Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 16, No. 4, August 2026, pp. 2074 ∼ 2086 ISSN: 2088-8708, DOI: 10.11591/ijece.v16i4.pp2074-2086 ❒ 2074 Integrating principal component analysis in spatial-spectral fusion models f or h yperspectral image segmentation Alexander Calvin, Laksmita Rahadianti F aculty of Computer Science, Uni v ersitas Indonesia, Depok, Indonesia Article Inf o Article history: Recei v ed Dec 6, 2025 Re vised Apr 12, 2026 Accepted Apr 26, 2026 K eyw ords: Hyperspectral imaging Land-use spatial-spectral fusion Principal component analysis Semantic se gmentation U A V remote sensing ABSTRA CT Hyperspectral imaging (HSI) from unmanned aerial v ehicles (U A Vs) pro vides rich spatial-spectral data, b ut its high dimensionality presents signicant com- putational challenges for semantic se gmentation. While state-of-the-art models lik e the transformer -based HSI-T ransUnet are often emplo yed, the y introduce massi v e computational o v erhead. This study adapts a lightweight, dual-tunnel deep con v olutional neural netw ork (DCNN) frame w ork for land-use se gmenta- tion on h yperspectral images by inte grating PCA-based spatial reduction in the spatial branch, and benchmarks it on the U A V -HSI-Crop dataset ag ainst HSI- T ransUnet. F or further analysis, an ablation study compares principal compo- nent analysis (PCA) and local similarity projection (LSP) as spatial feature e x- tractors. The results demonstrate a signicant performance and ef cienc y adv an- tage. Our proposed PCA-based model (271.1K parameters) obtained a Kappa ( κ ) of 0.8582, o v erall accurac y (O A) of 0.8800, and a v erage accurac y (AA) of 0.4918, outperforming the LSP-based model by 0.65% in κ , 0.51% in O A, and 2.16% in AA and the HSI-T ransUnet baseline by 2.35% in κ , 1.95% in O A, and 8.10% in AA. On our e xperimental setup, this res ult w as achie v ed with a 152.7-fold reduction in model size, a 14.2-fold decrease in training time, and a 4.6-fold speedup in inference relati v e to the reported HSI-T ransUnet baseline. These ndings sho w that the PCA-based dual-tunnel DCNN pro vides a f a v or - able trade-of f between class-balanced accurac y and computational ef cienc y for this HSI se gmentation task. This is an open access article under the CC BY -SA license . Corresponding A uthor: Ale xander Calvin F aculty of Computer Science, Uni v ersitas Indonesia UI Depok Campus, Depok, Indonesia Email: ale xander .calvin@ui.ac.id 1. INTR ODUCTION Hyperspectral imaging (HSI) in remote sensing pro vides the capability to record high-resolution spec- tral information in a lar ge number of narro w and contiguous spectral bands. The use of unmanned aerial v ehi- cles (U A Vs) with h yperspectral imaging technology has a lso enabled ne-scale observ ation of surf ace materials for lar ge re gions [1]. Unlik e re gular RGB images, HSI pro vides recorded reectance information in hundreds of narro w spectral bands to acquire infor mation related to material composition [2]. This spectral detail allo ws for precise identication of surf ace materials, supporting applications such as land use mapping [3], en vironmental monitoring [4], precision agriculture [5], and urban analysis [6]. Ho we v er , the se gmentation of h yperspectral aerial imagery remains a non-tri vial task due to the high dimensionality of the data, redundanc y among spectral bands, limited a v ailability of annotated samples, and the frequent presence of noise and en vironmental artif acts [7]. J ournal homepage: http://ijece .iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 2075 The high-dimensionality of HSI data poses considerable computational challenges. A lar ge number of spectral bands causes high computational comple xity with the potential for model o v ertting. This is particu- larly problematic gi v en the scarcity of labeled data in the conte xt of agriculture [8]. T o address such problems, se v eral methods of dimensionality reduction and feature e xtraction ha v e been proposed to eliminate redundant information with the intention of retaining spectral-spatial discriminati v e features. These include methods such as principal component analysis (PCA) [9] and linear discriminant analysis (LD A) [10]. T raditional machine learning (ML) techniques initially impro v ed classication accurac y by reduc- ing features dimensionality . Ho we v er , these methods relied hea vily on manual feature engineering and often struggled to model the comple x, strongly nonlinear spectral relationships inherent in h yperspectral imagery [11]-[13]. While deep learning ( DL) techniques using con v olutional neural netw orks (CNNs) o v ercame earlier limitations by automating feature learning [14], these spectrally focused methods often ignore spatial conte xt [15], leading to nois y salt-and-pepper classication maps [16]. T o address these limitati ons, recent literature emphasizes spatial-spectral feature e xtraction methods [17], e xploiting high-dimensional spectral features and neighboring spatial information to balance spectral redundanc y with spatial coherence. Recent studies ha v e further rened this frame w ork by combining spatial re gularization and ltering processes, including the use of guided or edge-a w are ltering [18], [19], te xture-based fusion with Gabor lters [20], and morphological or state-space models to ef ciently capture spatial-spectral dependencies [21]. It is important to note that these methods do not solely focus on DL-based pipelines , b ut also on inte- gration or comparison with traditional ML methods. T raditional ML methods generally struggle to model com- ple x spatial-spectral relationships in h yperspectral data, perform poorly on patch-based schemes, and f ace high computational demands for full-scene processing. In comparison, patch-wise CNN approaches e xtract local features from o v erlapping patches centered on each pix el, capturing ne spatial-spectral patterns b ut introduc- ing redundant computation and high training costs [22]. Meanwhile, U-Net-style architectures perform pix el- wise prediction on non-o v erlapping spectral–spatial patches, reducing redundanc y while maintaining broader conte xtual a w areness, achie ving a more ef cient balance between accurac y and comple xity [23]. This article focuses on e v aluating spatial-spectral feature e xtraction approaches that inte grate spatial conte xt within deep learning frame w orks implem ented on crop h yperspectral data, due to the scene’ s distinct area cate gories and te xtural characterist ics that reect v ariations in v e getation type, gro wth stage, and canop y structure, which collecti v ely inuence the spatial-spectral patterns captured by the imaging sensor [24], [25]. Simply adding spatial cues to spectral features is insuf cient. High-dimensional HSI data can cause spatial lters to be o v ershado wed by spectral redundanc y , while purely spectral learning often o v erlooks important local consistenc y [26]. While r ecent T ransformer models are po werful, hea vy computational costs often limit their real-w orld appli cation [27]. This highlights an important res earch g ap for lightweight, spatial-spectral architectures that can accurately process comple x agricultural data without the massi v e o v erhead. A balanced approach is therefore needed to ensure that both components contrib ute meaningfully . T o address this, the study adopts a dual-tunnel st rate gy in which the spatial pathw ay acts as a spectrally a w are lter . Dimensionality reduction is applied before spatial ltering, allo wing the model to reduce redundanc y and e xtract clean, te xture-ori ented information. This design ensures that spa tial conte xt becomes a rened e xtension of the spectral representation, ef fecti v ely complementing the deeper spectral features learned through the parallel pathw ay . The main contrib utions of this article are summarized as follo ws. − Ev aluation of the spatial-spectral DCNN frame w ork: This study implements and tests a spatial-spectral feature e xtraction method inte grated within a dual-tunnel deep con v olutional neural netw ork (DCNN) [20]. The frame w ork is e v aluated for its ability to enhance se gmentation accurac y on comple x U A V h yperspectral crop data by e xplicitly separating and fusing spectral and spatial process path. − Comparati v e study of dimensionality reduction methods: Ablation study is conducted to e v aluate dimen- sionality reduction techniques within the spatial lter tunnel. Specically , the study compares the ef fec- ti v eness of local similarity projection (LSP) [20] and PCA [9] to determine which method best preserv es discriminati v e spatial-spectral features. The rest of this article is or g anized as follo ws. Section 2 introduces the e v aluated spatial-spect ral methods, the dataset description, parameter settings, e v aluation metrics, and the proposed e xperimental frame- w ork. Section 3 discusses the results and comparati v e analysis with e xisting benchmarks. Finally , section 4 concludes the study and outlines future directions. Inte gr ating principal component analysis in spatial-spectr al ... (Ale xander Calvin) Evaluation Warning : The document was created with Spire.PDF for Python.
2076 ❒ ISSN: 2088-8708 2. METHOD This study adopts an e xperimental design to e v aluate the ef fect of spatial-spectral feature e xtraction for crop se gmentation of U A V -based h yperspectral images. The o v erall w orko w follo ws the general spatial- spectral fusion frame w ork inspired by recent local similarity-based methods, b ut adapts it for semantic se gmen- tation instead of patch-based classication [20]. The processing pipeline inte grates dimensionality reduction, spatial ltering, feature fusion, and semantic se gmentation within a unied structure. 2.1. Ov erall framew ork of the spatial-spectral pipeline The o v erall frame w ork of the method is illustrated in Figure 1. The frame w ork is di vided into tw o parts, namely the spatial tunnel and the spectral tunnel. The spatial tunnel e xtracts structural and te xture information using LSP and Gabor ltering, which together enhance spatial coherence and edge representa- tion. The spectral tunnel emplo ys a DCNN to learn discrimi nati v e spectral characteristics directly from pix el- wise reectance v ectors. The outputs from both tunnels are fused via concatenation to form a joint repre- sentation which is processed by a dual-optimized clas sier head to generate the nal pix el-wise se gmentation [20]. Figure 1. The dual-tunnel DCNN architecture for spatial-spectral feature learning. A ra w h yperspectral patch ( 96 × 96 × 200 ) is processed through parallel branches: a spatial tunnel (dimensionality reduction and Gabor ltering) and a spectral tunnel (2-D con v olutions). The e xtracted spatial and spectral features are concatenated into a joint representation and fed into a dual classier to output the nal 96 × 96 predicted patch 2.1.1. Local similarity and Gabor -based spatial featur e extraction The spatial tunnel, sho wn in white in Figure 1, combines LSP and 2-D Gabor ltering to deri v e spatially coherent and te xture-sensiti v e representations from h yperspectral data. The method addresses the challenges posed by the high dimensionality and limited labeled samples of h yperspectral data by rst apply- ing LSP to reduce dimensionality while preserving neighborhood relationships. Spatial features are e xtracted through 2-D Gabor ltering, which captures edge- and frequenc y-oriented te xtures, while spectral features are learned using a CNN applied to the original h yperspectral cube. These tw o feature streams are then fused and input into a deeper CNN, follo wed by classication using a dual-optimization classier . LSP focuses on projecting the local similarity of HSI data by ensuring that spectrally similar neigh- boring pix els remain close in the reduced feature space [20]. This mak es LSP particularly suitable for HSIs, where class distrib utions are often non-Gaussian, multi-modal, and spatially dependent. LSP operates under the assumption that adjacent pix els, especially those belonging to the same material class, share strong spectral similarity and should therefore be represented closely in the projected subspace. F ormally , gi v en a set of train- ing samples { x i ∈ R d } n i =1 with corresponding class labels y i ∈ { 1 , . . . , c } , where c represents the number of classes. Int J Elec & Comp Eng, V ol. 16, No. 4, August 2026: 2074-2086 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 2077 LSP denes the similarity weight between tw o samples x i and x j as in (1). A i,j = exp − ∥ x i − x j ∥ 2 γ i · γ j (1) where γ i and γ j denote the local scaling f actors determined from distances to their respecti v e m -nearest neigh- bors. This adapti v e k ernel ensures that local neighborhoods are preserv ed, e v en in re gions with v arying class. LSP uses local inter -class and intra-class scatter matrices to construct a transformation matrix T LSP that si- multaneously maxi mizes class separability and minimizes intra-class v ariance. This is achie v ed through the optimization of the Fisher criterion in (2). T LSP = arg max T tr h T ⊤ L l w T − 1 T ⊤ L lb T i (2) where L lb and L l w represent the local inter - and intra-class scatter matrices. These matrices are constructed using the similarity weights A i,j and are dened as in (3) and (4) respecti v ely . L lb = 1 2 X i,j W i,j ( x i − x j )( x i − x j ) ⊤ (3) L l w = 1 2 X i,j W ′ i,j ( x i − x j )( x i − x j ) ⊤ (4) By preserving these local relationships, LSP ensures that adjacent pix els of the same cl ass are tightly clustered, while those of dif ferent classes remain wel l-separated in the reduced space. This struc- ture leads to lo wer classication error , especially in challenging scenarios with o v erlapping classes and spatial noise. Meanwhile, the Gabor lter serv es as a spatial tunnel that encodes directional te xture patterns around each pix el, pro viding spatial information that enhances the spectral tunnel deri v ed from the original h yperspec- tral cube [20]. The 2-D Gabor transform function is dened as a sinusoidal w a v e modulated by a Gaussian en v elope, e xpressed as (5). g ( x, y ; . . . ) = exp − x ′ 2 + γ 2 y ′ 2 2 σ 2 exp j 2 π x ′ ϑ + ϕ  (5) where the rotated coordinates x ′ and y ′ are dened in (6). x ′ = x cos θ + y sin θ y ′ = y cos θ − x sin θ (6) In this formulation, ( x, y ) represent the horizontal and v ertical spatial coordinates, and θ denotes the orientation of the lter in radians. The parameter ϑ controls the w a v elength of the sinusoidal f actor , while ϕ species the phase of fset. Additionally , σ denes the standard de viation of the Gaussian en v elope, and γ represents the spatial aspect ratio of the Gaussian function. This conguration allo ws the Gabor lter to selecti v ely enhance spatial structures aligned with a specic direction and scale, ef fecti v ely encoding local te xtures and edges. By applying this ltering on the LSP-reduced data, the model preserv es local feature consistenc y while reducing redundanc y across bands. The resulting Gabor -ltered output is passed into the spatial tunnel for fusion with spectral features. 2.1.2. PCA-based spatial r eduction T o e v aluate a simpler alternati v e to the original LSP module [20], a PCA-based v ariant is introduced in the spatial tunnel. In this v ariant, PCA replaces only the dimensionality-reduction stage [28], while the subsequent Gabor ltering, feature fusion, and se gmentation components remain unchanged. Gi v en an input ra w HSI with B spectral bands, PCA is rst applied as a pre-processing step to project the ra w HSI into k principal components. The s ame Gabor -based spatial ltering module then processes the resulting principal component images to form the spatial feature representation, which is concatenated with the spectral features Inte gr ating principal component analysis in spatial-spectr al ... (Ale xander Calvin) Evaluation Warning : The document was created with Spire.PDF for Python.
2078 ❒ ISSN: 2088-8708 from the parallel DCNN branch before pix el-wise classication. The procedural o w for this PCA-based spatial-spectral classication approach is summarized in Algorithm 1. Algorithm 1 . PCA-Based Spatial-Spectral Classication Requir e: Ra w HSI X with B spectral bands, tar get components k Ensur e: Pix el-wise classication map M 1: // 1. Spatial T unnel Pr ocessing 2: X pca ← PCA ( X , k ) { Project B bands into k principal components } 3: F spatial ← GaborFilter ( X pca ) { Extract spatial features } 4: // 2. Spectral T unnel Pr ocessing (in parallel) 5: F spectr al ← DCNN ( X ) { Extract deep spectral features } 6: // 3. F eatur e Fusion & Classication 7: F f used ← Concatenate ( F spatial , F spectr al ) { Combine spatial and spectral features } 8: M ← Classify ( F f used ) { Perform pix el-wise classication } 9: r etur n M 2.1.3. Spectral featur e extraction tunnel T o impro v e the utilization of spectral data features, a spectral feature e xtraction tunnel, sho wn in blue in Figure 1, is designed to run in parallel with the spatial process ing. This tunnel focuses on the central pix el Z ij at position P ij and its immediate neighborhood, taking a 3-D patch of size k × k × B as input, where B is the number of spectral bands and k = 2 r + 1 is determined by the radius r . The e xtraction process in v olv es a series of 2-D con v olutions and non-linear transformations designed to compress the spectral depth while preserving discriminati v e signatures. The original data Z ij is rst processed by a 2-D con v olutional layer follo wed by batch normalization (BN) to standardize features. This is follo wed by a rectied linear unit (R eLU) acti v ation function. A second con v olutional layer further renes these features. 2.1.4. DCNN f or spectral-spatial featur e extraction The DCNN architecture [20], sho wn in purple in Figure 1, follo ws an encoder -decoder structure de- signed to e xtract features and subsequently restore spatial resolution. The encoding path be gins with the C1 layer , which applies a 5 × 5 con v olution k ernel to the input image patch, producing 50 feature maps. These feature maps are do wnsampled by the S1 layer using a 2 × 2 sampling windo w for maximum pooling. F ollo w- ing this, t h e C2 layer applies another 5 × 5 con v olution k ernel to the output of S1, ag ain resulting in 50 feature maps. Finally , the S2 layer performs a second do wnsampling operation with a 2 × 2 maximum pooling windo w , outputting 50 feature maps with encoded semantic information. T o generate the dense pix el-wise classication map, the decoding path emplo ys upsampling layers via transposed con v olutions. The encoded features are rst upsampled by a transposed con v olutional layer using a 3 × 3 k ernel with a stride of 2, doubling the spatial di- mension. A subsequent transposed con v olution layer res tores the feature maps to the original input resolution, mitig ating the loss of spatial details caused by the pooling operations. 2.1.5. Dual-optimized classier f or Pixel-wise segmentation The features e xtracted by the DCNN are then concatenated to form t he input for the crop se gmentation step through the dual classier step. This step is sho wn in green in Figure 1. T o adapt the dual-classi er concept for pix el-le v el se gmentation output, a stack ed feature renement strate gy is emplo yed: − First layer: feature reconstruction. The DCNN output feature maps are reconstructed for each pix el, forming a rened spectral–spatial representation. − Second layer: pix el-wise classication head. The reconstructed features are concatenated with the original DCNN outputs and passed through a 1×1 con v olutional layer with softmax acti v ation to generate per -pix el class probabilities. This ef fecti v ely e xpands the feature representation and impro v es se gmentation accurac y , especially for minority classes. This approach allo ws the netw ork to maintain the adv anta ges of dual optimization, including enhanced feature representation and impro v ed discrimination, while producing dense pix el-le v el outputs suitable for se gmenta- tion rather than center -patch labels used in classical h yperspectral classication. Int J Elec & Comp Eng, V ol. 16, No. 4, August 2026: 2074-2086 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 2079 2.2. Dataset The U A V -HSI-Crop dataset is a publicly a v ailable datase t which s erv es as a standardized be n c hmark for e v aluating h yperspectral se gmentation methods proposed by Niu et al. [23]. The dataset w as collected in Shenzhou City , China. The dataset consists of tw o scene s: Area A ( 864 × 1618 pix els) and Area B ( 2332 × 959 pix els), captured using a Pika L sensor with 200 spectral bands (400–1000 nm) and a 0.1 m ground sampling distance (GSD). As illustrated in Figure 2, the dataset reects a comple x real-w orld agricultural scenario containing 27 crop cate gories ( e .g . , corn, cabbage) planted in irre gular , smallholder plots. This spatial fragmentation and high spectral similarity among classes pro vide a rigorous en vironment for benchmarking spatial feature e xtraction ag ainst models lik e HSI-T ransUNet. Figure 2. V isualization of the U A V -HSI-Crop dataset sho ws the U A V images, ground truth annotation, and class label map Inte gr ating principal component analysis in spatial-spectr al ... (Ale xander Calvin) Evaluation Warning : The document was created with Spire.PDF for Python.
2080 ❒ ISSN: 2088-8708 2.3. Ev aluation metrics The model performance e v aluated both in term of accurac y as well as performance. The se gme n t ation accurac y w as e v aluated using three commonly adopted metrics in h yperspectral image analysis, e .g . , Kappa coef cient ( κ ), o v erall accurac y (O A), and a v erage accurac y (AA). These metrics pro vide complementary perspecti v es on model performance, considering both class-wise accurac y and o v erall agreement. The Kappa coef cient quanties the agreement between predicted and true classications, adjusted for the agreement e xpected by chance, e xpressed as in (7) κ = P o − P e 1 − P e (7) where P o = P C i =1 n ii N is the observ ed agreement (equi v alent to O A), and P e = P C i =1 n i + N · n + i N is the e xpected agreement by random chance. Here, n + i is the number of samples predicted to be in class i . A higher Kappa v alue indicates stronger agreement be yond chance [29]. The O A measures the proportion of correctly classied samples out of the total number of s amples, e xpressed as in (8) O A = P C i =1 n ii N (8) where n ii is the number of correctly predicted samples for class i , C is the total number of classes, and N is the total number of test samples. Whil e O A is straightforw ard to interpret, it may be biased in the presence of class imbalance [30]. The AA computes the mean of indi vidual class accuracies, e xpressed as in (9) AA = 1 C C X i =1 n ii n i + (9) where n i + is the total number of ground truth samples in class i . This metric pro vides a balanced vie w of classication performance across all classes, re g ardless of class frequenc y [21]. Model ef cienc y w as also e v aluated using four metrics, e .g . , total parameters , peak memory usage, to- tal training time, and inference time. T otal parameters and peak memory serv e as proxies for model comple xity and hardw are requirements. T emporal ef cienc y is measured by the total traini ng duration until EarlyStopping and the a v erage inference time, which determines suitability for real-time deplo yment. 2.4. Experimental setup T o determine the optimal h yperparameters, a series of preliminary e xperiments w as rst conducted. A critical parameter for the spatial tunnel is the number of components, k , retained after dimensionality reduction. The model’ s performance w as e v aluated with k set to { 3 , 5 , 10 , 20 } . The results, summarized in T able 1, sho w that k = 3 achie v es the highest κ (0.8557) and O A (0.8784) while also being the most ef cient in terms of memory usage and training duration. While higher v alues of k retain more v ariance, the y appear to introduce redundant or noisy information that de grades classication performance for this dataset. Therefore, k = 3 w as selected for all subsequent e xperiments. T able 1. Performance comparison on dif ferent number of components Components P aram. Memory κ O A AA 3 271.1K 9.1 GB 0.8557 0.8784 0.4614 5 273.6K 9.1 GB 0.8551 0.8779 0.4562 10 279.9K 9.1 GB 0.8553 0.8780 0.4634 20 292.4K 9.1 GB 0.8530 0.8761 0.4528 Based on the tuning process and the model architecture, the nal training conguration w as estab- lished. F or all subsequent results, each e xperiment w as repeated three times to collect the mean and standard Int J Elec & Comp Eng, V ol. 16, No. 4, August 2026: 2074-2086 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 2081 de viation. All netw ork v ariants were trained using the Adam optimizer with an initial learning rate of 1 × 10 − 3 , and a learning rate scheduler that reduced the rate by a f actor of 0.8 on plateau do wn to a minimum of 1 × 10 − 6 with a patience of 10 steps. The model w as trained for a maximum of 300 epochs with a batch size of 32. T o pre v ent o v ertting and ensure the best model state w as sa v ed, an EarlyStopping callback w as added to monitor the v alidation loss with a patience of 50 epochs. T o ensure reproducibility and f acilitate direct comparison with prior w ork, the of cial training, v alidation, and test splits pro vided by the dataset’ s authors in [23] are used. 2.5. Ablation experiments T o v alidate the contrib ution of each component in the architecture, a detailed ablation study w as conducted. The e xperiments are designed to isolate the performance impact of the spatial lt er choice and the indi vidual feature tunnels. The specic congurations are detailed in T able 2. T able 2. Ablation conguration of the netw ork V ariant Spatial lter Spatial tunnel Spectral tunnel LSP Fusion LSP Y es Y es PCA Fusion PCA Y es Y es Spatial T unnel LSP LSP Y es No Spatial T unnel PCA PCA Y es No Spectral T unnel - No Y es − Performance analysis of feature tunnels: T o v erify the indi vidual contrib utions of the spatial and spectral tunnels, by comparing the full dual-tunnel netw ork ag ainst v ariants that utilize only the spectral tunnel or only the spatial tunnel. This analysis is designed to conrm whether the tw o streams pro vide complementary information that impro v es o v erall classication accurac y when fused. − Performance analysis of spatial lter: T o e v aluate the impact of the spatial pre-processing module, the ”LSP Fusion” is compared ag ainst the ”PCA Fusion” v ariant. This e xperiment directly measures the ef- fecti v eness of LSP’ s local similarity projecti on ag ainst PCA ’ s global v ariance-based approach in creating a discriminati v e representation for the spatial tunnel. 3. RESUL TS AND DISCUSSION This section presents the se gmentation results and a detailed analysis of the computational costs. The performance of the dual-tunnel models is compared ag ainst the HSI-T ransUnet baseline, follo wed by an ablation study to dissect the contrib ution of each architectural component. 3.1. P erf ormance and efciency comparison The primary quantitati v e results for all e v aluated models are presented in T able 3. A clear performance and ef cienc y g ap is observ ed between the dual-tunnel architectures and the HSI-T ransUnet baseline, with the dual-tunnel designs consistently achie ving stronger c lass-le v el rob ustness and substantially lo wer model comple xity . These impro v ements reect the adv antage of lightweight spectra l–spatial feature fusion when compared to high-parameter transformer -based models. T able 3. Se gmentation performance comparison Method κ O A AA HSI-T ransUnet 0.8347 ± 0.0033 0.8605 ± 0.0026 0.4108 Spectral tunnel 0.8624 ± 0.0033 0.8838 ± 0.0027 0.4736 ± 0.0174 LSP fusion 0.8517 ± 0.0016 0.8749 ± 0.0012 0.4702 ± 0.0142 Spatial tunnel LSP 0.7797 ± 0.0020 0.8157 ± 0.0014 0.3631 ± 0.0074 PCA fusion 0.8582 ± 0.0036 0.8800 ± 0.0033 0.4918 ± 0.0020 Spatial tunnel PCA 0.8076 ± 0.0024 0.8386 ± 0.0019 0.3978 ± 0.0006 The PCA fusion model, which represents the best-performing proposed conguration, achie v es a κ of 0.8582, an O A of 0.8800, and an AA of 0.4918. This surpasses the HSI-T ransUnet baseline by approximately 2.4% in κ , 2.0% in O A, and most notably , o v er 8.1% in AA. This substantial increase in AA indicates that t h e Inte gr ating principal component analysis in spatial-spectr al ... (Ale xander Calvin) Evaluation Warning : The document was created with Spire.PDF for Python.
2082 ❒ ISSN: 2088-8708 proposed method is especially ef fecti v e at impro ving class-balanced performance relati v e to the T ransformer - based baseline. Furthermore, this accurac y is achie v ed with a signicantly more ef cient model. The PCA Fusion model contains only 271.1K parameters, a reduction of o v er 152.7-fold compared to the 41.4M param- eters of the HSI-T ransUnet. Although the number of parameters is greatly reduced, peak memory usage stays the same at 9.1 GB for all dual-tunnel v ariants. This is mainly because the high-dimensional h yperspectral cube must still be loaded into GPU memory for f ast patch e xtraction during training, not because the models themselv es are more comple x. As a result, most of the memory is consumed by data loading and frame w ork o v erhead rather than the netw ork architecture. It is important to ackno wledge the hardw are conditions re g arding the computational ef cienc y com- parison in T able 4. The HSI-T ransUNet baseline results are reported based on an NVIDIA GeF orce R TX 2080 T i GPU (11 GB memory) [23], while our e xperiments were trained on an NVIDIA T esla T4 (16 GB memory). Notably , the R TX 2080 T i typically of fers higher oating-point performance ( ∼ 13.4 TFLOPS) compared to the T esla T4 ( ∼ 8.1 TFLOPS). Despite being trained on inferior hardw are, our PCA Fusion model still achie v ed a 14.2-fold reduction in training time (507 s vs. 7200 s) and a 4.6-fold speedup in inference (26 s vs. 120 s). Therefore, the ef cienc y comparison should be interpreted as indicating a f a v orable practical trade-of f rather than a hardw are-normalized runtime benchmark. T able 4. Computational cost and ef cienc y comparison Note: Baseline HSI-T ransUnet v alues based on NVIDIA R TX 2080 T i [23]; Models trained on NVIDIA T4 Method P aram. Memory T rain time Infer . time HSI-T ransUnet 41.4M 11.0 GB 7200 s 120 s LSP fusion 271.1K 9.1 GB 507 s 26 s PCA fusion 271.1K 9.1 GB 507 s 26 s Spatial tunnel 182.5K 9.1 GB 473 s 23 s Spectral tunnel 222.4K 9.1 GB 504 s 24 s 3.2. Ablation study T o v alidate the dual-tunnel design and understand the contrib ution of each component, an ablation study w as performed by e v aluating the indi vidual tunnels and the choice of spatial lter . 3.2.1. Analysis of featur e tunnels The spectral tunnel v ariant, which represents the spectral-only tunnel, achie v es a κ of 0.8624 and an O A of 0.8838. This result is e xceptionally strong on its o wn, demonstrating that the spectral stream serv es as the primary ”w orkhorse” of the model, pro vidi ng the v ast majority of the discriminati v e po wer . When this spectral tunnel is complemented by the spatial tunnel (PCA Fusion), the global metrics sho w a slight v ariation with a κ of 0.8582 and an O A of 0.8800. This slight drop is a common trade-of f of se v ere class imbalance. Because scarce cate gories lack suf cient statistical v olume, a purely spectral model naturally biases its predictions to w ard the dominant majority classes to maximize o v erall accurac y . Ho we v er , the inclusion of the spatial stream is critical for class-wise balance. As sho wn in T able 3, the AA i ncreases from 0.4736 (spectral tunnel) to 0.4918 (PCA Fusion). This v alidates the h ypothesis that the tw o streams pro vide complementary information, pro vided the appropriate spatial lter is applied. While the spectral tunnel e xcels at pix el-le v el identication, the spatial tunnel pro vides conte xtual information that helps rene the recognition of minority classes, as illustrated in Figure 3. Specically , the re gions indicated by red rectangles demonstrate ho w the PCA-based spatial tunnel corrects ne structural details and pre v ents minority classes from being absorbed by dominant neighbors, directly dri ving the impro v ement in AA. 3.2.2. Analysis of spatial lter (PCA vs. LSP) A direct comparison between the tw o spatial ltering methods re v eals a consi stent adv antage for PCA, particularly in impro ving A v erage Accurac y . − In the full model: The PCA Fusion ( κ = 0.8582, O A = 0.8800, AA = 0.4918) outperforms the LSP Fusion ( κ = 0.8517, O A = 0.8749, AA = 0.4702). − As a standalone tunnel: The Spatial T unnel PCA model achie v es reliable performance ( κ = 0.8076, O A = 0.8386, AA = 0.3978), pro ving it is a strong feature e xtractor . In contrast, the Spatial T unnel LSP performs signicantly w orse across all metrics ( κ = 0.7797, O A = 0.8157, AA = 0.3631). Int J Elec & Comp Eng, V ol. 16, No. 4, August 2026: 2074-2086 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 ❒ 2083 Figure 3. Qualitati v e se gmentation results of the ablation study across the three e xperimental re gions Inte gr ating principal component analysis in spatial-spectr al ... (Ale xander Calvin) Evaluation Warning : The document was created with Spire.PDF for Python.