eess.IVApr 5, 2026

Non-Destructive Prediction of Fruit Ripeness and Firmness Using Hyperspectral Imaging and Lightweight Machine Learning Models

Authors: Phongsakon Mark Konrad, Casper Kunstmann-Olsen, Jacek Fiutowski, Serkan Ayvaz

Organizations: The Maersk Mc-Kinney Moller Institute, SDU Centre for Industrial Software, University of Southern Denmark, Sønderborg, 6400, Denmark · Mads Clausen Institute, SDU NanoSyd, University of Southern Denmark, Sønderborg, 6400, Denmark

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Post-harvest fruit quality assessment is essential for reducing food waste, yet reliable non-destructive methods typically depend on expensive hyperspectral cameras and computationally intensive deep learning models. These systems typically require GPU resources, large-scale training data, and domain expertise, limiting their feasibility for many real-world agricultural settings. This study systematically evaluates 20 classical machine learning algorithms on hyperspectral imaging data for simultaneous ripeness classification and firmness prediction across five fruit species, using cross-validated experimental design with Bayesian hyperparameter optimization. Data preprocessing strategy, particularly class balancing and spectral transformations, contributes as much to prediction accuracy as algorithm choice. Our results show that tree-based machine learning models can outperform state-of-the-art deep earning models reported in Fruit-HSNet. Moreover, the findings indicate that only three visible-range wavelengths are needed to recover over 94% of full-spectrum accuracy, demonstrating that low-cost multispectral sensors combined with lightweight machine learning models can serve as practical alternatives to expensive hyperspectral cameras and complex deep learning approaches for practical fruit quality sorting.

Figures & tables

Figure 1: Dataset overview. (a) Sample distribution across five fruit types and three ripeness classes. (b) PCA cumulative variance: 18 components capture 95% of variance in the 1,120-dimensional feature space. (c) PCA projection and (d) t-SNE projection colored by fruit type show species-driven cluster structure rather than quality-state separation.
Figure 2: Five spectral preprocessing types (rows) across the 224 bands (400–1000 nm), grouped by fruit type, ripeness state, and firmness class (columns). Shaded bands: mean ±\pm std. Rows: (a–c) raw, (d–f) 1st derivative, (g–i) continuum removed, (j–l) SNV, (m–o) D1(CR). Inter-species variation dominates; class separation is clearest in derivative features near the red edge ( ∼\sim 700 nm).
Figure 3: Six-phase benchmark pipeline. Each phase trains independent ripeness and firmness models, combined via overall accuracy (OA). Dashed arrows indicate configuration propagation between phases.
RkModelFamilyP1 ↑\uparrowP2 ↑\uparrowP2 ConfigP3 ↑\uparrowP3 CIP4 CV (%) ↑\uparrowTime ↓\downarrowSize ↓\downarrow
   OAOA OA[95%] (s)(MB)
1XGBoostBoosting67.467.4SMOTE / No PCA65.6[57.0, 72.7]82.3 ±\pm 5.45.90.72
2HistGradientBoostingBoosting65.268.1SMOTE / No PCA67.4[59.2, 74.6]82.2 ±\pm 5.44.65.05
3LGBMBoosting68.169.2Strat. Resplit / No PCA69.2[61.4, 76.6]82.2 ±\pm 4.94.71.64
4ExtraTreesTree-based71.775.0Strat. Resplit / No PCA75.0[67.6, 81.8]82.2 ±\pm 4.70.22.77
5SVC_RBFSVM56.257.6SMOTE / No PCA56.9[48.2, 64.5]81.1 ±\pm 4.00.65.16
Table 1: Model performance and efficiency across pipeline phases (full-spectrum, 1,120 features). Models sorted by Phase 4 cross-validated overall accuracy. 11 11 11 P3 and P4 rankings correlate at Spearman ρ=0.51\rho=0.51 ( p=0.020p=0.020 ); P4 cross-validated rankings are authoritative. P3 CIs are Wilson 95% intervals ( nn =138). Protocol differences with Fruit-HSNet are documented in Section 3 . CO 2 emissions negligible ( << 0.1 g per model); column omitted. The P2 Config column shows the per-model best Phase 2 configuration; Phase 3 optimization uses a single global best configuration (stratified resplit / no PCA) for all models. PLS-DA P4 CV metrics should be interpreted with caution: a pipeline anomaly produced identical per-fold ripeness and firmness accuracies, likely due to incorrect task separation during cross-validation (see Section 5 ). P4 shows 10-fold stratified cross-validation accuracy (mean ±\pm std). P3 Wilson 95% CI (binomial, nn =138) shown in brackets. Bold = best in column; underline = second-best. Arrows indicate preferred direction.
RankModelAvg. Rank ↓\downarrowMean F1 ↑\uparrowFamily
1ET3.70.792 ±\pm 0.077Tree-based
2SVM-R4.70.787 ±\pm 0.077SVM
3KNN5.40.771 ±\pm 0.061Other
4GBC5.80.760 ±\pm 0.064Boosting
5RF6.20.758 ±\pm 0.063Tree-based
6LGBM6.20.755 ±\pm 0.049Boosting
Table 2: Model rankings from 10-fold cross-validation (Friedman/Nemenyi, α=0.05\alpha=0.05 ). Rank = mean rank across folds (lower is better). Models connected by a bar in Fig. 1 are not significantly different.
ComparisonSummaryddEffect
Friedman (20 classifiers)χ2(19)=144.8\chi^{2}(19)=144.8 , p<10−20p<10^{-20}——
Nemenyi pairs ( α\alpha =0.05)46 of 190 pairs differ——
Firmness >> Ripeness19/20 models; mean gap 25.0 pp1.84Large
(IQR: 21.4–32.8 pp; exception: LDA)   
HSI vs. RGBCohen’s dd from Phase 4 fold F11.17Large
HSI vs. VIS-3Phase 3 OA, 20 models; p=0.46p=0.460.32Small
Table 3: Summary of statistical tests and descriptive comparisons. Effect sizes follow Cohen’s conventions: ∣d∣<0.2|d|<0.2 negligible, 0.20.2 – 0.50.5 small, 0.50.5 – 0.80.8 medium, >0.8>0.8 large. Friedman and Nemenyi tests are appropriate because models share the same CV folds. Firmness vs. ripeness comparisons involve the same 20 models trained on shared data; formal significance tests are not used because the observations are not independent. Phase 3 modality comparisons (HSI vs. VIS-3, HSI vs. RGB, VIS-3 vs. RGB) use paired Wilcoxon signed-rank tests on per-model overall accuracy.
Figure 4: Pipeline performance. (a) Phase progression for all 20 models: connected dots trace each model’s overall accuracy from baseline (P1) through ablation (P2, best config), optimization (P3), and 10-fold cross-validation (P4). Top-5 models highlighted; remaining in gray. The largest gains occur between P1 and P2, confirming that preprocessing contributes more than hyperparameter tuning for competitive models. (b) Training-time efficiency (log scale) vs overall accuracy (Phase 3), colored by algorithmic family. Marker size reflects model storage size. Star markers denote Pareto-optimal models in model size vs accuracy space (6 models: ExtraTrees, RandomForest, XGBoost, GradientBoosting, AdaBoost, DecisionTree), representing the best size-accuracy tradeoff frontier. Tree-based methods achieve the highest accuracy with the shortest training times.
Balancing strategyNo PCAPCAΔ\Delta (pp)
Stratified resplit †59.08 ±\pm 9.2657.41 ±\pm 4.57+1.67
SMOTE57.14 ±\pm 10.6153.84 ±\pm 7.18+3.30
Original unbalanced58.32 ±\pm 9.3856.29 ±\pm 6.59+2.03
Random oversampling57.01 ±\pm 9.8558.12 ±\pm 4.47−- 1.11
Random undersampling46.56 ±\pm 8.9340.13 ±\pm 7.47+6.43
Overall mean55.6253.16+2.46
Table 4: Phase 2 ablation: mean ±\pm SD overall accuracy (%) ↑\uparrow across 20 models by balancing strategy and PCA setting. Δ\Delta = no-PCA minus PCA difference. Best strategy in bold. Large SDs (4–11 pp) reflect substantial model-to-model variance within each configuration. † Stratified resplit replaces the training set ( nn =414 vs nn =381); strategies below the rule augment the original training set.
Figure 5: Phase 2 preprocessing ablation across 200 configurations. Overall accuracy (%) for 20 models (rows, sorted by best OA) across 10 preprocessing configurations (5 balancing strategies ×\times 2 PCA settings). Vertical white lines separate balancing strategies; within each pair, the left column is without PCA and the right is with PCA. The stratified resplit / no PCA column consistently achieves the highest values.
  Full-spectrumReduced-band OA ↑\uparrow 
RkModelRA ↑\uparrowFA ↑\uparrowOA ↑\uparrowVIS-3RGBΔ\Delta OA
1ExtraTrees63.087.075.071.072.1+2.9
2RandomForest60.986.273.667.871.4+2.2
3LGBM58.080.469.268.168.8+0.4
4HistGradBoosting54.380.467.467.469.9−- 2.5
5XGBoost52.279.065.668.568.5−- 2.9
Table 5: Task asymmetry and modality comparison of Phase 3 models. RA = ripeness accuracy, FA = firmness accuracy, OA = overall accuracy. Full-spectrum has 1,120 features; VIS-3 and RGB each use 15 features. Δ\Delta OA = Full OA −- max(VIS-3, RGB).
Figure 6: Task asymmetry, modality comparison, and spectral importance for Phase 3 optimized models. (a) Task asymmetry: ripeness (pink) vs firmness (green) accuracy per model, revealing a persistent gap across all 20 algorithms. (b) Modality comparison: full-spectrum HSI (blue) vs VIS-3 subset (orange) overall accuracy, with the Fruit-HSNet benchmark (70.73%) shown as a dashed line. (c, d) Spectral band importance (rolling mean ±\pm std) for ExtraTrees across 400–1000 nm; shaded regions highlight biologically meaningful zones (pigment changes, red-edge chlorophyll absorption, NIR cellular/water signatures). Ripeness importance concentrates in visible wavelengths (c), while firmness importance shifts toward NIR features (d).
Fruit ( nn )ExtraTreesRandomForestLightGBMHistGradBoostingXGBoost
Avocado (170)67.965.444.448.138.3
Kaki (68)41.741.733.341.741.7
Kiwi (162)58.354.262.570.850.0
Mango (68)66.775.041.766.750.0
Papaya (51)33.333.333.333.344.4
Table 6: Per-fruit ripeness accuracy (%) for the top-five models (Phase 4 single-split holdout, nn =138; ranked by Phase 3 overall accuracy). Bold = best per fruit; underline = second-best.
TaskClassPrecisionRecallF1
RipenessOverripe0.5610.6960.621
RipenessPerfect0.5690.6380.602
RipenessUnripe ∗1.0000.4710.640
Firmness groupRipeness accuracy
 Soft ( nn =27)0.741
 Medium ( nn =68)0.544
Table 7: ExtraTrees Phase 4 single-split holdout analysis ( nn =138). Top : ripeness per-class metrics from the confusion matrix (Fig. 7 a). Bottom : ripeness accuracy stratified by firmness group (4-bin analysis on the original unbalanced split). Extreme-firmness groups achieve higher ripeness accuracy than intermediate groups. ∗ Unripe precision (1.000) with 47.1% recall means the model rarely predicts unripe; most unripe fruit are misclassified as perfect (41.2%) or overripe (11.8%).
Figure 7: Classification patterns for the best model (ExtraTrees, single-split holdout, nn =138). (a) Normalized ripeness confusion matrix showing primary confusion between adjacent maturation states. (b) Ripeness accuracy stratified by firmness group (4-bin); extreme-firmness groups achieve higher ripeness accuracy than intermediate groups.
Figure 8: Cross-validation stability (Phase 4, 10-fold stratified CV). (a) Overall accuracy with 95% confidence intervals (mean ±\pm t9,0.025t_{9,0.025} ⋅\cdot std, where t9,0.025≈2.262t_{9,0.025}\approx 2.262 ) for all 20 models, sorted by mean. (b) Task-separated view: ripeness vs firmness accuracy with CIs. Firmness predictions are consistently more accurate and stable.
StrategyData StateRipe ↑\uparrowFirm ↑\uparrowOA ↑\uparrowΔ\Delta ET
Hard VotingOriginal52.985.569.2-5.8
Soft VotingOriginal50.084.867.4-7.6
StackingOriginal31.275.453.3-21.7
BlendingOriginal49.372.560.9-14.1
Hard VotingStrat. Resplit57.284.170.7-4.3
Soft VotingStrat. Resplit58.083.370.7-4.3
Table 8: Ensemble performance comparison. All ensembles use the top-5 models from Phase 3. Δ\Delta ET shows difference from best single model (ExtraTrees, 75.0%).
Figure 9: Ensemble performance decomposition by task. All eight ensemble configurations maintain high firmness accuracy ( >> 72%) but catastrophically degrade ripeness accuracy (stacking drops to 31.2%). Dashed line: ExtraTrees single-model baseline (75.0%); dotted line: Fruit-HSNet benchmark (70.73%).
Figure 10: Feature group ablation: relative performance drop (%) when each spectral transformation group (224 bands) is removed. (a) Ripeness: first derivatives are most critical for LightGBM (8.3% drop); other groups are redundant or harmful for ExtraTrees. (b) Firmness: continuum-removed features dominate ExtraTrees (16.6% drop), while LightGBM shows uniform sensitivity across groups ( ≤\leq 5.0%). Positive values indicate removing the group hurts performance; negative values indicate removal improves it. Task- and model-dependent contributions explain why PCA, which compresses all groups uniformly, degrades classification.
Figure 11: Preprocessing gain vs optimization gain for all 20 models. Each model’s accuracy improvement is decomposed into preprocessing gain (best Phase 2 config minus Phase 1 baseline, teal) and optimization gain (Phase 3 minus best Phase 2, purple). Models sorted by preprocessing gain. Dotted lines indicate means. For most models, preprocessing contributes the dominant share of improvement; optimization gain is near-zero or negative for many.
Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: Critical difference diagram from Nemenyi post-hoc test ( α=0.05\alpha=0.05 ) on 20-model ×\times 10-fold overall F1 scores. Models connected by a horizontal bar are not significantly different. Models colored by algorithm family.
Figure 2: (a) Per-fold overall F1 score distributions from 10-fold stratified cross-validation, sorted by mean F1. Box plots show median and interquartile range; individual fold scores are overlaid as jittered points. Dashed line marks the Fruit-HSNet benchmark (70.7%). (b) Per-model Cohen’s dd effect sizes for the HSI vs. VIS-3 subset modality comparison, with 95% confidence intervals. Shaded bands indicate negligible ( ∣d∣<0.2|d|<0.2 ), small ( 0.20.2 – 0.50.5 ), and medium ( 0.50.5 – 0.80.8 ) effect regions. Models colored by algorithm family.
Figure 3: Per-fruit overall accuracy for the top-five models (Phase 4). Avocado and mango show highest performance; papaya consistently underperforms due to smallest sample size ( nn =51). Performance variation across species highlights deployment considerations for multi-fruit sorting systems.
Abbr.ModelHPFILib.Optuna Search Space
ETExtra Trees5✓sklearnn_est ∈\in [50,500], depth ∈\in [3,30], split ∈\in [2,20], leaf ∈\in [1,20], crit ∈\in {gini,ent}
RFRandom Forest6✓sklearnn_est ∈\in [50,500], depth ∈\in [3,30], split ∈\in [2,20], leaf ∈\in [1,20], feat ∈\in {sqrt,log2, ∅\emptyset }, crit ∈\in {gini,ent}
DTDecision Tree4✓sklearndepth ∈\in [3,30], split ∈\in [2,20], leaf ∈\in [1,20], crit ∈\in {gini,ent}
XGBXGBoost9✓xgboostn_est ∈\in [50,500], lr ∈\in [.01,.3], depth ∈\in [3,10], child_w ∈\in [1,10], γ\gamma ∈\in [0,1], sub ∈\in [.5,1], col ∈\in [.5,1], α\alpha ∈\in [0,1], λ\lambda ∈\in [0,1]
LGBMLightGBM8✓lightgbmn_est ∈\in [50,500], lr ∈\in [.01,.3], depth ∈\in [3,15], leaves ∈\in [10,300], child ∈\in [5,100], α\alpha ∈\in [0,1], λ\lambda ∈\in [0,1], col ∈\in [.5,1]
GBCGradient Boosting5✓sklearnn_est ∈\in [50,500], lr ∈\in [.01,.3], depth ∈\in [3,10], split ∈\in [2,20], leaf ∈\in [1,20]
Appendix
Table 1: Overview of 20 evaluated models with Optuna hyperparameter search spaces. HP = number of tuned hyperparameters; FI = native feature importance. All ranges define the Bayesian optimization bounds (100 trials per model).
RkModelBest Hyperparameters
1ExtraTreesn_estimators=277, max_depth=14, min_samples_split=6, min_samples_leaf=1, criterion=gini
2RandomForestn_estimators=285, max_depth=10, min_samples_split=5, min_samples_leaf=2, max_features=log2, criterion=entropy
3LGBMn_estimators=470, learning_rate=0.2874, max_depth=14, num_leaves=10, min_child_samples=90, reg_alpha=0.2373, reg_lambda=0.4189, colsample_bytree=0.6112
4HistGradientBoostinglearning_rate=0.1795, max_iter=123, max_depth=5, min_samples_leaf=41, l2_regularization=0.5475
5XGBoostn_estimators=400, learning_rate=0.1357, max_depth=10, min_child_weight=8, gamma=0.1722, subsample=0.7181, colsample_bytree=0.5824, reg_alpha=0.3973, reg_lambda=0.5746
6KNeighborsn_neighbors=3, weights=distance, algorithm=ball_tree, p=1
Appendix
Table 2: Best hyperparameters found by Optuna (100 Bayesian TPE trials per model). Parameters listed after stripping the classifier__ prefix. Models ranked by Phase 3 overall accuracy.
ComponentSpecification
Hardware
ProcessorApple M4 (10-core CPU: 4 performance + 6 efficiency)
Memory24 GB unified LPDDR5X
StorageSSD
GPUNone (CPU-only training)
Software Environment
Appendix
Table 3: Hardware and Software Environment Specifications. All 20 models were trained on CPU only; no GPU acceleration was used.

Explore similar work

CardsList
Aug 2, 2026

Fruit-HSNet: A Machine Learning Approach for Hyperspectral Image-Based Fruit Ripeness Prediction

Fruit ripeness prediction (FRP) is a classification-based agricultural computer vision task that has attracted much attention, thanks to its wide-ranging advantages in agriculture field for both pre-harvest and post-harvest management. Accurate and timely FRP can be achieved using machine/deep learning-based hyperspectral image classification techniques. However, challenges including the limited availability of labeled data and the lack of robust methods generalizable to various hyperspectral cameras and fruit types can compromise the effectiveness of hyperspectral image-based FRP. Addressing these challenges, this paper introduces Fruit-HSNet, a machine learning architecture specifically designed for hyperspectral classification of fruit ripeness. Fruit-HSNet incorporates a spatio-spectral feature extraction module based on Fourier Transform and central pixel spectral signature followed by learnable feature fusion and a classifier optimized for ripeness classification. The proposed architecture was evaluated using the DeepHS Fruit dataset, the largest publicly available labeled real-world hyperspectral dataset for predicting fruit ripeness, which includes five different types of fruits-avocado, kiwi, mango, kaki, and papaya-captured with three distinct hyperspectral cameras at various stages of ripeness. Experimental results highlight that Fruit-HSNet substantially outperforms existing deep learning methods, from baseline to state-of-the-art models, with improvements of 12%, achieving a new state-of-the-art overall accuracy of 70.73%.

Ahmed Baha Ben Jmaa, Faten Chaieb, Anna Fabijańska

1Efrei Research Lab, Paris Panthéon-Assas University, Paris, France · Institute of Applied Computer Science, Lodz University of Technology, Łódź, Poland

Agricultural Image AnalysisHyperspectral Image Classification
AskPDF
Jun 24, 2026

Predicting Fruit Quality with a Hybrid Machine Learning and Image Processing Approach

Fruit spoilage is a significant issue in agriculture, leading to substantial economic losses. Addressing this, our study introduces a hybrid approach combining image processing and deep learning to assess fruit freshness. We developed an image processing algorithm that quantifies spoilage on a scale from 0 (fully fresh) to 100 (fully rotten). Alongside, we trained a convolutional neural network (CNN) to perform binary classification (fresh or rotten) using a large dataset of fruit images. The outcomes of both methods were synthesized using logistic regression to enhance the accuracy of freshness predictions. Subsequently, this logistic regression model was utilized to enable the image processing algorithm to provide binary classification based on its percentage output, thus eliminating the need for the CNN in real-time applications. Our approach, which does not require high computational resources, achieved real-time performance and was validated with over 90% accuracy on a dataset comprising apples and oranges. The primary limitation lies in the requirement for fruits to be isolated on a background that must be either white or transparent, suggesting future improvements could include advanced segmentation models to automate background removal. This study's results highlight the potential of integrating simple image processing techniques with machine learning to provide practical solutions in the agricultural sector.

Amir Reza Hashemi, Shahram Amiri

Department of Computer Science, Science and Research Branch, Islamic Azad University, Tehran, Iran · Department of Decision and Information Sciences, School of Business Administration, Stetson University, DeLand, Florida, USA

Agricultural Image AnalysisImage Classification
AskPDF
Oct 9, 2026

Selective Spatio-Temporal Fruit Inspector (S2^2FI): A Mamba-based Framework for Rotational Fruit Quality Assessment

The global fruit and vegetable industry confronts severe post-harvest sorting challenges for small and medium-scale growers. Large intelligent sorting systems are too costly, while manual sorting suffers from low efficiency, inconsistent standards, and high labor demand. To bridge this gap between expensive equipment and inefficient manual work, this study develops a low-cost, high-efficiency, lightweight fruit sorting system driven by deep learning, suitable for small and medium growers. We propose the Selective Spatio-Temporal Fruit Inspector (S2^2FI) architecture, which combines a 3D convolutional encoder with a Mamba-based backbone.It efficiently extracts spatiotemporal features, adaptively weighs informative frames during fruit rotation, and models long-range dependencies with linear complexity, reducing motion blur and occlusion. The hardware uses a dual-conveyor and curved rail design, supporting household power, fast deployment, and lightweight embedded deployment to improve affordability. We also built a conveyor-based fruit quality video dataset with 6,460 samples. Experiments on the held-out test set show that our method achieves the best trade-off among accuracy, efficiency, and practicality, outperforming the compared baselines on the test set, achieving an excellent balance of accuracy, speed, and practicality for real-world sorting.

Hao Tian, Jie Hu, Guoqi Shan +3

School of Artificial Intelligence, Chuzhou University, Anhui, China · School of Computer and Information Engineering, Chuzhou University, China

Agricultural Image Analysis
AskPDF