Inter national J our nal of Inf ormatics and Communication T echnology (IJ-ICT) V ol. 14, No. 2, August 2025, pp. 393 ∼ 404 ISSN: 2252-8776, DOI: 10.11591/ijict.v14i2.pp393-404 ❒ 393 Categorizing h yperspectral imagery using con v olutional neural netw orks f or land co v er analysis Assia Nouna 1,2 , Soumaya Nouna 2 , Mohamed Mansouri 2 , Achchab Boujamaa 2 1 Laboratory LAMSAD, ENSA Berrechid, Department of Mathematics and Informatics, Hassan First Uni v ersity of Settat, Berrechid, Morocco 2 Laboratory LAMSAD, ENSA Berrechid, Hassan First Uni v ersity of Settat, Berrechid, Morocco Article Inf o Article history: Recei v ed Aug 30, 2024 Re vised No v 28, 2024 Accepted Dec 15, 2024 K eyw ords: Con v olutional neural netw ork Deep learning HSI classication Hyperspectral images K-nearest neighbors Support v ector machine ABSTRA CT Cate gorizing h yperspectral imagery (HSI) is crucial in v arious remot e sensing applications, i ncluding en vironmental monitoring, agriculture, and urban plan- ning. Recently , numerous approaches ha v e emer ged, with con v olutional neu- ral netw ork (CNN)-based algorithms demonstrating remarkable performance in HSI classication due to their ability to learn comple x spatial-spectral features. Ho we v er , these algori thms often require signicant computational resources and storage capacity , which can be limiting in practical applications. In this study , we propose a no v el CNN architecture tailored for HSI classi cation within the spectral domain, focusing on optimizing computational ef cienc y without com- promising accurac y . The architecture l e v erages adv anced spectral feature e xtrac- tion techniques to enhance classication performance. Experimental e v aluations on multiple benchmark h yperspectral datasets re v eal that the proposed approach not only impro v es classication accurac y b ut also achie v es a superior balance between performance and computational demand compared to traditional meth- ods lik e K-nearest neighbors (KNN) and other deep learning-based techniques. Our results demonstrate the potential of the proposed CNN model in adv ancing the eld of HSI classication, of fering a via ble solution for real-w orld applica- tions with constrained computational resources. This is an open access article under the CC BY -SA license . Corresponding A uthor: Assia Nouna Laboratory LAMSAD, ENSA Berrechid, Department of Mathematics and Informatics Hassan First Uni v ersity of Settat Berrechid, Morocco Email: a.nouna@uhp.ac.ma 1. INTR ODUCTION Remote sensors capture h yperspectral imagery (HSI) [1], which contain se v eral hundred as man y channels of observ ation at highly spe ctral resolution. The richness of the spectral information present in HSI allo ws for the de v elopment of numerous traditional c lassication approaches, including K-nearest neighbors (KNN), logistic re gression, and minimum distance [2]. In recent times, there ha v e been proposed some im- pro v ed methods for e xtracting features and adv anced classiers, for instance, the spectral and spatial classi- cation [3] and Fisher’ s local discriminant analysis, which ha v e sho wn to be more ef fecti v e. The support v ector machine (SVM) [4] is widely re g arded as a reliable and ef fecti v e approach to the tasks of h yperspectral classication, mainly when dealing with limited training samples. SVM functions by seeking an optimal decision h yperplane that can separate tw o-class data by utilizing a high-dimensional feature J ournal homepage: http://ijict.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
394 ❒ ISSN: 2252-8776 space included in the k ernel. Se v eral adaptations to the SVM for the classication of HSIs ha v e been introduced in the current literature to enhance classication performance [5], [6]. The classication of remote sensing data has been e xplored using neural netw orks, lik e the multi- layer perceptron (MLP) [7] and the radial basis function neural netw orks. Ratle et al. [8], a s emisupervised neural netw ork frame w ork w as propos ed for HSI classication on a lar ge scale. Ho we v er , in remote sensing classication tasks, SVM has been sho wn to outperform classi cal neural netw orks in classication accurac y and computational cost. Despite this, [9] considers that deep neural netw ork (DNN) architecture as a po werful classication model that can compete with SVM in terms of classication performance. Deep learning methods ha v e sho wn great potential in v arious domains. Con v olutional neural netw orks (CNNs) [9] are particularly ef fecti v e for proce ssing visual-related tasks within deep learning. CNNs are a class of multilayer models that tak e inspiration from biological neural netw orksthat can be end-to-end trained, starting with the ra w pix els of the image and ending with the classie r . The CNN concept w as initially presented in [10], further de v eloped by [11], and renement and simplication in subsequent w ork [12]. Recently , CNNs ha v e achie v ed better results some of the con v entional approaches, including the hu- man performers [13], in a v ariety of tasks related to vision, such as classication of images [14], [15], detection of objects, the labeling of scenes, classication of house numbers, and recognition of f aces. Additionally , CNNs ha v e also applied t o speech recognition [16] and pro v ed to be ef fecti v e models for understanding visual image content. In a study by [17], researchers utilized CNNs for HSI classication [18], [19], specically using stack ed a utoencoders (SAEs) to e xtract distincti v e features. These ndings conrm that CNNs are an ef cient class of models for solving visual-related problems and can produce state-of-the-art results in visual image classication. It has been sho wn that CNNs are superior in classication performance compared to traditional SVM and DNNs on tasks related to images [20], [21]. There is a lack of literature on the appli cation of CNNs with multiple layers for HSI classication [22], as CNNs ha v e primarily been used in visual-related problems [23]-[25]. The paper demonstrates that h yperspectral data can be accurately classied using CNNs with a suit- able layer architecture. Through e xperiments, it w as observ ed that traditional CNN models, lik e the LeNet-5 based on tw o con v olutional layers, are unsuitable. Instead, a 20-layer CNN architecture with supervised HSI classication weights w as proposed, which w as pro v en to be a more ef fecti v e and straightforw ard approach. Our proposed method has been pro v en to outperform the traditional KNN and con v entional deep l earning ar - chitectures t h r ou gh v arious e xperiments. T o our kno wledge, this is -one of the rst times a multi-layer CNN has been utilized for the HSI classication, resulting in outstanding results. This paper is structured into dif ferent sections. Section 2 pro vides a concise o v ervie w of CNNs, where the traditional CNN structure and -their for - mation process are discussed. In s ection 3, we conduct e xperiments to compare the ef cienc y of our approach to K-nearest neighbor (KNN) and DNN using commonly used datasets. The paper concludes by summarizing the obtained results in section 4. 2. DESCRIPTION OF METHODOLOGY This study presents a no v el CNN architecture specically designed for HSI classication wit h i n the spectral domain. The methodology consists of three primary phases: (i) An introduction to CNNs, empha- sizing their suitability for processing spectral data and their adv antages o v er traditional classication methods lik e KNN; (ii) A detailed e xploration of HSI classication techniques using CNNs, highlighting the inte gration of spectral features; and (iii) A comprehensi v e description of the proposed CNN model, including its archi- tecture, training strate gy , forw ard propag ation, and backpropag ation mechanisms. The training phase utilizes a care fully curated set of h yperspectral datasets, applying rigorous preprocessing a n d e v aluation protocols to ensure high accurac y and rob ust performance. Experimental results demonstrate the proposed method’ s supe- rior clas sication accurac y and ef cienc y compared to con v entional approaches, es tablishing its potential for adv anced remote sensing applications. This section pro vides a step-by-step account of the e xperimental setup, implementation details, and v alidation metrics to ensure replicability and transparenc y of the research ndings. 2.1. Con v olution neural netw ork CNNs produce feed-forw ard neural netw orks composed of di v erse layers of con v ol ution, layers of maximum pooling, and layers of fully connected. A CNN can benet from local spatial correlation by imple- menting a specic pattern of the local connections among the neurons in the adjacent layers. The con v oluted layers are alternated by layers of maximum pooling, reproducing the comple x and simple nature of the cells of Int J Inf & Commun T echnol, V ol. 14, No. 2, August 2025: 393–404 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 395 the mammalian visual corte x. A CNN is composed of a pair or se v eral pairs of maximum pooling and con v o- lution layers and results in a fully connected neural netw ork. The con v olutional netw ork architecture can be found in Figure 1. In other w ords, inside the DNN, e v ery hidden acti v ation x i can be calculated simply by taking the entire input X and multiplying it by the W weights inside this layer . But inside CNN, e v ery hidden acti v ating is cal culated using the multiplication of the short input by the W weights. Then W weights are then distrib uted o v er all the input areas. The neurons which are on the identical layer are gi v en identical weights. W eight sharing is a fundament of CNNs, as it reduces the o v erall number of training components and results in better model performance and more ef cient training. Usually , the con v olutional layer is succeeded by a maximum pooling layer . A CNN detects features across the input data by replicating weights. Ho we v er , when an input image is shifted, the feature detecting neuron also shifts. T o address this, we use pooling to render characteristics in v ariant to location. This is achie v ed by summarizing multiple neuron outputs in con v olutional layers with the pooling function, typically maximal. The max pooling function retrie v es a maximum data v alue of an input, partitioning the input data into the non-o v erlapping windo w and outputting the maximal v alues within e v ery sub-re gion. This enabl es the comple xity of the calculation for the higher layers to be reduced and the in v ariance of the translation to be ensured. Finally , to enable the classication, a CNN’ s calculation chain ends with a fully connected netw ork inte grating the information on all the locations of all the feature maps in the lo wer layer . Figure 1. U-Net architecture 2.2. HSI classication based on the CNN CNN HSI classication is a technique used to classify HSIs using CNN. HSIs are multidimensional data that contain information about the reectance or emission of light at dif ferent w a v elengths of the electro- magnetic spectrum. These images ha v e man y applications in areas such as en vironmental monitoring, mining e xploration and urban planning. The CNN model is trained to classify each pix el in the HSI int o one of se v eral predened classes. The model tak es the HSI as input and applies con v olutional lters to e xtract features from the spectral data. These features are then passed through se v eral layers of the netw ork, including pooling layers, fully connected layers, and acti v ation functions, to reduce the dimensionality and classify the pix els into their respecti v e classes. The main adv antage of usi ng CNN-Based HSI Classication is that it can automatically learn rel e v ant features from the input data, eliminating the need for manual feature e xtraction. This approach allo ws for impro v ed accurac y in classication, as the model can identify subtle dif ferences in the spectral characteristics of dif ferent classes. Additionally , CNN-Based HSI Classication c an handle high-dimensional data, which mak es it well-suited for processing HSIs. Ov erall, the HSI classication based on CNN is a po werful and ef cient technique for the analysis of HSIs and has man y practical applications in v arious elds, including remote sensing, agriculture, and geology . 2.3. The pr oposed HSI classication based on CNN 2.3.1. The ar chitectur e of pr oposed CNN The CNN depends on the w ay in which con v olutional and maximum pooling layers are constructed and the w ay in which netw orks are formed. So, the procedure adopted for HSI class ication in the proposed w ork can be illustrated by using Figure 2. Cate gorizing hyper spectr al ima g ery using con volutional neur al networks . . . (Assia Nouna) Evaluation Warning : The document was created with Spire.PDF for Python.
396 ❒ ISSN: 2252-8776 Figure 2. The proposed CNN classier architecture W ithin our architecture, the HSI data is rst tak en into account and normalized. Hyperspectral data are split into training and test data with mer ged features and are deli v ered as inputs to the CNN algorithm to achie v e classication. The CNN architecture is composed primarily of tree groups ( C 1 , C 2 , C 3 ) of con v olution layers, each g r ou p C 1 contains four layers, so in total we ha v e twelv e layers ( c 1 , c 2 .....c 12 ) and tree pooling layers ( P 1 , P 2 , P 3 ) . Finaly the lters used in the 3 groups are currently K 1 = 128 , K 2 = 64 , and K 3 = 32 respecti v ely . W e considered the input training data size as ( X 1 , Y 1 , 1) , when we applied the rst step with the rst group of con v olutional layer C 1 with K 1 lter , the data are transformed into ( X 1 , Y 1 , K 1 ) and ( X 2 , Y 2 , K 2 ) where X 2 = X 1 2 and Y 2 = Y 1 2 . After applying the second step with the second group of con v olutional layer C 2 with K 2 lter , the data are transformed into ( X 2 , Y 2 , K 2 ) and ( X 3 , Y 3 , K 2 ) where X 3 = X 2 2 and Y 3 = Y 3 2 . Lastly , the data become ( X 3 , Y 3 , K 3 ) after applying the third step with the third group of con v olutional layer C 3 and pooling layers P 3 . The nished output is classied with the SoftMax function. Moreo v er , the dropout and batch normal- ization (BN) layers are emplo yed on the suggested netw ork. 2.3.2. T raining appr oaches In this section, we outline the techniques we emplo yed to train the s pace of parameters for the algo- rithm proposed for the CNN classier . Firstly , e v ery trained parameter with our CNN is randomly initialized to v alues from − 1 to 1 . The formation process comprises tw o main stage s: forw ard propag ation and Backprop- ag ation. During a forw ard propag ation stage, CNN computes a classication output of an input data based on the current set of parameters. Inside a subsequent Back propag ation stage, trainable parameters are updated to minimize the dif ference between the real and desired classication outputs. The ultimate aim of this process is to reduce the discrepanc y between the tw o outputs as much as possible. A. F orward pr opagation: The input layer of our CNN (with L + 1 layers and L = 19 ) consists of v 1 elements, while the output layer comprises v 20 elements. Additionally , there are a number of hidden elements present in the con v olutional layer C , pooling layer P , and fully connected layer F . These layers are commonly referred to as hidden layers. Int J Inf & Commun T echnol, V ol. 14, No. 2, August 2025: 393–404 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 397 Supposing that a i is the input of the i layer , so we calculate a i +1 as, a i +1 = g i ( z i ) (1) where z i = W T i a i + b i (2) and the W T i weight matrix for the i layer operates for the entry data, while the b i v ector of additi v e bias is used for the same layer . The g acti v ation function used in the C -layers and F -layer is the ReLu function. The max( z ) function is applied in the P -layer . Since our CNN classier is designed to classify multiple classes. Finaly the output of layer F used a SoftMax function. This model is dened as a SoftMax re gression model. y = 1 P v 11 h =1 e W T L,h a L + b L,h e W T L, 1 a L + b L, 1 e W T L, 2 a L + b L, 2 . . . e W T L,v 20 a L + b L,v 20 (3) In each iteration, the probability of all classes is represented by the nal probability indicated in the output layer as the y output v ector , which is obtained by multiplying the a input v ector with L + 1 ( y = a L +1 ). B. Back-pr opagation: In the process of back-propag ation, the trainable parameters under go updates through the utilization of the gradient descent technique. This approach in v olv es the minimization of a cost function, coupled with the computation of the partial deri v ati v e of said function concerning each trainable parameter . Once a CNN classier’ s architecture and associated trainable parameters are dened, we ha v e the ability to construct it and reloa d an y sa v ed parameters for the purpose of classifying HSI data. The classication process is analogous to the forw ard propag ation step, whereby we can determine the classication outcome. 3. EXPERIMENT AL EV ALUTION AND AN AL YSES This study focused on e v aluating the ef fecti v eness of a no v el CNN architecture for HSI classication, addressing a g ap in pre vious research where computational ef cienc y and spectral feature e xtraction were often o v erlook ed. The rst section, study area and data collection, describes the use of the P a via Uni v ersity dataset, predominantly consisting of v e getation re gions, to assess the clas sication performance of the proposed model. The second section, results and analysis, demonstrates that the proposed CNN method signicantly impro v es classication acc urac y for 5 out of 9 classes, particularly for v e getation types, compared to tradit ional KNNs methods. These results suggest a strong correlation between spectral feature e xtraction and impro v ed classi- cation outcomes, without compromising computational ef cienc y . The nal section, ndings and comparisons, sho ws that the proposed method outperforms e xisting techni q ue s in terms of accurac y and computational re- quirements, while addressing the limitations of focusing on a single dataset. Future studies could e xtend this approach to more di v erse en vironments, conrming its rob ustness and e xploring optimization strate gies under resource-constrained conditions. Ov erall, this w ork highlights the potential for CNN-based HSI classication to re v olutionize remote sensing applications by enhancing accurac y without increasing resource demands. 3.1. Study ar ea and data collection 3.1.1. Hyperspectral imaging Hyperspectral remote sensing is the e xtraction of information from objects or sceneries on the Earth’ s surf ace using light collected by airborne or spaceborne sensors. Small, commercial, high spatial, and spectral resolution instruments ha v e been increasingly used in lab-scale applications (industries also including en viron- mental management, agriculture, urban planning, and military applications) using h yperspectral sensing and its imaging modality , h yperspectral imaging. Millions of spatial core gistered images corresponding to v arious spectral channels compensate h y- perspectral data. A HSIs structure is as follo ws: each pix el is represented lik e v ector of the B-dimensional Cate gorizing hyper spectr al ima g ery using con volutional neur al networks . . . (Assia Nouna) Evaluation Warning : The document was created with Spire.PDF for Python.
398 ❒ ISSN: 2252-8776 characteristics along the spectral dimension, which is referred to as the substance spectrum within that pix el. This ab undance of data in each spatial area impro v es an ability to discern between v arious ph ysical materials. As a result, h yperspectral photograph y opens up ne w a v enues for the class ication of pictures, a critical step in a v ast range of applications such as precision agriculture. 3.1.2. Data set of h yperspectral images The studies used P a via data sets, each presenting an urban re gion, a spatial resolution of 1.3m per pix el, and se v eral bands of 103 bands. The follo wing section presents this data set. P a via is a dataset acquired o v er the city of P a via, Italy , utilizing the R OSIS sensor and a ground sample distance (GSD) of 1.3 m. In Figure 3, P a via Uni v ersity (103 bands, 610 340px) and P a via Center (102 bands, 1096 715px) are the tw o sections. In this art, we will use the rst part of the data (the P a via Uni v ersity) which contains 9 cate gories of interest labeled throughout half of the surf ace (see in T able 1). The y are made up of v arious urban components (including bricks, asphalt, and metals), as well as w ater and plants. As it is one of the lar gest labeled HSI datasets and enables the e v aluation of the usage of HSI for future applications, it has long been one of the k e y reference datasets. Figure 3. P a via Uni v ersity composite image T able 1. 9 classes of P a via Label Class Samples 1 Asphalt 6631 2 Meado ws 18649 3 Gra v el 2099 4 T rees 3064 5 P ainted metal sheets 1345 6 Bare soil 5029 7 Bitumen 1330 8 Self-blocking bricks 3682 9 Shado ws 947 In the original recorded image, there are 115 data channels (with a spectral range of 0.43 to 0.86 m). The studies were conducted using the remaining 103 bands after the 12 most noisy channels were deleted (see in Figure 4). Int J Inf & Commun T echnol, V ol. 14, No. 2, August 2025: 393–404 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 399 Figure 4. Bands of P a via Uni v ersity 3.2. Result and analysis 3.2.1. Get the data r eady f or training and testing Once the data has been inte grated, it must be fed into the CNN netw ork. Before this step, it is important to split the data into training and testing sets and to remo v e an y background pix el samples. In the case of the P a via Uni v ersity data, 1, 64, 624 pix els were remo v ed for each band. This preparation process helps reduce the computing b urden and more information can be found in T able 2. T able 2. The number of training and test samples in the P a via dataset Label Class Samples T raining T est 1 Asphalt 6631 4641 1990 2 Meado ws 18649 13054 5604 3 Gra v el 2099 1469 630 4 T rees 3064 2144 920 5 P ainted metal sheets 1345 941 404 6 Bare soil 5029 3521 1508 7 Bitumen 1330 932 398 8 Self-blocking bricks 3682 2578 1104 9 Shado ws 947 663 284 T otal 42776 29943 12833 3.2.2. Description of pr ogram The program is written in Python using the K eras API, which is a high-le v el neural netw orks API that can run on top of T ensorFlo w . The Sequential class is imported from the K eras library and is used to initialize a sequential model. The model consists of multiple layers, including Con v1D, BN, MaxPooling1D, Dropout, and Dense (see in T able 3). Cate gorizing hyper spectr al ima g ery using con volutional neur al networks . . . (Assia Nouna) Evaluation Warning : The document was created with Spire.PDF for Python.
400 ❒ ISSN: 2252-8776 T able 3. Model layers Layer (type) Output shape P aram # inputLayer (Con v1D) (None, 101, 128) 512 Batch-normalization (Batch-normalization) (None, 101, 128) 512 Layer1 (Con v1D) (None, 99, 128) 49,280 Layer2 (Con v1D) (None, 97, 128) 49,280 Layer3 (Con v1D) (None, 95, 128) 49,280 Layer4 (Con v1D) (None, 93, 128) 49,280 MaxPooling Layer1 (MaxPooling1D) (None, 46, 128) 0 Dropout1 (Dropout) (None, 46, 128) 0 Layer5 (Con v1D) (None, 44, 64) 24,640 Layer6 (Con v1D) (None, 42, 64) 12,352 Layer7 (Con v1D) (None, 40, 64) 12,352 Layer8 (Con v1D) (None, 38, 64) 12,352 MaxPooling Layer2 (MaxPooling1D) (None, 19, 64) 0 Dropout2 (Dropout) (None, 19, 64) 0 Layer9 (Con v1D) (None, 17, 32) 6,176 Layer10 (Con v1D) (None, 15, 32) 3,104 Layer11 (Con v1D) (None, 13, 32) 3,104 Layer12 (Con v1D) (None, 11, 32) 3,104 MaxPooling Layer3 (MaxPooling1D) (None, 5, 32) 0 Dropout3 (Dropout) (None, 5, 32) 0 Flatten (Flatten) (None, 160) 0 DenseLayer (Dense) (None, 25) 4,025 OutputLayer (Dense) (None, 10) 260 Con v1D is a one-dimensional con v olutional layer that applies a con v olution operation on the input data with a specied number of lters and k ernel size. The acti v ation function used for the con v olutional layers is ReLU. BN is used to normalize the input data by adjusting and scaling the acti v ations of the pre vious layers. MaxPooling1D is used to reduce the spatial size of the data by taking the maximum v alue within a specied pool size. Dropout is used to pre v ent o v er tting by randomly dropping out a certain percentage of the input units. Flatten is used to con v ert the multidimensional input data into a one-dimensional array . Dense is a fully connected layer that applies a linear operation on the input data. The output layer uses the soft-max acti v ation function to output the probability distrib ution of the classes. The model is then summarized using the model.summary() function, which outputs the layers, shapes, and parameters of the model. 3.2.3. Findings and comparisons This study in v estig ated the ef fecti v eness of a no v el CNN architecture for HSI classication, focusing on impro ving accurac y and computational ef cienc y . While earlier studies ha v e successfully applied CNNs to HSI classication, the y often focus on spatial feature e xtraction or dimensionality reduction without addressing the balance between high classication accurac y and resource ef cienc y . This g ap becomes critical when applying these methods to lar ge-scale datasets or edge computing en vironments, where computational resources are limited. Once the data w as input into the CNN architecture (Figure 2), the model follo wed a standard train- ing procedure with 100 epochs, a batch size of 256, cate gorical cross entrop y loss, and an Adam optimizer . T able 3 outlines the specications of each CNN layer , pro viding a detailed breakdo wn of the architecture. The P a via Uni v ersity dataset w as used for e v aluating the classication accurac y , and the results were compared ag ainst the traditional KNN classier . T ables 4 and 5 summarize the classication accurac y for each class, sho wing that the proposed method signicantly outperformed KNN in 5 out of 9 classes, achie ving comparable results in 2 classes, while no notable lo w accurac y w as observ ed in the remaining classes. Our k e y ndings indicate that the proposed CNN method consistently yields higher classication ac- curac y for v e getation classes, such as wheat and corn, compared to KNN. This higher accurac y correlates with the ability of CNNs to e xtract comple x spectral features, which is critical for distinguishing subtle v ariations in crop types. The classication maps (Figures 5 and 6) demonstrate a clear adv antage of the proposed method, particularly in i d e ntifying and classifying v e getation areas with high precision, as conrmed by the ground truth data. Int J Inf & Commun T echnol, V ol. 14, No. 2, August 2025: 393–404 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 401 T able 4. Accurac y for proposed method on P a via data sets Method Accurac y Our proposed method CNN 0.95 KNN 0.87 T able 5. Accurac y per class for the data of the Uni v ersity of P a via in comparison of our method with the KNN method Label Class KNN Porposed method 1 Asphalt 0.92 0.96 2 Meado ws 0.94 0.97 3 Gra v el 0.75 0.83 4 T rees 0.91 0.96 5 P aint ed metal sheets 0.95 1.0 6 Bare soil 0.74 0.93 7 Bitumen 0.83 0.91 8 Self-blocking bricks 0.81 0.88 9 Shado ws 0.98 1.0 Figure 5. P a via Uni v ersity classication map Figure 6. Clasication for our proposed method map When comparing these results to other studies, our method sho ws that impro v ed spectral feature e x- traction does not ne g ati v ely impact computational ef cienc y , a k e y limitation in pre vious w orks. F or e xample, unlik e studies focusing on spatial features alone, our method e xploits spectral domain data, enabling superior performance in classication without the hea vy computational b urden typically associated with CNNs. Ho we v er , this study w as limited to the P a via Uni v ersity dataset, which primarily contains v e geta- tion data. The potential impact of this limitation is that our ndings may not generalize to dat asets con- taining more di v erse or urban en vironments. Further research is needed to v alidate the method’ s rob ustness across dif ferent types of h yperspect ral data and en vironments. Our results s uggest that the proposed CNN method is more resilient to spectral noise and outperforms traditional classiers in v e getation-related classi- cation tasks. Future studies could e xplore the application of this architecture to other h yperspectral datasets, fo- cusing on optimizing CNN performance under limited computational resources while maintaining high accurac y . Cate gorizing hyper spectr al ima g ery using con volutional neur al networks . . . (Assia Nouna) Evaluation Warning : The document was created with Spire.PDF for Python.
402 ❒ ISSN: 2252-8776 In conclusion, the ndings from this study pro vide strong e vidence that the proposed CNN ar chitecture for HSI classication of fers a balanced approach, enhanci ng classication accurac y in comple x spectral datasets without increasing computational costs. These impro v ements could signicantly benet remote sensing appli- cations, particularly in agriculture and en vironmental monitoring, where ef cient and accurate classication is crucial. 4. CONCLUSION In this study , a ne w technique using CNNs for classifying HSIs is presented. First, the data is normal- ized to retrie v e both spatial and spectral features. A resulting HSI image is then combined with the original input HSI and fed into a proposed CNN, which comprises three sets of pooling and con v olution layers. T o im- pro v e the method’ s accurac y , we ha v e incorporated BN and dropout mechanisms. The classication approach is e v aluated on three standard datasets and has been sho wn to outperform e xisting state-of-the-art approaches. Future w ork will concentrate on reducing an algorithm’ s running time and applying a proposed method to a broader range of HSI datasets using 2D and 3D CNNs. FUNDING INFORMA TION Authors state no funding in v olv ed. A UTHOR CONTRIB UTIONS ST A TEMENT Name of A uthor C M So V a F o I R D O E V i Su P Fu Assia Nouna ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Soumaya Nouna ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Mohamed Mansouri ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Achchab Boujamaa ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ 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 Authors state no conict of interest. D A T A A V AILABILITY he data that support the ndings of this study are openly a v ailable in the P a via Uni v ersity repository at [https://www .ehu.eus/ccwintco/inde x.php/Hyperspectral Remote Sensing Scenes]. REFERENCES [1] D. Landgrebe, “Hyperspectral image data analysis, ” IEEE Signal Processing Mag azi ne , v ol. 19, no. 1, pp. 17–28, 2002, doi: 10.1109/79.974718. [2] G. M. F oody and A. Mathur , “ A relati v e e v aluation of multiclass image cl assication by support v ector machines, ” IEEE T ransactions on Geoscience and Remote Sensing , v ol. 42, no. 6, pp. 1335–1343, Jun. 2004, doi: 10.1109/TGRS.2004.827257. [3] Y . T arabalka, J. A. Benedikts son, a nd J. Chanus sot, “Spectral-spatial classication of h yperspectral imagery based on parti- tional clustering techniques, ” IEEE T ransactions on Geoscience and Remote Sensing , v ol . 47, no. 8, pp. 2973–2987, Aug. 2009, doi: 10.1109/TGRS.2009.2016214. [4] J. A. Gualtieri and S. Chettri, “Support v ector machines for classi cation of h yperspectral data, ” in International Geoscience and Remote Sensing Symposium (IGARSS) , 2000, v ol. 2, pp. 813–815, doi: 10.1109/ig arss.2000.861712. Int J Inf & Commun T echnol, V ol. 14, No. 2, August 2025: 393–404 Evaluation Warning : The document was created with Spire.PDF for Python.