pith. sign in

arxiv: 2602.10434 · v3 · pith:HAJRRVRDnew · submitted 2026-02-11 · 📡 eess.IV

Benchmarking Deep Learning and Statistical Target Detection Methods for PFM-1 Landmine Detection in UAV Hyperspectral Imagery

Sagar Lekhak , Prasanna Reddy Pulakurthi , Ramesh Bhatta , Emmett J. Ientilucci This is my paper

Pith reviewed 2026-05-16 04:02 UTC · model grok-4.3

classification 📡 eess.IV
keywords landmine detectionhyperspectral imagingUAVtarget detectionspectral neural networkbenchmarkprecision-recallPFM-1
0
0 comments X

The pith

A lightweight spectral neural network achieves the highest average precision for detecting sparse PFM-1 landmine pixels in UAV hyperspectral imagery, outperforming classical detectors under precision-focused metrics.

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

This paper benchmarks four classical statistical target detectors—Spectral Angle Mapper, Matched Filter, Adaptive Cosine Estimator, and Constrained Energy Minimization—against a proposed lightweight Spectral Neural Network with Parametric Mish activations for identifying PFM-1 landmines in UAV-collected hyperspectral data. All methods reach high ROC-AUC on an independent test set, with ACE at 0.989, yet the extreme sparsity of target pixels relative to background makes ROC-AUC potentially misleading. Under precision-recall and average precision evaluation, the Spectral-NN leads, showing why precision-oriented metrics better suit rare-target remote sensing tasks. The release of pixel-level ground truth masks enables reproducible comparisons that could support safer, automated large-area surveys with reduced human risk.

Core claim

While the Adaptive Cosine Estimator reaches the highest ROC-AUC of 0.989, the Spectral Neural Network delivers the highest average precision on the precision-recall curve because target pixels are extremely sparse; this demonstrates that precision-focused evaluation is required for reliable performance assessment in UAV hyperspectral landmine detection.

What carries the argument

Lightweight Spectral Neural Network with Parametric Mish activations, which classifies individual hyperspectral pixels as target or background based on learned spectral signatures.

If this is right

  • Precision-recall metrics must replace or supplement ROC-AUC when targets occupy only a tiny fraction of pixels.
  • Learning-based spectral models can exceed classical detectors in precision for this sparse detection problem.
  • Releasing pixel-level annotations creates a reproducible benchmark that future UAV hyperspectral studies can use directly.
  • Scene-aware evaluation prevents over-optimistic claims when background dominates the data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar lightweight networks may improve detection of other rare spectral signatures, such as specific minerals or vegetation stress, in airborne imagery.
  • Domain adaptation or online fine-tuning would likely be needed before deploying the model on live minefields or different geographic regions.
  • Embedding the detector in real-time UAV flight software could allow adaptive path planning that spends more time over high-probability patches.

Load-bearing premise

Results measured on inert PFM-1 targets in the released VNIR dataset will generalize to live mines, different soils, lighting, or other landmine types.

What would settle it

Retraining or testing the Spectral-NN and ACE on a dataset of live PFM-1 mines in new soil backgrounds and observing that the neural network loses its lead in average precision.

Figures

Figures reproduced from arXiv: 2602.10434 by Emmett J. Ientilucci, Prasanna Reddy Pulakurthi, Ramesh Bhatta, Sagar Lekhak.

Figure 1
Figure 1. Figure 1: Overview of PFM-1 mines (highlighted in red) and other target types [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cropped regions used in this study from the original hyperspectral [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: ROC curves and corresponding AUCs for all algorithms in the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Precision-recall curve and corresponding APs for a) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Aeropoints in the scene appearing as the biggest false alarms for [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
read the original abstract

In recent years, unmanned aerial vehicles (UAVs) equipped with imaging sensors and automated processing algorithms have emerged as a promising tool to accelerate large-area surveys while reducing risk to human operators. Although hyperspectral imaging (HSI) enables material discrimination using spectral signatures, standardized benchmarks for UAV-based landmine detection remain scarce. In this work, we present a systematic benchmark of four classical statistical detection algorithms, including Spectral Angle Mapper (SAM), Matched Filter (MF), Adaptive Cosine Estimator (ACE), and Constrained Energy Minimization (CEM), alongside a proposed lightweight Spectral Neural Network utilizing Parametric Mish activations for PFM-1 landmine detection. We also release pixel-level binary ground truth masks (target/background) to enable standardized, reproducible evaluation. Evaluations were conducted on inert PFM-1 targets across multiple scene crops using a recently released VNIR hyperspectral dataset. Metrics such as receiver operating characteristic (ROC) curve, area under the curve (AUC), precision-recall (PR) curve, and average precision (AP) were used. While all methods achieve high ROC-AUC on an independent test set, the ACE method observes the highest AUC of 0.989. However, because target pixels are extremely sparse relative to background, ROC-AUC alone can be misleading; under precision-focused evaluation (PR and AP), the Spectral-NN outperforms classical detectors, achieving the highest AP. These results emphasize the need for precision-focused evaluation, scene-aware benchmarking, and learning-based spectral models for reliable UAV-based hyperspectral landmine detection. The code and pixel-level annotations will be released.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript benchmarks four classical statistical target detection algorithms (SAM, MF, ACE, CEM) against a proposed lightweight Spectral Neural Network with Parametric Mish activations for PFM-1 landmine detection in UAV hyperspectral imagery. It releases pixel-level ground-truth masks and evaluates all methods on an independent test set from a VNIR dataset, reporting that ACE achieves the highest ROC-AUC (0.989) while the Spectral-NN achieves the highest average precision under PR evaluation.

Significance. If the results hold, the work is significant for establishing a reproducible benchmark and releasing annotations for UAV-based HSI landmine detection. The explicit demonstration that ROC-AUC can mislead on sparse targets while AP favors the learning-based model provides actionable guidance for metric selection in imbalanced detection tasks.

major comments (2)
  1. The section describing the Spectral Neural Network does not provide sufficient detail on architecture (layer count, input/output dimensions), the precise definition of Parametric Mish, loss function, optimizer, learning rate schedule, or training/validation split procedure. These elements are load-bearing for reproducing the reported AP superiority and for determining whether the gain stems from the architecture or from dataset-specific tuning.
  2. Experiments section: the evaluation and claims of reliability for UAV-based detection rest exclusively on inert PFM-1 targets. The manuscript should add explicit discussion of how spectral signatures may shift for live mines (explosive fill, casing corrosion, soil interaction) and whether the trained NN would require retraining or domain adaptation, as this directly affects the operational relevance of the precision-focused ranking.
minor comments (2)
  1. Abstract: the statement that code and annotations 'will be released' should be updated with a concrete repository link or DOI if available at submission time.
  2. Ensure all result tables or figures report the number of independent training runs (if any) and any variance for the Spectral-NN metrics to allow assessment of stability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which have improved the clarity and reproducibility of our work. We address each major comment below and have revised the manuscript accordingly.

read point-by-point responses

  1. Referee: The section describing the Spectral Neural Network does not provide sufficient detail on architecture (layer count, input/output dimensions), the precise definition of Parametric Mish, loss function, optimizer, learning rate schedule, or training/validation split procedure. These elements are load-bearing for reproducing the reported AP superiority and for determining whether the gain stems from the architecture or from dataset-specific tuning.

    Authors: We agree that these implementation details are necessary for full reproducibility. In the revised manuscript we have expanded the Spectral Neural Network section to specify the exact layer count and input/output dimensions, the mathematical definition of Parametric Mish, the loss function (binary cross-entropy), the optimizer (Adam), the learning-rate schedule, and the training/validation split procedure. These additions make clear that the reported AP improvement arises from the proposed architecture rather than undisclosed hyper-parameter choices. revision: yes

  2. Referee: Experiments section: the evaluation and claims of reliability for UAV-based detection rest exclusively on inert PFM-1 targets. The manuscript should add explicit discussion of how spectral signatures may shift for live mines (explosive fill, casing corrosion, soil interaction) and whether the trained NN would require retraining or domain adaptation, as this directly affects the operational relevance of the precision-focused ranking.

    Authors: We acknowledge the distinction between inert and live targets. Although the released dataset contains only inert PFM-1 mines, the revised manuscript now includes an explicit discussion of possible spectral-signature shifts caused by live explosive fill, casing corrosion, and soil interactions. We also note that the trained Spectral-NN may require retraining or domain adaptation for operational live-mine scenarios. This addition places the current precision-focused benchmark in proper operational context without overstating generalizability. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmarking on released ground-truth data

full rationale

The paper performs direct empirical comparison of four classical detectors (SAM, MF, ACE, CEM) and one lightweight Spectral-NN on a released VNIR hyperspectral dataset with pixel-level binary ground-truth masks. All reported metrics (ROC-AUC, PR, AP) are computed on an independent test set of inert PFM-1 targets. No equations, parameter fits, or self-citations are invoked as load-bearing steps in any derivation; the central claims rest on observable performance numbers against external annotations. This is the standard non-circular case for a benchmarking study.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that hyperspectral signatures remain sufficiently distinct for inert PFM-1 mines under the tested conditions, plus the standard assumption that the chosen train/test split is representative. No new physical entities are postulated.

free parameters (1)
  • Spectral-NN weights and Parametric Mish parameters
    Learned during training on the provided dataset; the claim that the NN outperforms classical methods depends on these fitted values.
axioms (1)
  • domain assumption Hyperspectral signatures of PFM-1 landmines are stable enough to serve as reliable targets across scene crops
    Invoked when applying all detectors including the NN to the VNIR dataset.

pith-pipeline@v0.9.0 · 5613 in / 1401 out tokens · 68390 ms · 2026-05-16T04:02:18.158867+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.