels, in: Advances in Neural Information Processing Systems, Vol. 31, 2018.100

  23. [31]

    H. Guo, Y. Mao, R. Zhang, Mixup as locally linear out-of-manifold regularization, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33, 2019, pp. 3714–3722

  24. [32]

    Zadrozny, J

    B. Zadrozny, J. Langford, N. Abe, Cost-sensitive learning by cost- proportionate example weighting, in: Proceedings of the Third IEEE105 69 International Conference on Data Mining, 2003, pp. 435–442. doi: 10.1109/ICDM.2003.1250950

  25. [33]

    H. He, E. A. Garcia, Learning from imbalanced data, IEEE Transactions on Knowledge and Data Engineering 21 (9) (2009) 1263–1284. doi: 10.1109/TKDE.2008.239.110

  26. [34]

    N. V. Chawla, K. W. Bowyer, L. O. Hall, W. P. Kegelmeyer, SMOTE: Synthetic minority over-sampling technique, Journal of Artificial Intelli- gence Research 16 (2002) 321–357.doi:10.1613/jair.953

  27. [35]

    Menardi, N

    G. Menardi, N. Torelli, Training and assessing classification rules with imbalanced data, Data Mining and Knowledge Discovery 28 (1) (2014)115 92–122.doi:10.1007/s10618-012-0295-5

  28. [36]

    B. C. Wallace, K. Small, C. E. Brodley, T. A. Trikalinos, Class imbalance, redux, in: Proceedings of the IEEE International Conference on Data Mining, 2011, pp. 754–763.doi:10.1109/ICDM.2011.33

  29. [37]

    C. K. Chow, An optimum character recognition system using decision120 functions, IRE Transactions on Electronic Computers EC-6 (4) (1957) 247–254.doi:10.1109/TEC.1957.5221465

  30. [38]

    Viola, M

    P. Viola, M. Jones, Rapid object detection using a boosted cascade of simple features, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Vol. 1, 2001, pp. 511–518. doi:10.125 1109/CVPR.2001.990517

  31. [39]

    H. Li, Z. Lin, X. Shen, J. Brandt, G. Hua, A convolutional neural network cascade for face detection, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 5325–5334. doi:10.1109/CVPR.2015.7299170.130 70

  32. [40]

    Graves, Adaptive computation time for recurrent neural networks, arXiv preprint arXiv:1603.08983 (2016)

    A. Graves, Adaptive computation time for recurrent neural networks, arXiv preprint arXiv:1603.08983 (2016)

  33. [41]

    Bolukbasi, J

    T. Bolukbasi, J. Wang, O. Dekel, V. Saligrama, Adaptive neural networks for efficient inference, in: Proceedings of the International Conference on Machine Learning, 2017, pp. 527–536.135

  34. [42]

    Teerapittayanon, B

    S. Teerapittayanon, B. McDanel, H. T. Kung, BranchyNet: Fast inference via early exiting from deep neural networks, in: Proceedings of the International Conference on Pattern Recognition, 2016, pp. 2464–2469. doi:10.1109/ICPR.2016.7900006

  35. [43]

    X. Wang, F. Yu, Z.-Y. Dou, T. Darrell, J. E. Gonzalez, SkipNet: Learning140 dynamic routing in convolutional neural networks, in: Proceedings of the European Conference on Computer Vision, 2018, pp. 409–424.doi: 10.1007/978-3-030-01261-8_25

  36. [44]

    Figurnov, M

    M. Figurnov, M. D. Collins, Y. Zhu, L. Zhang, J. Huang, D. Vetrov, R. Salakhutdinov, Spatially adaptive computation time for residual145 networks, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 1039–1048.doi:10.1109/CVPR.2017. 194

  37. [45]

    Huang, D

    G. Huang, D. Chen, T. Li, F. Wu, L. van der Maaten, K. Q. Weinberger, Multi-scale dense networks for resource efficient image classification, in:150 International Conference on Learning Representations, 2018

  38. [46]

    Panda, A

    P. Panda, A. Sengupta, K. Roy, Conditional deep learning for energy- efficient inference of neural networks, in: Proceedings of the Design, Automation and Test in Europe Conference, 2016, pp. 475–480.doi: 10.3850/9783981537079_0564.155 71

  39. [47]

    Recht, R

    B. Recht, R. Roelofs, L. Schmidt, V. Shankar, Do ImageNet classifiers generalize to ImageNet?, in: Proceedings of the International Conference on Machine Learning, 2019, pp. 5389–5400

  40. [48]

    Geirhos, P

    R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, W. Brendel, ImageNet-trained CNNs are biased towards texture; in-160 creasing shape bias improves accuracy and robustness, in: International Conference on Learning Representations, 2019

  41. [49]

    Shorten, T

    C. Shorten, T. M. Khoshgoftaar, A survey on image data augmentation for deep learning, Journal of Big Data 6 (1) (2019) 60.doi:10.1186/ s40537-019-0197-0.165

  42. [50]

    Minderer, J

    M. Minderer, J. Djolonga, R. Romijnders, F. Hubis, X. Zhai, N. Houlsby, D. Tran, M. Lucic, Revisiting the calibration of modern neural networks, in: Advances in Neural Information Processing Systems, Vol. 34, 2021, pp. 15682–15694

  43. [51]

    M. H. Saleem, J. Potgieter, K. M. Arif, Plant disease detection and170 classification by deep learning, Plants 8 (11) (2019) 468.doi:10.3390/ plants8110468

  44. [52]

    Y. Cui, M. Jia, T.-Y. Lin, Y. Song, S. Belongie, Class-balanced loss based on effective number of samples, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 9268–9277.175 doi:10.1109/CVPR.2019.00949

  45. [53]

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Dollár, Focal loss for dense object detection, in: Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 2980–2988.doi:10.1109/ICCV.2017.324. 72

  46. [54]

    Davis, M

    J. Davis, M. Goadrich, The relationship between precision-recall and180 ROC curves, in: Proceedings of the International Conference on Machine Learning, 2006, pp. 233–240.doi:10.1145/1143844.1143874

  47. [55]

    Howard, M

    A. Howard, M. Sandler, G. Chu, L.-C. Chen, B. Chen, M. Tan, W. Wang, Y. Zhu, R. Pang, V. Vasudevan, Q. V. Le, H. Adam, Searching for MobileNetV3, in: Proceedings of the IEEE International Conference185 on Computer Vision, 2019, pp. 1314–1324. doi:10.1109/ICCV.2019. 00140

  48. [56]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, ImageNet: A large-scale hierarchical image database, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp.190 248–255.doi:10.1109/CVPR.2009.5206848

  49. [57]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, in: International Conference on Learning Representations, 2019

  50. [58]

    Fuentes, S

    A. Fuentes, S. Yoon, S. C. Kim, D. S. Park, A robust deep-learning- based detector for real-time tomato plant diseases and pests recognition,195 Sensors 17 (9) (2017) 2022.doi:10.3390/s17092022

  51. [59]

    D. M. W. Powers, Evaluation: From precision, recall and F-Measure to ROC, informedness, markedness and correlation, Journal of Machine Learning Technologies 2 (1) (2011) 37–63

  52. [60]

    B. J. Barnett, C. B. Barrett, J. R. Skees, Poverty traps and index-based200 risk transfer products, World Development 36 (10) (2008) 1766–1785. doi:10.1016/j.worlddev.2007.10.016. 73

  53. [61]

    M. Tan, Q. V. Le, EfficientNet: Rethinking model scaling for convolu- tional neural networks, in: Proceedings of the International Conference on Machine Learning, 2019, pp. 6105–6114.205

  54. [62]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778.doi:10.1109/CVPR.2016. 90. 74

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.