IAES Inter national J our nal of Articial Intelligence (IJ-AI) V ol. 15, No. 5, October 2026, pp. 4512 ∼ 4522 ISSN: 2252-8938, DOI: 10.11591/ijai.v15.i5.pp4512-4522 ❒ 4512 Deep multi-block con v olutional neural netw ork f or quantitati v e near -infrar ed spectral analysis Hilmy Abidzar T awakal 1,2 , Y ohanes Aris Purwanto 3 , Sony Hartono W ijaya 1 , Shelvie Nidya Neyman 1 1 School of Data Science, Mathematics, and Informatics, IPB Uni v ersity , Bogor , Indonesia 2 Department of Informatics, Nurul Fikri Colle ge of T echnology , Depok, Indonesia 3 Di vision of Biosystem Engineering, F aculty of Engineering and T echnology , IPB Uni v ersity , Bogor , Indonesia Article Inf o Article history: Recei v ed Jul 16, 2025 Re vised Jul 17, 2026 Accepted Aug 12, 2026 K eyw ords: Deep learning Multi-block con v olutional neural netw ork Near -infrared spectra Preprocessing Rice quality Spectral analysis ABSTRA CT Near -infrared (NIR) spectroscop y is a po werful analytical technique, b ut its ef fecti v eness is often hi ndered by inst rumental noise and a depe ndenc y on e xpert-dri v en preprocessing and calibration. This study introduces a no v el, intelligent deep learning architecture, a deep multi-block con v olutional neural netw ork (CNN) that addresses these challenges by performing end-to-end quantitati v e analysis directly on ra w spectral data. By se gmenti ng the NIR spectrum into multiple blocks and processing them in parallel, the model mimics multi-scale feature learning, automatically identifying and le v eraging localized spectral re gions rele v ant to tar get properties. This approach eliminates the need for manual preprocessing, a signicant step to w ard creating autonomous analytical systems. This st udy e v aluated the model on four rice quality datasets (moisture, protein, f atty acids, and total f at) and benchmark ed it ag ainst traditional partial least squares re gression (PLSR) and other deep learning models. The proposed multi-block CNN consistently achie v ed superior performance, demonstrating its ability to handle ra w spectral data rob ustly . F or instance , it achie v ed coef cient of determination ( R 2 ) v alues of 0 . 996 ± 0 . 003 for total f at and 0 . 997 ± 0 . 001 for f atty acids. Statistical tests conrm that the model’ s performance on ra w data is not signicantly dif ferent from its performance wit h optimized preprocessing, v alidating its potential for real-w orld articial intelligence applications in agriculture, food quality monitoring, and biosystems engineering where rapid, automated analysis is critical. This is an open access article under the CC BY -SA license . Corresponding A uthor: Y ohanes Aris Purw anto Di vision of Biosystem Engineering, F aculty of Agricultural T echnology , IPB Uni v ersity Bogor , Indonesia Email: arispurw anto@apps.ipb .ac.id 1. INTR ODUCTION Near -infrared (NIR) spectroscop y of fers rapid, non-destructi v e, and safe chemical analysis, making it in v aluable in elds lik e agriculture and food science [1]–[3]. Ho we v er , ra w NIR spectra are frequently corrupted by noise, baseline drift, and light scattering ef fects, which can obscure the underlying chemical information [4]. T raditionally , this chall enge is addressed through a multi-step pipeline that in v olv es e xpert-dri v en preprocessing techniques follo wed by the construction of a calibration model, typically using methods s uch as parti al least squares re gression (PLSR). The success of this pipeline is highly dependent on human e xpertise, and an inappropriate choice of preprocessing can e v en de grade model performance more than J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4513 using no preprocessing at all [5]. Deep learning has emer ged as a po werful alternati v e, capable of learning hierarchical features directl y from data and thus automating the feature engineering process [6], a trend that has been e xtensi v ely re vie wed in recent literature [7]. This has led to a recent sur ge of interest in the eld, alongside a critical discussion of its benets and potential pitf alls [8]. Studies ha v e sho wn that con v olutional neural netw ork (CNN) can ef fecti v ely replace traditional preprocessing steps and achie v e promising results in NIR spectral analysis [9]. Se v eral researchers ha v e successfully de v eloped CNN architectures for NIR data that perform better on ra w spectra than on preprocessed data [5], [10]–[13]. Ho we v er , an important research g ap remains: e xisting CNN-based spectral models often treat the spectrum as a monolithi c whole and f ail to capture localized spectral rele v ance ef fecti v ely , e v en though chemical information in NIR spectra is not uniformly distrib uted. Specic or g anic compounds correlate strongly with narro w , distinct w a v elength re gions, while other re gions may contain only noise. T o address this g ap, this study proposes a no v el deep multi-block CNN architecture that se gments the NIR spectrum and learns localized features in parallel, b uilding on recent adv ancements in parallel input models for spectral analysis [14]. Inspired by the inception layer concept and attention mechanisms [15], the architecture functions as an intelligent system that autonomously identies and weighs the importance of dif ferent spectral re gions. This study’ s primary objecti v e is to de v elop and v alidate a preprocessing-free, end-to-end CNN architec ture for rob ust NIR spe ctral re gression. Thi s study h ypothesizes that by enabling the model to focus on localized spectral se gments, it can enhance its ability to e xtract meaningful features, reduce the impact of noise, and b uild more accurate and generalizable calibration models directly from ra w data. This w ork adv ances intelligent systems by demonstrating ho w a modular , block-based architecture can enable automated learning directly from ra w sensor data, highlighting its potential for real-w orld articial intelligence applications in quality control and process monitoring. 2. MA TERIALS AND METHODS 2.1. Materials This study used NIR spectra of rice to estim ate essential properties: moisture, protein, f atty aci ds, and total f at, which are k e y indicators of rice quality [16]–[18]. Rice samples are placed in a glass container and sealed with dark paper to minimize en vironmental interference, such as ambient light, during measurement. Rice samples were stored in a container and measured monthly o v er a 5-month period. NIR spectra were collected using a spectrometer with a w a v elength range of 1,350 to 2,500 nm, resulting in 257 spectral features per sample. F or calibration, reference v alues for rice properties were obtained from certied laboratories. 2.2. Methods 2.2.1. Pr epr ocessing methods T o establish a performance benchmark, this study compared models trained on ra w data with those trained on data preprocessed with v arious con v entional techniques, including scatter correction (standard normal v ariate (SNV), multiplicati v e scatter correction (MSC), e xtended multiplicati v e scatter correction (EMSC)), asymmetric least squares baseline correction (ALS), and smoothing (Sa vitzk y-Golay). These methods address common issues in NIR s pectra, such as baseline drift, noise, and scattering ef fects. The y were selected based on prior research demonstrating their ef fecti v eness [19]–[21]. 2.2.2. Baseline calibration models T o e v aluate the proposed method, this study benchmark ed its performance ag ainst a di v erse set of traditional and deep learning-based models. This study included PLSR as a representati v e of the standard linear chemometric methods widely used in spect ral data analysis. T o pro vide a rationale for using more comple x architectures, tw o baseline models were implemented: a simple deep neural netw ork (DNN) and a simple CNN. The simple DNN consisted of tw o hidden layers with 64 neurons each using rectied linear unit (ReLU) acti v ation, designed to assess performance without con v olutional feature e xtraction. The simple CNN featured a Con v1D layer with 64 lters , fol lo wed by MaxP o ol ing1D, a se cond Con v1D l ayer with 128 lters, and a nal dense output layer . Finally , the proposed multi-block architecture w as compared with state-of-the-art models in the literature, specically DeepSpectra [5] and spectral-visual geometry group (SpectraV GG) [13], which represent more adv anced, deeper CNN architectures. Deep multi-bloc k con volutional neur al network for quantitative near -infr ar ed ... (Hilmy Abidzar T awakal) Evaluation Warning : The document was created with Spire.PDF for Python.
4514 ❒ ISSN: 2252-8938 2.2.3. Pr oposed multi-block con v olutional neural netw ork ar chitectur e The proposed multi-block CNN is an intelligent architecture designed for end-to-end analysis, inte grating feature e xtraction and re gression to operate directly on ra w spectra. The design in this study is inspired by the ef fecti v eness of parallel processing of spectral se gments, a concept that has pro v en successful in recent chemometric models [14]. As illustrated in the conceptual diagram in Figure 1, the ar chitecture’ s logic comprises three k e y stages: spectral se gmentation, parallel feature e xtraction, and feature fusion for re gression. This gure sho ws the o w from ra w spectra → block se gmentation → parallel CNN blocks → concatenation and dense layers → output. i) Block se gmentation strate gy: the input spectrum (1 × 257 v ector) is rst di vided into B contiguous, non-o v erlapping blocks of equal width. This se gmentation allo ws the model to learn localized features in parallel, since dif ferent spectral re gions encode distinct chemical information. The number of blocks, B , is a k e y h yperparameter determined during optimization. ii) P arallel CNN feature e xtraction blocks: each spectral block is processed in parallel by an identical CNN sub-netw ork designed to act as an automated feature e xtractor for its assigned re gion. Each sub-net w ork consists of a 1D con v olutional layer with a ReLU acti v ation function, follo wed by a MaxPooling1D layer to reduce dimensionality . This is follo wed by a second 1D con v olutional layer with ReLU act i v ation and an A v eragePooling1D layer , concluding with a atten layer to produce the block’ s specic fea ture v ector . The optimal number of lters and k ernel sizes for these con v olutional layers are deter mined through h yperparameter optimization. iii) Feature fusion and re gression: the attened feature v ectors from all B blocks are concatenated into a single, comprehensi v e feature v ector . This v ector is then processed by a deep, dense netw ork (compri sing multiple fully connected layers with ReLU acti v ation) that models the non-linear relationship between the spectral features and the tar get property . Batch normalization and dropout are applied to enhance generalization before the nal single-neuron output layer . Figure 1. Conceptual diagram of the proposed multi-block CNN architecture 2.2.4. Intelligent h yper parameter optimization and training T o autonomously det ermine the optimal model conguration, this study emplo yed an i ntelligent h yperparameter optimization strate gy using the Optuna frame w ork [22]. This automated approach is a practical implementation for tuni ng deep spectral models, as detailed in recent hands-on tutorials for the chemometrics community [23], [24]. The tree-structured P arzen estimator (TPE) algorithm w as used to ef cient ly search the h yperparameter space o v er 100 trials, with the objecti v e of maximizing the v alidation coef cient of determination ( R 2 ) score. The search space w as dened to e xplore k e y architectural and training parameters, including the number of blocks ( B ∈ { 2 , 4 , 8 } ), the number of lters in con v olutional layers ( l ay er s ∈ [16 , 64] ), k ernel sizes ( k er nel ∈ { 3 , 5 , 7 } ), the number of neurons in dense layers ( node ∈ [8 , 32] ), the dropout rate ( do ∈ [0 . 1 , 0 . 5] ), and the learning rate ( l r ∈ [10 − 4 , 10 − 2 ] ). The optimal h yperparameters for the multi-block CNN were determined using Optuna, yielding B = 8 blocks. Each block consists of tw o con v olutional layers: the rst with 32 lters (k ernel size 5) and the second with 16 lters (k ernel size 5). The subsequent dense layers contain 16, 8, and 4 neurons, respecti v ely . T o pre v ent o v ertting, a dropout rate of 0.2 w as applied, and the model w as trained with a learning rate of 0.001. Int J Artif Intell, V ol. 15, No. 5, October 2026: 4512–4522 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4515 2.2.5. Experimental setup and e v aluation T o ensure rigorous and unbiased performance estimation, we discarded the traditional train-test spl it in f a v or of a nested cross-v alidation (Nested CV) protocol. The Nested CV consisted of an outer loop (5 folds) dedicated strictly to nal model e v aluation and an inner loop (3 folds) dedicated to h yperparameter tuning. This guarantees that the h yperparameters are optimized wi thout an y data leakage from the test set. All reported performance metrics represent the mean ± standard de viation (SD) across the 5 outer test folds. Model performance w as e v aluated using six k e y metrics: R 2 , mean absolute error (MAE), mean absolute percentage error (MAPE), cross-v alidation root mean squared error ( RMSE cv ), prediction root mean squared error ( RMSE p ), rat io of performance to de viation (RPD), and ratio of error to range (RER). The RMSE cv is calculated during the inner tuning loop to rank h yperparameters, whereas RMSE p measures the true generalization error on the held-out outer folds. In chemometrics, an RPD > 3 is generally considered acceptable for s creening purposes, and an RPD > 5 is considered suit able for quality control. T o ensure methodological transparenc y and f acilitate reproducibility , all e xperimental code, h yperparameter search spaces, and model weights ha v e been documented. Importantly , the test set in the outer loop of the nested cross-v alidation w as strictly held out and completely isolated from the h yperparameter search process. Scripts for reproducing the results and generating synthetic data schemas (to protect proprietary datasets) are a v ailable upon request. 3. RESUL TS AND DISCUSSION 3.1. Model perf ormance on raw spectra First, this study e v aluated all models on ra w spectral data to establish a baseline for their intri nsic learning capabilities without manual feature engineering. As sho wn in the heatmap in Figure 2, the proposed multi-block CNN achie v ed the highest R 2 scores across all four tar get properties. This gure demonstrates the superior performance of the multi-block CNN on ra w spectral data. This result highlights its po werful, b uilt-in ability to e xtract meaningful features directly from unprocessed data. In contrast, the simple DNN performed the w orst, conrming that con v olutional layers are essential for ef fecti v e feature e xtraction from 1D spectral data. While the simple CNN and more comple x architectures li k e DeepSpectra and SpectraV GG performed well, the y were consistently outperformed by our multi-block approach, underscoring its g ains. Figure 2. Heatmap of mean R 2 v alues for each model without an y preprocessing 3.2. Impact of pr epr ocessing Ne xt, this study e v aluated each model using its optimal preprocessing method, identied through e xhausti v e testing. T able 1 summarizes both the selected best preprocessing method and the resulting mean R 2 ± SD for each tar get. The multi-block CNN, when combined with its optimal preprocessing, ag ain outperformed all other models and sho wed lo w v ariability . More importantly , its performance on ra w data (MultiBlock RA W) remained superior to all other models, e v en when the y used their optimal preprocessing. Deep multi-bloc k con volutional neur al network for quantitative near -infr ar ed ... (Hilmy Abidzar T awakal) Evaluation Warning : The document was created with Spire.PDF for Python.
4516 ❒ ISSN: 2252-8938 T able 1. Summary of optimal preprocessing and corresponding mean R 2 ± SD for each model and dataset Method T otal f at F atty acid Moisture Protein PLSR 0 . 655 ± 0 . 046 [SG] 0 . 940 ± 0 . 059 [SG] 0 . 754 ± 0 . 040 [SG] 0 . 715 ± 0 . 055 [SG] DNN 0 . 966 ± 0 . 010 [Cmb] 0 . 963 ± 0 . 008 [Cmb] 0 . 940 ± 0 . 043 [Cmb] 0 . 854 ± 0 . 024 [Cmb] CNN 0 . 964 ± 0 . 008 [MMS] 0 . 959 ± 0 . 011 [MMS] 0 . 940 ± 0 . 038 [MMS] 0 . 862 ± 0 . 031 [MMS] DeepSpectra 0 . 948 ± 0 . 012 [AS] 0 . 954 ± 0 . 010 [Cmb] 0 . 905 ± 0 . 030 [SG] 0 . 857 ± 0 . 033 [SG] SpectraV GG 0 . 983 ± 0 . 007 [MMS] 0 . 978 ± 0 . 009 [MMS] 0 . 966 ± 0 . 025 [MMS] 0 . 896 ± 0 . 023 [SS] MultiBlock RA W 0 . 995 ± 0 . 004 [Ra w] 0 . 997 ± 0 . 001 [Ra w] 0 . 991 ± 0 . 014 [Ra w] 0 . 939 ± 0 . 019 [Ra w] MultiBlock 0 . 996 ± 0 . 003 [AS] 0 . 997 ± 0 . 001 [Cmb] 0 . 991 ± 0 . 014 [Ra w] 0 . 941 ± 0 . 017 [AS] SG: Sa vitzk y-Golay , Cmb: combination, MMS: min-max scaling, AS: auto scaling, SS: standard scaling 3.3. Deep di v e into multi-block con v olutional neural netw ork perf ormance 3.3.1. Nested cr oss-v alidation perf ormance T able 2 presents the rigorous nested cross-v alidation results for the optimal multi-block CNN across all four tar get properties. The model demonstrated e xceptional predicti v e accurac y . The reported v alues represent the mean and SD across the v e outer folds of the nested cross-v alidation procedure. The 95% condence interv als (CI) were computed using the Student’ s t-distrib ution with four de grees of freedom ( t 0 . 975 , 4 = 2 . 776 ), according to CI = t · SD √ n , where n = 5 denotes the number of outer folds. The relati v ely narro w CI, particularly for total f at and f atty acid prediction, indicate stable generalization performance across folds and conrm the rob ustness of the proposed multi-block CNN under strict test-set isolation. From an application perspecti v e, the model achie v ed RPD scores ranging from 4 . 18 (protein) to 22 . 02 (f atty acid). According to standard industry thresholds, an RPD > 5 indicates that the model is highly suitable for strict quality control, while an RPD > 3 is e xcellent for screening. Therefore, the proposed multi-block CNN is fully capable of replacing m anual chemical assays for rigorous postharv est rice e v aluation. T able 2. Nested cross-v alidation performance of the proposed multi-block CNN (mean ± SD; 95% CI sho wn in brack ets) T ar get R 2 (Mean ± SD) RMSEcv RMSEp RPD RER T otal f at 0.9954 ± 0.0038 0.0858 ± 0.0139 0.0191 ± 0.0075 17.35 ± 5.23 45.94 ± 13.89 95% CI [ ± 0.0047] [ ± 0.0172] [ ± 0.0093] [ ± 6.50] [ ± 17.24] F atty acid 0.9974 ± 0.0013 0.0133 ± 0.0020 0.00216 ± 0.00063 22.02 ± 7.12 65.70 ± 21.33 95% CI [ ± 0.0017] [ ± 0.0025] [ ± 0.00078] [ ± 8.84] [ ± 26.48] Moisture 0.9912 ± 0.0139 0.2845 ± 0.0371 0.0775 ± 0.0589 17.07 ± 6.57 46.47 ± 17.94 95% CI [ ± 0.0173] [ ± 0.0460] [ ± 0.0732] [ ± 8.15] [ ± 22.27] Protein 0.9385 ± 0.0188 0.1305 ± 0.0147 0.0754 ± 0.0078 4.18 ± 0.65 15.41 ± 1.46 95% CI [ ± 0.0234] [ ± 0.0182] [ ± 0.0097] [ ± 0.80] [ ± 1.81] 3.3.2. Signicance test This study e v aluated the paired cross-v alidation folds globally ( N = 20 ) using a tw o-sided W ilcoxon signed-rank test. T o pre v ent type I errors across the six e v aluation metrics, we applied the Benjamini-Hochber g procedure to control the f alse disco v ery rate ( α = 0 . 05 ). The statistical analysis re v ealed no signicant dif ference in performance between the ra w and preprocessed pipelines. As detailed in T able 3, the corrected p -v alues for primary metrics such as R 2 ( q = 0 . 5429 ) and RPD ( q = 0 . 1322 ) conrm that the addition of preprocessing does not yield a statistically signicant impro v ement. In this table, corrected p -v alues were adjusted using the Benjamini-Hochber g procedure to control the f alse disco v ery rate. T able 3. P aired W ilcoxon signed-rank test results: preprocessed vs. ra w data ( N = 20 ) Metric W i lcoxon stat Ra w p -v alue Corrected p -v alue Signicant ( α = 0 . 05 ) R 2 82.0 0.4091 0.5429 F alse RMSEp 80.0 0.3683 0.5429 F alse MAE 96.0 0.7562 0.7562 F alse MAPE 84.0 0.4524 0.5429 F alse RPD 51.0 0.0441 0.1322 F alse RER 51.0 0.0441 0.1322 F alse 3.3.3. Data efciency T o assess model generalization as a function of data a v ailability , this study conducted a data-ef cienc y analysis by v arying the training set size (25%, 50%, 75%, and 100%). As detailed in T able 4, performance Int J Artif Intell, V ol. 15, No. 5, October 2026: 4512–4522 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4517 scales consistently with a v ailable data. W ith 25% of the data, the model under generalized ( R 2 v alues ranging from 0 . 48 for moisture to 0 . 73 for total f at). Performance impro v ed sharply as the sample size increased, and with 75% of the data used for training, all tar gets achie v ed R 2 ≥ 0 . 90 . Using 100% of the data produced peak performance ( R 2 > 0 . 93 for all tar gets, with f at and f atty acid abo v e 0 . 99 ), while g ains from 75% to 100% became smaller , indicating an approaching plateau. T able 4. Summary of R 2 scores across v arying training data fractions T ar get 25% 50% 75% 100% T otal f at 0 . 73 ± 0 . 08 0 . 87 ± 0 . 04 0 . 96 ± 0 . 01 0 . 995 ± 0 . 004 F atty acid 0 . 61 ± 0 . 13 0 . 79 ± 0 . 12 0 . 94 ± 0 . 04 0 . 997 ± 0 . 001 Moisture 0 . 48 ± 0 . 22 0 . 85 ± 0 . 09 0 . 92 ± 0 . 04 0 . 991 ± 0 . 014 Protein 0 . 49 ± 0 . 17 0 . 74 ± 0 . 16 0 . 90 ± 0 . 01 0 . 939 ± 0 . 019 3.3.4. Ablation study T o systematically v erify the architectural contrib utions of the proposed model, an ablation study w as conducted. This study compared the optimal multi-block CNN ag ainst a monolithic CNN v ariant (single block) with a comparable parameter count, which ef fecti v ely r emo v es the spectral se gmentation mechanism. As sho wn in T able 5, the monolithic CNN (1 block) f ailed to model the data ef fecti v ely under the nested CV protocol. This drastic performance g ap conrms that se gmenting the spectrum enabl es the netw ork to learn localized, distinct feature maps t hat are otherwise lost when global pooling is applied to the entire ra w spectrum. T able 5. Impact of block se gmentation on model performance (total f at dataset) Architecture R 2 (Mean ± SD) R M S E p RPD RER Multi-block (proposed) 0 . 995 ± 0 . 004 0 . 019 ± 0 . 007 17 . 35 ± 5 . 23 45 . 94 ± 13 . 89 Monolithic (single block) 0 . 267 ± 0 . 068 0 . 250 ± 0 . 012 1 . 18 ± 0 . 05 3 . 21 ± 0 . 15 3.3.5. P ermutation test f or model r ob ustness T o v erify that the proposed deep multi-block CNN learns genuine w a v elength–chemical relationshi ps rather than spurious correlations or dataset artif acts, a tar get permutation test (y-randomization) w as performed. In this e xperiment, the tar get v ariables (f at, f atty acid, moisture, and protein) were randomly permuted with respect to the NIR spectral inputs, thereby completely destro ying the structural relationship between the predictor matrix ( X ) and the response v ector ( y ). Using the pre viously optimized h yperparameter congurations, the model w as re-e v aluated under identical 5-fold cross-v alidation conditions. Under proper model beha vior , remo v al of the X – y dependenc y should lead to a collapse in predicti v e performance. The results are summarized in T able 6. T able 6. Model performance follo wing y-randomization across 5-fold cross-v alidation (mean ± SD) T ar get R 2 R M S E p RPD RER T otal f at 0 . 0087 ± 0 . 0182 0 . 2983 ± 0 . 0167 1 . 0124 ± 0 . 0093 2 . 6841 ± 0 . 1513 F atty acid − 0 . 0089 ± 0 . 0330 0 . 0436 ± 0 . 0021 1 . 0038 ± 0 . 0160 3 . 0099 ± 0 . 1439 Moisture 0 . 0035 ± 0 . 0637 1 . 0006 ± 0 . 0360 1 . 0108 ± 0 . 0306 2 . 7647 ± 0 . 0771 Protein − 0 . 0020 ± 0 . 0353 0 . 3114 ± 0 . 0336 1 . 0073 ± 0 . 0176 3 . 7476 ± 0 . 3962 As e xpected, eliminating the true X – y relationship reduced predicti v e performance to near -random le v els across all tar gets. The R 2 approached zero or became ne g ati v e, indicating no e xplanatory po wer . Lik e wise, the RPD con v er ged to approximately 1.0, which in chemometric interpretation corresponds to performance equi v alent to predicting the mean v alue for all samples. These results conrm that the high predicti v e accurac y observ ed on the structured dataset arises from meaningful spectral–chemi cal correlations rather than o v ertting, memorization, or background artif acts. The permutation analysis therefore pro vides strong statistical e vidence of genuine model generalization. 3.3.6. Str ess testing via articial noise injection T o further quantify the model’ s resil ience to instrumental de gradation, a systematic stress test w as conducted by injecting articial Gaussian noise into the test set spectra. Gaussian noise, dened as N (0 , σ 2 ) , w as added to the standardized spectral inputs, where σ represents the SD of the noise le v el relati v e to the original spectral v ariance. The noise intensity w as incrementally scaled from σ = 0 . 01 (lo w noise) to σ = 0 . 10 Deep multi-bloc k con volutional neur al network for quantitative near -infr ar ed ... (Hilmy Abidzar T awakal) Evaluation Warning : The document was created with Spire.PDF for Python.
4518 ❒ ISSN: 2252-8938 (se v ere noise). The pre-trained multi-block CNN w as then e v aluated on these corrupted datasets across all 5-folds without retraining the model. The resulting a v erage R 2 for each property is presented in T able 7. T able 7. Model performance de gradation ( R 2 ) under increasing le v els of articial Gaussian noise Noise le v el ( σ ) T otal f at ( R 2 ) F atty acid ( R 2 ) Moisture ( R 2 ) Protein ( R 2 ) 0.00 (baseline) 0.9924 0.9951 0.9888 0.9368 0.01 (lo w) 0.9891 0.9934 0.9882 0.9318 0.05 (moderate) 0.8351 0.8576 0.8887 0.8209 0.10 (se v ere) 0.4614 0.3385 0.6592 0.2465 As sho wn in T able 7, the multi-block CNN demonstrates e xceptional rob ustness at lo w noise le v els ( σ = 0 . 01 ). The predi cti v e performance remains virtually identical to the baseline, with R 2 de grading by less than 1% across all tar get properties. Ev en under moderate noise corruption ( σ = 0 . 05 ), the architecture e xhibits graceful de gradation, maintaining strong predicti v e capabilities with R 2 v alues remaining abo v e 0.82 for all properties. A substantial collapse in predicti v e performance is only observ ed under se v ere articial corruption ( σ = 0 . 10 ), where the injec ted noise v ariance be gins to hea vily mask the underlying spectral v ariance. Ho we v er , such e xtreme de gradation generally e xceeds the typical operational v ariance observ ed in modern, calibrated F ourier transform near -infrared (FT -NIR) spectrometers. Ultimately , these results empiricall y v alidate the model’ s operational reliability in real-w orld scanning en vironments, conrming its ability to maintain high accurac y without the strict need for traditional scatter -correction preprocessing algorithms. 3.4. Discussion The e xperimental results demonstrate that the proposed multi-block CNN model signicantly outperforms both traditional and other deep l earning-based approaches for quantitati v e NIR analysis. Its ability to operate directly on ra w spectra presents a major adv antage. This is demonstrated by its e xceptional performance on the ra w f atty acid dataset ( R 2 = 0 . 997 ± 0 . 001 ). A k e y architectural adv antage is the balance between model capacity and generalizat ion. While deeper architectures lik e DeepSpectra and SpectraV GG ha v e sho wn strong performance, the y can risk o v er tting on smaller datasets [5], [13]. The multi-block CNN combines depth (in its dense layers) with width (via spectral se gmentation), which pro v ed to be a more rob ust strate gy for this task. Despite its success, the model has se v eral limitations. The lo west performance w as observ ed for protein prediction ( R 2 = 0 . 939 ± 0 . 019 ), a well-kno wn challenge in traditional NIR analysis due to subtle, often noisy spectral signals. T o in v estig ate ho w our model handled this challenge, we generated salienc y maps to visualize its internal feature-ranking process [5]. The results, sho wn in Figure 3, pro vide a clear visual e xplanation for the performance disparity . The model in the gure identies multiple, distrib uted, and subtle spectral re gions for the more challenging protein prediction. In contrast, it focuses on a single, highly concentrated, and dominant re gion for the high-performing f atty acid prediction. Bright yello w indicates high importance. The salienc y map for the protein sho wn in Figure 3(a) re v eals that the model identies multiple dif fuse re gions of importance across the spectrum. Th i s conrms that the spectral signature for protein is comple x and lacks a single dominant feature. In stark contrast, Figure 3(b) presents the salienc y map for f atty acid, one of the best-performing models. The model focuses its attention on a single, highly concentrated, and dominant re gion at the start of the spectrum. This vi sual e vidence strongly suggests that the model’ s superior performance on f atty acids is due to its ability to identify a l ocalized, information-rich spectral s ignature, a task that is signicantly more challenging for the subtle, distrib uted features of proteins. In terms of generalizability , the modula r , block-based architecture is not inherently limited to rice spectra. Its ability to learn localized features from 1D data mak es it a strong candidate for other quantitati v e spectral analyses, b uilding on the gro wing body of w ork applying CNNs to v arious spectral challenges [25]. The core principle of se gmenting a signal to isolat e localized information is broadly applicable. In addition to absolute performance, the SDs across runs were lo w , conrming the model’ s reproducibilit y and stability . The lo w root mean squared error (RMSE) and MAPE v alues for f atty acid and moisture predictions, in particular , emphasize the model’ s suitability for precise, real-time applications such as food quality monitoring or on-site agricultural screening. Int J Artif Intell, V ol. 15, No. 5, October 2026: 4512–4522 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4519 (a) (b) Figure 3. Feature importance maps for (a) protein prediction and (b) f atty acid prediction 3.5. Futur e w ork Future w ork should focus on impro ving performance for challenging tar gets lik e proteins. This could in v olv e e xploring h ybrid architectures that inte grate the multi-block CNN with transformer or attention mechanisms to better model long-range spectral dependencies and emphasize subtle features. Additionally , in v estig ating transfer learning to adapt the model across dif ferent crops or instruments and deplo ying the model on portable NIR de vices for edge articial intelligence applications represent promising directions for making this technology more accessible for real-time quality control. 4. CONCLUSION This study introduced a no v el deep multi-bl o c k CNN architecture for quantitati v e NIR spe ctral analysis, inte grating feature e xtraction and calibration into a single, intelligent system. The proposed model Deep multi-bloc k con volutional neur al network for quantitative near -infr ar ed ... (Hilmy Abidzar T awakal) Evaluation Warning : The document was created with Spire.PDF for Python.
4520 ❒ ISSN: 2252-8938 demonstrated superior performance o v er both traditional chemometric methods and other deep learning architectures, achie ving state-of-the-art results in predicting f atty acid, moisture, and t otal f at in rice. The k e y contrib ution of this w ork i s an end-to-end model that can operate directly on ra w spectra without con v entional preprocessing, thereby simplifying the analysis pipeline and reducing dependence on do m ain e xpertise. This study statistically v alidated that the model’ s performance on ra w data is equi v alent to its performance with optimized preprocessing. The architecture’ s no v elty lies in its use of spectral se gmentation and parallel processing, which enables it to learn localized features ef fecti v ely and autonomously , a signicant impro v ement o v er monolithic, end-to-end models. FUNDING INFORMA TION This research is funded by the Ministry of Research, T echnology , and Education through the doctoral dissertation research program (PDD), with the title ”Preprocessing method of NIR spectrum based on CNN for estimation of rice quality” and the number 0536/E5/PG.02.00/2023. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib utions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Hilmy Abidzar T a w akal ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Y ohanes Aris Purw anto ✓ ✓ ✓ ✓ ✓ ✓ Son y Hartono W ijaya ✓ ✓ ✓ ✓ Shelvie Nidya Ne yman ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT The authors declare that the y ha v e no kno wn competing nancial interests or personal relat ionships that could ha v e appeared to inuence the w ork reported in this paper . Furthermore, the authors conrm that this study w as conducted independently and that no e xternal interests, af liations, or personal relationships inuenced the research design, data analysis, interpretation of the results, or preparation of the manuscript. INFORMED CONSENT This section is not applicable. This study did not in v olv e human participants, human data, or an y personally identiable informat ion. All data used were either publicly a v ailable, fully anon ymized, or deri v ed from non-human sources, and therefore no informed consent w as required from indi viduals. ETHICAL APPR O V AL This section is not applicable. This research did not in v olv e human subjects, human biological materials, or e xperimental procedures on animals. The w ork w as conducted solely on computational models, publicly a v ailabl e datasets, or non-sensiti v e data that did not require interv ention with li ving or g anisms. Therefore, ethical appro v al from an institutional re vie w board or animal ethics committee w as not necessary for this study . Int J Artif Intell, V ol. 15, No. 5, October 2026: 4512–4522 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4521 D A T A A V AILABILITY The data that support the ndings of this study are a v ailable from the corresponding author , [Y AP], upon reasonable request. The data, which were collected from proprietary rice samples measured with portable NIR spectroscop y , are not publ icly a v ailable due to restrictions related to ongoing research and institutional data-sharing policies. Synthetic data schemas replicating the struct u r e of the original datasets, along with the e xperimental code used to produce results in this paper , are a v ailable upon request to f acilitate reproducibility . REFERENCES [1] K. B. W alsh, V . A. McGlone, and D. H. Han, “The uses of near infra-red spectroscop y in postharv est decision support: a re vie w , ” P ostharvest Biolo gy and T ec hnolo gy , v ol. 163, May 2020, doi: 10.1016/j.postharvbio.2020.111139. [2] Y . Liu, H. Pu, and D.-W . Sun, “Ef cient e xtraction of deep image features using con v olutional neural netw ork (CNN) for applications in detecting and analysing comple x food matrices, ” T r ends in F ood Science & T ec hnolo gy , v ol. 113, pp. 193–204, Jul. 2021, doi: 10.1016/j.tifs.2021.04.042. [3] M. Nadimi and J. P aliw al, “Recent applications of near -infrared spectroscop y in food quality anal ysis, ” F oods , v ol. 13, no. 16, Aug. 2024, doi: 10.3390/foods13162633. [4] J. W estermayr and P . Marquetand, “Machine learning spectroscop y to adv ance computation and analysis, ” Chemical Science , v ol. 16, no. 46, pp. 21660–21676, 2025, doi: 10.1039/D5SC05628D. [5] X. Zhang, T . Lin, J. Xu, X. Luo, and Y . Y ing, “DeepSpectra: an end-to-end deep learning approach for quantitati v e spectral analysis, ” Analytica Chimica Acta , v ol. 1058, pp. 48–57, Jun. 2019, doi: 10.1016/j.aca.2019.01.002. [6] Y . Xiao, L. Zhou, Y . Zhao, H. Qi, Y . Pu, and C. Zhang, “Deep learning-based re gression of food quality attrib utes using near -i nfrared spectroscop y and h yperspectral imaging: a re vie w , ” F ood Chemis try , v ol. 493, No. 4, No v . 2025, doi: 10.1016/j.foodchem.2025.145932. [7] W . Zhang, L. C. Kasun, Q. J. W ang, Y . Zheng, and Z. Lin, “ A re vie w of machine learning for near -infrared spectroscop y , ” Sensor s , v ol. 22, no. 24, Dec. 2022, doi: 10.3390/s22249764. [8] P . Mishra et al ., “Deep learning for near -infrared spectral data modelling: h ypes and benets, ” T rA C T r ends in Analytical Chemistry , v ol. 157, Dec. 2022, doi: 10.1016/j.trac.2022.116804. [9] C. Zhang, L. Zhou, Y . Zhao, S. Zhu, F . Liu, and Y . He, “Noise reduction in the spectral domain of h yperspectral images using denoising autoencoder methods, ” Chemometrics and Intellig ent Labor ator y Systems , v ol. 203, Aug. 2020, doi: 10.1016/j.chemolab .2020.104063. [10] P . Nie, J. Zhang, X. Feng, C. Y u, and Y . He, “Classication of h ybrid seeds using near -infrared h yperspectral imaging technology combined with deep learning, ” Sensor s and Actuator s B: Chemical , v ol. 296, Oct. 2019, doi: 10.1016/j.snb .2019.126630. [11] C. Zhang, W . W u, L. Zhou, H. Cheng, X. Y e, and Y . He, “De v eloping deep learning based re gression approaches for determination of chemical compositions in dry black goji berries ( L ycium rutheni cum Murr .) using near -infrared h yperspectral imaging, ” F ood Chemistry , v ol. 319, Jul. 2020, doi: 10.1016/j.foodchem.2020.126536. [12] X. Zhang, J. Y ang, T . Lin, and Y . Y ing, “F ood and agro-product quality e v aluation based on spectroscop y and deep learning: a re vie w , ” T r ends in F ood Science & T ec hnolo gy , v ol. 112, pp. 431–441, Jun. 2021, doi: 10.1016/j.tifs.2021.04.008. [13] U. Blazhk o, V . Shapa v al, V . K o v ale v , and A. K ohler , “Comparison of augmentation and pre-processing for deep learning and chemometric classication of infrared spectra, ” Chemometrics and Intellig ent Labor atory Systems , v ol. 215, Aug. 2021, doi: 10.1016/j.chemolab .2021.104367. [14] P . Mis hra and D. P ass os, “D eep m ultiblock predicti v e modelling using parallel input con v olutional neural netw orks, ” Analytica Chimica Acta , v ol. 1163, Jun. 2021, doi: 10.1016/j.aca.2021.338520. [15] L. Shuai, Z. Li, Z. Chen, D. Luo, and J. Mu, “ A research re vie w on deep learning combined with h yperspectral imaging in multiscale agricultural sensing, ” Computer s and Electr onics in Agricultur e , v ol. 217, Feb . 2024, doi: 10.1016/j.compag.2023.108577. [16] L. Lin, Y . He, Z. Xiao, K. Zhao, T . Dong, and P . Nie, “Rapid-detection sensor for rice grain moisture based on NIR spectroscop y , ” Applied Sciences, v ol. 9, no. 8, Apr . 2019, doi: 10.3390/app9081654. [17] X. Miao et al ., “Classication of rice based on storage time by using near infrared spectroscop y and chemometric methods, ” Micr oc hemical J ournal , v ol. 171, Dec. 2021, doi: 10.1016/j.microc.2021.106841. [18] P . S. Sampaio, B. Carbas, A. Soares, I. Sousa, and C. Brites, “Spectral mark ers and machine learning: re v olutionizing rice e v aluation with near infrared spectroscop y , ” F ood Chemistry , v ol. 492, No v . 2025, doi: 10.1016/j.foodchem.2025.145569. [19] P . Mishra, J. M. Roger , D. N. Rutledge, and E. W oltering, “SPOR T pre-processing can impro v e near -infrared quality prediction models for fresh fruits and agro-materials, ” P ostharvest Biolo gy and T ec hnolo gy , v ol. 168, Oct. 2020, doi: 10.1016/j.postharvbio.2020.111271. [20] J .-M. Roger , A. Biancolill o, and F . Marini, “Sequential preprocessing through OR Thogonalization (SPOR T) and its application to near infrared spectroscop y , ” Chemometrics and Intellig ent Labor atory Systems , v ol. 199, Apr . 2020, doi: 10.1016/j.chemolab .2020.103975. [21] L.- H. Xie et al ., “Simultaneous determination of apparent amylose, am ylose and amylopectin content and classication of w axy rice using near -infrared spectroscop y (NIRS), ” F ood Chemistry , v ol. 388, Sep. 2022, doi: 10.1016/j.foodchem.2022.132944. [22] T . Akiba, S. Sano, T . Y anase, T . Ohta, and M. K o yama, “Optuna: a ne xt-generation h yperparameter optimization frame w ork, ” in 25th A CM SIGKDD International Confer ence on Knowledg e Disco very & Data Mining , Jul. 2019, pp. 2623–2631, doi: 10.1145/3292500.3330701. [23] D. P assos and P . Mis hra, “ A tutorial on automatic h yperparameter tuning of deep spectral modelling for re gression and classication tasks, ” Chemometrics and Intellig ent Labor atory Systems , v ol. 223, Apr . 2022, doi: 10.1016/j.chemolab .2022.104520. [24] F . Ding, M. ´ A. R.Delg ado, R. Zha, and J. F . G.-Mart ´ ın, “Optuna-optimized boosting models for predicting quality traits in multiple juice types using NIRS: interpretability analysis via SHAP , ” F ood Contr ol , v ol. 182, Apr . 2026, doi: 10.1016/j.foodcont.2025.111878. Deep multi-bloc k con volutional neur al network for quantitative near -infr ar ed ... (Hilmy Abidzar T awakal) Evaluation Warning : The document was created with Spire.PDF for Python.