Utafiti huu unaunganisha bendi za B8, B11 na B12 za Sentinel-2; usanifu wa payload ya optik-elektroniki, uundaji wa modeli ya sensa ya InGaAs, usanifu wa deep learning wa Hybrid CNN–ResNet50 na mbinu ya Grey Wolf Optimization (GWO) katika mfumo mmoja kwa lengo la kutofautisha moto wa misitu kutoka picha za satellite za short-wave infrared (Short-Wave Infrared, SWIR). Waandishi wameripoti %91,03 accuracy, %91,27 precision, %91,03 recall na %91,01 F1-score kwa mfumo kamili. Hata hivyo, nambari katika confusion matrix ya sampuli 465 iliyoonyeshwa na utafiti hailingani kimahesabu na performance metrics hizi; overall accuracy inayokokotolewa kutoka matrix cells ni takriban %94,62. Kwa hiyo, ingawa inaweza kusemwa kwamba mbinu imetoa matokeo ya classification yanayoahidi, asilimia halisi ya mafanikio haiwezi kuthibitishwa kwa consistency ndani ya chanzo.

Input kuu ya picha iliyotumika katika utafiti haikutoka real-time kwenye SWIR camera iliyotengenezwa na watafiti. Experimental classification ilifanywa kwenye offline satellite images katika BC Wildfire Sentinel-2 dataset. Optical lens system, SWIR filters, field of view, electronic signal chain na behavior ya InGaAs sensor ziliundwa kwa dhana na kwa hisabati kwa SWIR optical-electronic payload inayoweza kutekelezwa baadaye. Kwa hiyo, utafiti hauthibitishi performance ya physical fire-detection payload inayofanya kazi orbit au kwenye aircraft.

Katika mbinu, bendi ya B8 near-infrared (NIR), B11 SWIR-1 na B12 SWIR-2 zilibadilishwa ukubwa hadi 128 × 128 pixels, zikanormalize na kuunganishwa katika three-channel input tensor ya 128 × 128 × 3. Fire-related spectral features zilishughulikiwa kwa Normalized Burn Ratio (NBR), Normalized Difference Vegetation Index (NDVI) na thermal anomaly approach inayotegemea tofauti ya B12–B11. Hybrid CNN–ResNet50 model iliunganisha local spatial features za CNN branch na deeper semantic features za ResNet50 branch; GWO ilitumika kutafuta hyperparameters kama learning rate, batch size, dropout na trainable ResNet50 layers.

Kwa mtazamo wa Uturuki: SWIR-based remote sensing inatoa mbinu ya thamani ya kutafiti kwa satellite, UAV au local observation payloads zitakazoendelezwa baadaye katika maeneo ya Uturuki yenye hatari kubwa ya moto wa misitu. Hata hivyo, matokeo ya utafiti huu yanategemea Sentinel-2 dataset ya British Columbia na binary-classification setup maalum. Kwa kuwa vegetation, topography, atmospheric conditions, fire types na sensor geometry nchini Uturuki zinaweza kutofautiana, performance percentages zilizoripotiwa haziwezi kuhamishwa moja kwa moja kwenda Uturuki; local data na validation kwa real hardware vinahitajika.

Tatizo ambalo utafiti unajaribu kutatua ni lipi?

Fire-monitoring systems zinazotegemea visible-spectrum images zinaweza kuwa na shida kutofautisha targets chini ya smoke, haze na low-light conditions. Mbinu kuu ya utafiti ni kutumia ukweli kwamba short-wave infrared spectral information inaweza kubeba taarifa tofauti na visible image kuhusu burned surface, vegetation moisture/stress na baadhi ya thermal anomalies.

Literature gap iliyobainishwa na waandishi si higher classification accuracy pekee. Wanasema tafiti nyingi zilizopo zinalenga software-level image classification; hazijaunganisha vya kutosha SWIR optical system, sensor, electronic signal chain, spectral feature extraction na AI optimization katika end-to-end architecture moja.

End-to-end system iliyopendekezwa inafanyaje kazi?

Katika Kielelezo 1 cha utafiti, mchakato umetolewa katika hatua tano kuu: kupata Sentinel-2-based images, preprocessing, SWIR spectral feature extraction, classification kwa Hybrid CNN–ResNet50 na GWO-based optimization. Katika hatua ya mwisho, model inatathminiwa kwa accuracy, precision, recall na F1-score.

Sambamba na software chain hii, watafiti wanafafanua payload architecture inayowakilisha upande wa kimwili wa future SWIR remote-sensing system: optical lens, SWIR filter, InGaAs detector, signal conditioning, analog-to-digital conversion, embedded controller na data-transmission module.

Picha za Sentinel-2 ziliandaliwaje?

Original dataset ina 516 PNG images: 258 wildfire na 258 non-wildfire samples. Baada ya data augmentation, jumla ya images iliongezwa hadi 3096. Katika Jedwali 1 la utafiti, images 2167 (%70) zimetolewa kama training set na images 464 (%15) kama validation set. Test row haipo kwenye table, lakini results section inaeleza kwamba 465 test samples zilitumika; 2167 + 464 + 465 ni 3096.

Kila bendi ya B8, B11 na B12 ilibadilishwa ukubwa kwanza hadi 128 × 128 pixels na kunormalize kando. Kisha bendi hizo tatu ziliunganishwa kwenye channel axis kuunda input tensor ya 128 × 128 × 3.

Image resizing na normalization zilifafanuliwaje?

Katika chanzo, resizing operation imeelezwa kwa uhusiano huu:

\[ I_r(x,y)=I\left(x\frac{W}{W_r},y\frac{H}{H_r}\right) \]

Hapa \(I_r(x,y)\) ni pixel husika ya resized image; \(I(x,y)\) ni original image; \(W\) na \(H\) ni original width na height; \(W_r\) na \(H_r\) ni target width na height.

Min–max normalization iliyotumika kuleta intensity scales kwenye common range ni:

\[ I_n=\frac{I-I_{min}}{I_{max}-I_{min}} \]

ambapo \(I_n\) ni normalized pixel value, \(I_{min}\) na \(I_{max}\) zinawakilisha minimum na maximum intensities za image husika.

Ni spectral bands zipi zilitumika?

Three-channel classification input iliundwa kutoka B8, B11 na B12:

\[ B_{SWIR}=\{B11,B12,B8\} \]

B8 hubeba near-infrared (NIR), B11 SWIR-1 na B12 SWIR-2 information. Utafiti unaeleza hasa kwamba B11 na B12 zilitumika kunasa spectral differences zinazohusiana na burned area, surface conditions na vegetation moisture/stress.

Burned area ilitathminiwaje kwa NBR?

Normalized Burn Ratio (NBR) inalinganisha B8 na B12:

\[ NBR=\frac{B8-B12}{B8+B12} \]

Kielelezo 7 cha utafiti kinaonyesha NBR maps za wildfire na non-wildfire scenes pamoja na difference map. Kwa tafsiri ya chanzo, lower NBR values zinapokaribia −1 zinahusishwa na burned areas, na higher values zinazoelekea +1 zinahusishwa na healthier vegetation.

Figure pia inaonyesha classes za high burn severity kwa NBR < −0,44; moderate burn kati ya −0,44 na −0,27; low burn kati ya −0,27 na −0,10; unburned area kati ya −0,10 na 0,10; na healthy vegetation juu ya 0,10. Thresholds hizi ni classification iliyotumika katika visual ya utafiti na hazipaswi ku-generalize automatically kwa ecosystems tofauti.

Vegetation stress na thermal anomaly zilishughulikiwaje?

Normalized Difference Vegetation Index ilitumika kwa vegetation analysis:

\[ NDVI=\frac{B8-B4}{B8+B4} \]

Hapa B8 ni NIR na B4 ni red visible band. Equation hii kwa namna ya kuzingatia hutumia B4 band ambayo haipo katika three-channel model input; NDVI imefafanuliwa ndani ya spectral feature analysis, huku classification tensor ikijengwa kwa B8/B11/B12.

Thermal anomaly indicator imeelezwa katika chanzo kwa simple SWIR band difference:

\[ T_a=B12-B11 \]

\(T_a\) ni thermal anomaly value, na B12 na B11 ni band intensities husika. Uhusiano huu si direct physical surface-temperature measurement; utafiti unautumia kama anomaly indicator inayotegemea tofauti kati ya SWIR band intensities.

SWIR optical payload ilimodeliwaje?

Payload iliyopendekezwa ina optical lens system, SWIR filters, imaging optics na field-of-view setup. Thin-lens relation ni:

\[ \frac{1}{f}=\frac{1}{d_o}+\frac{1}{d_i} \]

ambapo \(f\) ni focal length, \(d_o\) ni object distance kati ya Earth surface na lens, na \(d_i\) ni image distance.

Optical magnification imefafanuliwa kama:

\[ M=\frac{d_i}{d_o} \]

Imefafanuliwa kwa uhusiano huu.

Katika chanzo, target wavelength range ya SWIR filter imetolewa kama:

\[ \lambda_{SWIR}=1.0\,\mu m-2.5\,\mu m \]

Spectral transmittance ya filter imeelezwa kwa:

\[ T_{SWIR}=\frac{I_t}{I_i} \]

ambapo \(I_t\) ni transmitted na \(I_i\) ni incident SWIR radiation intensity.

Field of view na ground sampling distance zinaamuliwaje?

Kwa Field of View (FOV):

\[ FOV=2\tan^{-1}\left(\frac{S}{2f}\right) \]

imetumika. \(S\) inawakilisha sensor size na \(f\) focal length. FOV pana hufunika ground area kubwa, huku FOV nyembamba inaweza kutoa higher target resolution.

Ground Sampling Distance (GSD) imeelezwa kwa:

\[ GSD=\frac{H p}{f} \]

ambapo \(H\) ni satellite altitude, \(p\) pixel size na \(f\) focal length. Katika results graph ya utafiti, GSD value ya developed sensor system imeonyeshwa kuwa 50,0 m.

Data inasongaje katika electronic subsystem?

SWIR detector hubadilisha incoming optical radiation kuwa electric current:

\[ I_d=R_{\lambda}P_{opt} \]

\(I_d\) ni detector output current, \(R_{\lambda}\) spectral responsivity na \(P_{opt}\) incoming optical power.

Weak signal kisha huamplify katika signal-conditioning circuit:

\[ V_o=A_vV_i \]

Quantization step ya analog-to-digital converter imefafanuliwa kama:

\[ Q=\frac{V_{max}-V_{min}}{2^n} \]

Imefafanuliwa kwa namna hiyo. Processing time kwa embedded controller:

\[ T_p=\frac{N_i}{f_c} \]

na data-transmission rate:

\[ R_t=\frac{D_t}{T_t} \]

zimeundwa. Hizi ni general engineering relations zinazoeleza system architecture; utafiti haukupima experimentally real-time performance ya electronic chain hii kwenye flight hardware maalum.

InGaAs sensor model inatumia parameters zipi?

InGaAs SWIR sensor model imefafanuliwa kwa spectral responsivity, noise, dynamic range na pixel resolution. Spectral responsivity:

\[ R_{\lambda}=\frac{I_o}{P_i} \]

noise-equivalent power:

\[ NEP=\frac{N}{R_{\lambda}} \]

na dynamic range:

\[ DR=20\log_{10}\left(\frac{S_{max}}{S_{min}}\right) \]

zimetolewa.

Kielelezo 8 cha utafiti kinaonyesha 70,5 dB dynamic range, 50,0 m GSD na 100,0 ke− full-well capacity kwa proposed sensor system. Hizi si values zilizopimwa kutoka real orbital test, bali system parameters zilizotengenezwa/kumodeliwa katika utafiti.

Kwa nini Hybrid CNN–ResNet50 imegawanywa katika branches mbili?

CNN branch ya model hutumia convolution, ReLU na pooling layers mfululizo ili kujifunza more local spatial features. Convolution operation imetolewa katika chanzo kama:

\[ F(i,j)=\sum_m\sum_n I(i-m,j-n)K(m,n) \]

Hapa \(I\) ni input image, \(K\) convolution kernel na \(F\) resulting feature map.

ReLU activation ni:

\[ A(x)=\max(0,x) \]

Imefafanuliwa kwa namna hiyo.

Katika ResNet50 branch, residual connection imeelezwa kwa:

\[ H(x)=F(x)+x \]

Muundo huu hutumika kusaidia feature transfer katika deeper layers na kupunguza vanishing gradient problem.

Feature vectors kutoka branches mbili:

\[ F_{fusion}=[F_{CNN},F_{ResNet50}] \]

huunganishwa na kisha kupelekwa kupitia dense layers kwenye two-class Softmax output.

Model ilifundishwa kwa hyperparameters zipi?

Imeripotiwa kwamba training computer ilitumia 12th-generation Intel Core i5-12400F processor, 16 GB RAM na 64-bit Windows 11 Pro. GPU information haijatolewa katika hardware table ya utafiti.

GWO inaoptimize nini hasa?

Katika Grey Wolf Optimization section, kila candidate solution imefafanuliwa kama:

\[ X_i=[LR_i,BS_i,DR_i,TL_i] \]

Hapa LR ni learning rate, BS batch size, DR dropout rate na TL number of trainable ResNet50 layers.

Search ranges:

• Learning rate: 1 × 10⁻⁵ – 1 × 10⁻³
• Batch size: 16 – 64
• Dropout rate: 0,2 – 0,5
• Trainable layer: 20 – 80

zimetolewa.

Kwa hiyo, kulingana na mathematical method ya utafiti, quantities ambazo GWO inaoptimize wazi ni deep-learning hyperparameters. Ingawa katika sehemu kadhaa text inasema GWO hufanya “payload design optimization” au “feature selection”, physical focal length, FOV, sensor pixel size au ADC parameters hazipo katika GWO solution vector. Kwa hiyo, claims hizi hazipaswi kuchukuliwa kama ushahidi wa physical payload optimization tofauti na hyperparameter optimization.

GWO convergence iliripotiwaje?

Katika Kielelezo 6, Alpha Wolf fitness value iko karibu 84,5 katika iterations za mwanzo; inafikia 89,0 katika iteration ya nne, 91,0 katika iteration ya tano na kubaki 91,0 kati ya iterations 5–10. Chanzo kinatafsiri hili kama fast convergence.

Matokeo ya classification yanaonyesha nini?

Waandishi wameripoti metrics zifuatazo kwa framework kamili:

Hata hivyo, confusion matrix katika results section hiyo hiyo inaonyesha 221 correct non-wildfire, 219 correct wildfire, na misclassifications 12 na 13 kwa test images 465. Ikikokotolewa moja kwa moja kutoka matrix cells:

\[ Accuracy=\frac{221+219}{465}\approx0.9462 \]

yaani takriban %94,62. Wildfire class ikichukuliwa kuwa positive, matrix hiyo hiyo inatoa takriban %94,8 precision, %94,4 recall na %94,6 F1-score. Values hizi haziendani na metrics za karibu %91 zilizoripotiwa na makala.

Kwa hiyo, kwa mtazamo wa Verianla, njia sahihi si kuwasilisha reported %91,03 accuracy kama performance ya uhakika na internally verified ya utafiti; bali kusema wazi kwamba waandishi wameripoti %91,03, lakini published confusion matrix inaonyesha performance tofauti.

Ni kutokuwiana gani kwa pili katika confusion matrix?

Kulingana na axes za Kielelezo 3, true non-wildfire row ina 221 correct non-wildfire na 12 incorrect wildfire classifications; true wildfire row ina 13 incorrect non-wildfire na 219 correct wildfire. Kwa upande mwingine, maelezo ya text kuhusu numbers 12 na 13 kuassigniwa vibaya katika class gani yanaonekana kwa order iliyogeuzwa. Matrix visual na textual description hazilingani kikamilifu katika hatua hii.

Matokeo ya spectral response yanaonyesha nini?

Kielelezo 10 cha utafiti kinatoa average normalized spectral response ya bendi tatu kwa numbers sahihi. B12 SWIR-2 ina highest average normalized response ya 0,6543; B8 NIR ina 0,5491; na B11 SWIR-1 ina 0,4467.

Average normalized spectral response ya juu ya B12 katika dataset hii imehusishwa na waandishi na higher sensitivity kwa surface conditions na vegetation changes. Hata hivyo, average values hizi tatu pekee hazithibitishi kwamba B12 ni bora kuliko bendi nyingine katika fire scenarios zote.

Matokeo ya SNR yanaonyesha nini?

Kielelezo 9 kinaripoti SNR value ya 1,884 kwa non-wildfire images na 2,378 kwa wildfire images. Waandishi wanatafsiri higher wildfire SNR kama support kwamba proposed SWIR approach inaweza kutofautisha fire-related signal.

Hata hivyo, inapaswa kuzingatiwa kwamba values hizi hazikutokana na experimental physical InGaAs payload measurement, bali kutoka study framework inayotegemea Sentinel-2 images. Kwa hiyo hazipaswi kutafsiriwa kama field signal-to-noise performance ya real sensor.

Ablation analysis inasema nini?

Katika Jedwali 4 la utafiti, reported accuracy ya full framework ni 0,9103. Baadhi ya components zinapoondolewa, chanzo kinatoa results zifuatazo:

Chanzo kinatafsiri table hii kama ushahidi kwamba modules zote zinachangia classification. Hata hivyo, kwa kuwa physical SWIR payload au real InGaAs camera haikutumika katika experiments halisi, rows za “bila payload design” na “bila sensor modeling” hazipaswi kutathminiwa kama hardware ablation ambapo physical hardware iliondolewa na experiment kurudiwa. Utafiti unaripoti role ya components hizi ndani ya software/modeling framework.

Matokeo yanayoungwa mkono na utafiti

• Three-channel data iliyoundwa kutoka Sentinel-2 B8, B11 na B12 images iliweza kutumiwa kwa wildfire/non-wildfire classification kwa Hybrid CNN–ResNet50.
• Waandishi wameripoti %91,03 accuracy, %91,27 precision, %91,03 recall na %91,01 F1-score kwa full model.
• Imeonyeshwa kwamba GWO inaoptimize learning rate, batch size, dropout na trainable ResNet50 layer count katika defined search space.
• B12 band ilionyesha highest average normalized spectral response ya 0,6543 kati ya bendi tatu zilizochaguliwa katika utafiti.
• Chanzo kinaripoti SNR value ya 2,378 kwa wildfire images na 1,884 kwa non-wildfire images.
• NBR maps zinaonyesha spectral difference kati ya wildfire na non-wildfire scenes.
• Proposed architecture inaunganisha image preprocessing, spectral features, CNN–ResNet50 na GWO katika software framework moja.

Matokeo ambayo utafiti hauungi mkono au haujathibitisha

• Haijathibitishwa kwamba real SWIR optical-electronic payload inatambua fire kwa %91,03 accuracy katika orbit au aircraft.
• Physical prototype ya InGaAs sensor haikutumiwa katika utafiti kuzalisha data badala ya real Sentinel-2.
• Onboard real-time processing performance haijathibitishwa experimentally.
• Haijaonyeshwa kwamba performance ileile itadumishwa chini ya heavy cloud, different atmospheric conditions na different geographies.
• Kwa kuwa reported %91,03 accuracy haiwezi kurekebishwa kimahesabu na confusion matrix, haijathibitishwa independently kama exact success rate.
• Haijaonyeshwa kwa mathematical solution vector kwamba GWO inaoptimize physical lens, FOV, sensor au electronic payload parameters.
• Haijathibitishwa kwamba B12 ni universal best band katika fire-detection systems zote.
• Matokeo ya utafiti hayawezi ku-generalize moja kwa moja kwa forest fires nchini Uturuki kwa accuracy values zilezile.

Mbinu na Matokeo ya Utafiti

Experimental data na preprocessing chain

Utafiti unatumia BC Wildfire Sentinel2 Dataset. Original images 516 ziliongezwa hadi samples 3096 baada ya data augmentation. Training, validation na test samples zilizotajwa katika results section ni 2167, 464 na 465 mtawalia. Kwa kila image, B8, B11 na B12 bands zililetwa hadi 128 × 128 pixels, zikanormalize na kuunda three-channel input.

Utafiti unaonyesha augmentation function kwa general form:

\[ I_a=T(I) \]

ambapo \(T\) ni transformation function inayotumika kwa image. Hata hivyo, chanzo hakitoi detailed types zote za augmentation operations na distribution ya kila transformation kama fully reproducible protocol zaidi ya mathematical equation.

SWIR sensor na electronic model

Katika model, conversion ya incoming SWIR radiation kuwa electrical signal na sensor, amplification ya signal, digitization kwa ADC, processing katika embedded controller na transmission zimefafanuliwa kama sequential electronic chain. InGaAs sensor model inajumuisha responsivity, noise-equivalent power, dynamic range na pixel resolution.

Hata hivyo, utafiti hauripoti laboratory calibration kwa real sensor part number, real optical-lens tolerance measurements, MTF test, detector-temperature-dependent dark-current measurement au orbit-level radiometric calibration. Kwa hiyo, section hii inapaswa kusomwa kama sensor/payload design model.

Deep learning na optimization

CNN na ResNet50 ni parallel feature-extraction branches mbili. ResNet50 ilianzishwa kwa ImageNet pretrained weights, na layers 60 za mwisho zikaachwa trainable ili kuadapt kwa wildfire data. Feature vectors mbili ziliunganishwa katika fusion layer na kubadilishwa kuwa wildfire/non-wildfire class kwa Softmax.

Cross-entropy loss:

\[ L=-\sum_{i=1}^{C}y_i\log(P_i) \]

iliminimize. \(y_i\) ni true class label, \(P_i\) prediction probability na \(C\) class count.

GWO fitness value imefafanuliwa katika chanzo kwa accuracy:

\[ F=\frac{TP+TN}{TP+TN+FP+FN} \]

Optimal hyperparameter solution ni:

\[ X_{best}=\arg\max(F_i) \]

.

Detailed check ya confusion matrix na reported metrics

Matrix iliyoonyeshwa katika Kielelezo 3 ni:

Cells hizi nne ni total samples 465 na 440 zimeclassify correctly. Kwa hiyo accuracy ya visual matrix yenyewe ni takriban %94,62. Haijaelezwa %91,03 value katika Figure 4 na text ya utafiti ilitoka kwa different prediction set au calculation stage gani.

Katika tathmini ya Verianla, tofauti hii imechukuliwa kuwa critical kwa sababu classification accuracy ni mojawapo ya main claims za utafiti. Badala ya kuunda assumption isiyoelezwa katika chanzo au kukubali kimya kimya moja ya numbers kama “sahihi”, data zote mbili zimetolewa pamoja.

Bibliographic problem katika comparison table

Table 5 ya chanzo inalinganisha proposed model na methods mbili za awali. Hata hivyo, comparison ya pili imewasilishwa kama “Deep Learning Wildfire Detection Using Multi-Source Remote Sensing [32]”, wakati reference 32 katika bibliography si wildfire classification model bali different study kuhusu fire management. Kwa hiyo bibliographic match ya reference ya pili katika Table 5 haiwezi kuthibitishwa ndani ya chanzo.

Aidha, ingawa column names za Table 5 zina percentage sign, values zimeandikwa kama 0,8840, 0,8932 na 0,9103 katika scale ya 0–1. Kimaana kuna uwezekano kwamba hizi ni %88,40, %89,32 na %91,03 mtawalia; lakini table format ina scale-label inconsistency kama ilivyo katika chanzo.

Nguvu za utafiti

• Hauchukulii fire detection kama RGB image classification pekee, bali unalenga SWIR spectral information.
• Unafafanua wazi Sentinel-2 B8/B11/B12 bands.
• Unajadili physical/spectral indicators kama NBR, NDVI na SWIR band difference katika framework ileile na deep learning approach.
• Unatoa architecture wazi inayounganisha Hybrid CNN na ResNet50 features.
• Unafafanua GWO search space na optimized hyperparameters kwa hisabati.
• Unatoa ablation analysis na kujaribu kulinganisha athari ya model components kwenye reported performance.
• Unaeleza wazi katika discussion section kwamba real-time hardware validation haikufanywa.

Mapungufu makuu ya utafiti

• Real SWIR payload hardware haikutengenezwa na kujaribiwa experimentally.
• Experiments zinategemea offline Sentinel-2 images.
• Real-time onboard processing validation haikufanywa.
• Heavy cloud na adverse atmospheric conditions zinaweza kupunguza performance.
• Computational cost ya Hybrid CNN–ResNet50 na GWO ni kubwa.
• Kuna serious numerical inconsistency kati ya confusion matrix na reported main performance metrics.
• Katika comparison table, angalau reference number moja hailingani na study iliyoelezwa.
• Baadhi ya claims kwamba GWO hufanya physical payload optimization hazijaungwa mkono moja kwa moja na defined solution variables.
• Katika ablation, effect ya physical payload/sensor components haijaonyeshwa kwa real hardware-removal experiment.

Mapendekezo ya future work

Waandishi wanapendekeza future experimental validation ya system kwenye real SWIR payload na embedded hardware, real-time implementation kwa lighter deep-learning architectures, multimodal sensor fusion na transformer-based models. Kuongeza UAV na hyperspectral images pia ni miongoni mwa development directions zilizotajwa.

Maelezo ya Chanzo na Mbinu

Jina kamili la utafiti asilia: Design and Development of a SWIR Optical-Electronic Payload for Earth Remote Sensing Applications

Waandishi: Ainur Zhetpisbayeva; Samal Kaliyeva; Berik Zhumazhanov; Almira Mukhamejanova; Ainur Satpayeva; Aliya Kargulova.

Mpangilio wa waandishi: Original order katika chanzo imehifadhiwa kama ilivyo.

Mchango sawa/mwandishi wa kwanza sawa: Haijatajwa katika chanzo.

Mwandishi wa mawasiliano: Samal Kaliyeva.

Taasisi: Department of Radio Engineering, Electronics and Telecommunications, Institute of Physical and Technical Sciences, L.N. Gumilyov Eurasian National University, Astana, Kazakhstan; Ghalam LLP, Astana, Kazakhstan; Department of Telecommunication Engineering, Almaty University of Power Engineering and Telecommunications Named After Gumarbek Daukeyev, Almaty, Kazakhstan; Department of Electric Power Supply NCJSC, S. Seifullin Kazakh Agro Technical Research University, Astana, Kazakhstan.

Aina ya chanzo: Makala ya utafiti iliyopitiwa na rika. Inaunganisha deep-learning classification inayotegemea offline Sentinel-2 images na conceptual SWIR optical-electronic payload pamoja na InGaAs sensor modeling.

Jarida: Aerospace

Mchapishaji: MDPI

Juzuu / nambari ya makala: 13, 649

Tarehe ya kuchapishwa: 17 Julai 2026

DOI: 10.3390/aerospace13070649

Kiungo rasmi cha uchapishaji: https://www.mdpi.com/2226-4310/13/7/649

Kiungo cha DOI: https://doi.org/10.3390/aerospace13070649

Hali ya mapitio: Ni makala ya utafiti iliyochapishwa katika jarida la Aerospace lililopitiwa na rika.

Leseni: Creative Commons Attribution (CC BY) open-access license.

Ufadhili: Umefadhiliwa na Science Committee of the Ministry of Science and Higher Education of the Republic of Kazakhstan chini ya Research Identification Registration Number (IRN) BR27198365.

Upatikanaji wa data: Imeelezwa kwamba data zilizowasilishwa katika utafiti zinaweza kupatikana kutoka corresponding authors kwa ombi.

Taarifa ya mgongano wa maslahi: Source text inatumia kauli “All authors were employed by the company Ghalam LLP” na kisha kusema hakuna commercial or financial relationship kwa “remaining authors”. Hata hivyo, affiliation list katika makala inaassign Ghalam LLP kwa baadhi ya authors pekee. Visible inconsistency hii haijasahihishwa kimya kimya.

Maelezo ya michango ya waandishi: Katika chanzo, investigation contribution imetolewa kama “S.K., A.S., B.Z., B.Z. and A.Z.” na B.Z. imerudiwa mara mbili. Verianla haikadirii repetition hii ilipaswa kuwa jina la author gani.

Performance consistency note: Waandishi wameripoti %91,03 accuracy, %91,27 precision, %91,03 recall na %91,01 F1-score. Kwa upande mwingine, confusion matrix ya 221/12/13/219 katika Kielelezo 3 inalingana na overall accuracy ya takriban %94,62. Chanzo hakielezi sababu ya difference hii. Kwa hiyo Verianla haijabadilisha kimya kimya reported value kwa matrix-derived value au kinyume chake.

Comparative-source note: Comparison ya “Deep Learning Wildfire Detection Using Multi-Source Remote Sensing” iliyoonyeshwa kwa [32] katika Table 5 inaonekana kutolingana bibliographically na reference 32 katika bibliography. Kwa hiyo comparison hiyo haijaandikwa upya kwa independent bibliographic accuracy assumption.

GWO scope note: GWO solution vector inajumuisha learning rate, batch size, dropout rate na trainable layers parameters. Kwa hiyo explicitly validated optimization scope ya method ni deep-learning hyperparameters; haiwezi kutolewa kutoka equation set hii kwamba physical SWIR payload parameters zimeoptimize na GWO.

Kikomo cha kisayansi: Utafiti haukufanya real-time SWIR payload implementation au onboard hardware validation. Classification experiments zilifanywa kwa offline Sentinel-2 images. Sehemu ya optical payload na InGaAs sensor ni model/conceptual architecture kwa future physical implementation.

Mbinu ya uzalishaji wa maudhui: Scientific methods, equations, numerical data, graph interpretations na limitations katika maelezo haya ya Verianla yanategemea source study iliyochunguzwa. Matumizi ya external sources yalipunguzwa kwa bibliographic verification ya study identity na peer-review status ya journal; hakuna new scientific finding au performance data kutoka nje iliyoongezwa kwenye main article.