Dimensionality reduction for AI based hyperspectral image classification based on XAI
Vladimir Zeljković, Branka Stojanović, Harald Ganster, Aleksandar Nešković
cs.CV · cs.AI
Quick verdict
Paper uses Grad-CAM on a small Conv1D to select key SWIR wavelengths for wood-recycling HSI; results plausible but lacks PCA/random baselines and variance.
Summary
This workshop-style paper proposes using Grad-CAM explanations for dimensionality reduction in hyperspectral image (HSI) classification for wood recycling. Data comes from a SWIR line scanner (960–2550 nm, 288 channels) imaging material on a conveyor belt, covering eight classes (raw wood, three processed-wood types, glass, inert, plastic, paper). A single-layer Conv1D network (kernel size 5, 20 filters), fed 5×5 patches reshaped to 25×288 (64 samples per class), produces Grad-CAM heatmaps whose per-class averages serve as wavelength-importance curves; local maxima become the selected wavelengths. A separate Conv2D classifier (architecture detailed only in an unpublished companion [9]) is then evaluated on three wavelength sets: the single most important wavelength per class, the top four, and all 288. Table 1 shows one wavelength already suffices for some classes (Paper 98.35% vs 99.28% with all channels) and four wavelengths lift most classes (Type 1 processed wood 78.27%→91.34%), while Plastic remains hard (38.94%→74.49%→82.71%). Notably, reduced sets beat the full benchmark in several rows (Inert 89.14% vs 80.24%; Type 2 processed wood 92.15% vs 88.04%), which the authors do not explain. Main limitations: no PCA or random-subset baselines, no variance estimates or test-set sizes, proprietary data, and no code release, leaving the central claim—Grad-CAM selects better wavelengths—unverified.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Grad-CAM heatmaps from a single-layer Conv1D identify the wavelengths most important for classifying each wood-recycling material class (§3.2).
- A single Grad-CAM-selected wavelength per class can achieve near-benchmark accuracy for some classes (Paper: 98.35% vs 99.28% with all 288 channels, Table 1).
- Selecting the top four wavelengths per class yields balanced accuracy improvements across material classes (§4.4).
- Pre-selecting important wavelengths significantly reduces the feature space while keeping classification competitive, lowering acquisition and processing cost (§4.2, §5).
Strengths
Real industrial data and practical motivation
SWIR line-scanner data (960–2550 nm, 288 channels) from a conveyor-belt wood-recycling scenario (§2.1), with nine classes including three processed-wood types — a harder, more applied setting than the usual public HSI benchmarks.
Simple, transparent reduction pipeline
§3.2 fully describes the Grad-CAM workflow — single-layer Conv1D (kernel 5, 20 filters), 5×5 patches reshaped to 25×288, class-level heatmaps averaged over 64 samples, local-maxima wavelength picks — making the core idea easy to follow and cheap to deploy.
Clean three-tier evaluation design
Honest headline cases
Weaknesses
No dimensionality-reduction baselines
No statistical reporting
Reproducibility deferred to an unpublished source
Discussion overreaches the evidence
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a short, workshop-style paper intersecting explainable AI (XAI) and hyperspectral imaging (HSI), self-described as a contribution to a broader research initiative on wood-recycling material characterization. The setting is industrial: a Short-Wave Infrared line scanner (960–2550 nm, 288 channels) images conveyor-belt material streams spanning raw wood, three processed-wood types, glass, inert matter, plastic, and paper (plus background — nine rows in Table 1). The central claim is that Grad-CAM saliency computed on a deliberately tiny Conv1D network identifies each class's important wavelengths, so a downstream Conv2D classifier can run on a drastically reduced wavelength set with little accuracy loss (§3–4). Secondary claims are practical: fewer channels mean cheaper acquisition, lower processing cost, and longer equipment life (§2.1). The authors frame the work as preliminary ("to initiate our preliminary assessments", §3.1), but even at that bar the evaluation leaves the core claim unverified.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is this paper about?
It shows how Grad-CAM, an explainable-AI technique, applied to a small 1D convolutional network can identify which of 288 short-wave-infrared wavelengths matter most for classifying wood-recycling materials, so a classifier can run on only a few selected wavelengths instead of the full spectrum.
How much accuracy is lost when using only a few wavelengths?
Per Table 1, losses are small for several classes — Paper reaches 98.35% with one wavelength vs 99.28% with all 288 — but Plastic falls from 82.71% (all) to 38.94% (one), recovering to 74.49% with four. The top-4 setting keeps most classes near the full-spectrum benchmark, though results vary and reduced sets sometimes score higher than all wavelengths.
Is the code or dataset public?
No. The hyperspectral data was captured with an industrial SWIR line scanner in a wood-recycling setting and is not released, no code repository is mentioned, and the companion paper containing the full classifier details is still unpublished.
How does it compare to PCA for wavelength selection?
It does not. Although the related-work section names PCA as the most common dimensionality-reduction method in hyperspectral imaging, the experiments only compare against using all 288 wavelengths — never against PCA or random wavelength subsets.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–2 (class heatmaps and visual classification maps) cannot be assessed from the extracted text; the qualitative claims of §4.2 rest on them.
- Classifier architecture, data splits, and the patch-size ablation live in companion study [9] (anonymous, unpublished), so the protocol cannot be fully audited from this paper.
- Table 1 lacks per-class test-set sizes, and §4.4's Paper claim conflicts with the table — one of the two may be a source/extraction error.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.