Generasi Citra Multispektral dari RGB menggunakan Modifikasi Model Pix2Next untuk Estimasi Biomassa, Klorofil, dan Deteksi Kesehatan pada Agroforestri Kakao

Rasyidah, Dalila Anwar Dini (2026) Generasi Citra Multispektral dari RGB menggunakan Modifikasi Model Pix2Next untuk Estimasi Biomassa, Klorofil, dan Deteksi Kesehatan pada Agroforestri Kakao. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241024-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy

Abstract

Akuisisi data multispektral, terutama pada kanal Near-Infrared (NIR) dan Red Edge, merupakan komponen penting dalam pemantauan vegetasi dan pertanian presisi, namun terkendala oleh biaya sensor yang tinggi dan keterbatasan akses di daerah terpencil. Penelitian ini mengusulkan pendekatan sintesis citra multispektral dari input RGB menggunakan arsitektur Pix2Next yang dimodifikasi berbasis Generative Adversarial Network (GAN). Modifikasi arsitektur mencakup tiga komponen utama: Serial Dual Attention untuk menangkap fitur spasial dan spektral secara berurutan, Physical-Aware Bottleneck untuk menjaga konsistensi spektral, dan Split Head Architecture dengan tiga output head terpisah untuk kelompok kanal spektral yang berbeda. Fungsi loss yang digunakan terdiri dari adversarial loss, feature matching loss, SSIM loss, dan physical loss berbasis indeks vegetasi. Dataset citra Unmanned Aerial Vehicle (UAV) DJI Phantom 4 Multispektral dari plot agroforestri kakao di Côte d'Ivoire pada ketinggian 80 meter digunakan untuk pelatihan dan evaluasi. Hasil evaluasi menunjukkan performa sintesis dengan SSIM rata-rata 0,90 pada lima kanal target. Kanal Red mencapai performa tertinggi (SSIM 0,96; PSNR 42,11 dB), diikuti oleh Blue (SSIM 0,94; PSNR 40,87 dB), Green (SSIM 0,91; PSNR 36,27 dB), Red Edge (SSIM 0,84; PSNR 30,11 dB), dan NIR (SSIM 0,73; PSNR 22,39 dB). Dibandingkan dengan baseline Pix2Next, model yang diusulkan menunjukkan peningkatan signifikan terutama pada kanal NIR (+21,2%) dan Blue (+2,5%). Perbandingan dengan metode lain (Pix2Pix dan TaijiGNN) menunjukkan bahwa model yang diusulkan menghasilkan performa terbaik pada semua kanal spektral. Validasi pada aplikasi downstream menunjukkan korelasi tinggi antara indeks vegetasi sintetik dan ground truth untuk viii NDVI (r = 0,90; R² = 0,80), SAVI (r = 0,90), dan CIgreen (r = 0,74). Klasifikasi kesehatan tanaman berbasis NDVI mencapai akurasi 71,20% dengan Cohen's Kappa 0,57 (moderate agreement), sedangkan estimasi kandungan klorofil berbasis CIgreen mencapai akurasi 64,70% dengan Cohen's Kappa 0,47. Hasil penelitian menunjukkan bahwa kombinasi Serial Dual Attention, Physical-Aware Bottleneck, dan Split Head Architecture efektif meningkatkan kualitas sintesis citra multispektral, khususnya pada kanal NIR yang memiliki keterbatasan informasi dalam input RGB. Citra multispektral sintetik yang dihasilkan dapat dimanfaatkan untuk aplikasi klasifikasi kesehatan vegetasi dan perhitungan indeks vegetasi dengan tingkat akurasi moderat. Pendekatan ini berpotensi menjadi alternatif bagi praktisi pertanian yang memiliki keterbatasan akses ke sensor multispektral.
========================================================================================================================
Multispectral data acquisition, particularly Near-Infrared (NIR) and Red Edge bands, is an essential component in vegetation monitoring and precision agriculture, yet remains constrained by high sensor costs and limited accessibility in remote areas. This research proposes a multispectral image synthesis approach from RGB input using a modified Pix2Next architecture based on Generative Adversarial Networks (GAN). The architectural modifications comprise three main components: Serial Dual Attention for capturing spatial and spectral features sequentially, Physical-Aware Bottleneck for maintaining spectral consistency, and Split Head Architecture with three separate output heads for different spectral band groups. The loss function consists of adversarial loss, feature matching loss, SSIM loss, and physical loss based on vegetation indices. Unmanned Aerial Vehicle (UAV) imagery from a DJI Phantom 4 Multispectral captured over cocoa agroforestry plots in Côte d'Ivoire at 80-meter altitude was used for training and evaluation. Evaluation results demonstrate synthesis performance with an average SSIM of 0.90 across five target bands. The Red band achieved the highest performance (SSIM 0.96; PSNR 42.11 dB), followed by Blue (SSIM 0.94; PSNR 40.87 dB), Green (SSIM 0.91; PSNR 36.27 dB), Red Edge (SSIM 0.84; PSNR 30.11 dB), and NIR (SSIM 0.73; PSNR 22.39 dB). Compared to the baseline Pix2Next, the proposed model showed significant improvements particularly in NIR (+21.2%) and Blue (+2.5%) bands. Comparison with other methods (Pix2Pix and TaijiGNN) demonstrated that the proposed model achieved the best performance across all spectral bands. Downstream application validation revealed high correlation between synthetic and ground truth vegetation indices for NDVI (r xi = 0.90; R² = 0.80), SAVI (r = 0.90), and CIgreen (r = 0.74). NDVI-based plant health lassification achieved 71.20% overall accuracy with Cohen's Kappa of 0.57 (moderate agreement), while CIgreen-based chlorophyll content estimation achieved 64.70% accuracy with Cohen's Kappa of 0.47. The findings demonstrate that the combination of Serial Dual Attention, Physical-Aware Bottleneck, and Split Head Architecture effectively improves multispectral image synthesis quality, particularly for the NIR band which has limited information in RGB inputs. The synthesized multispectral imagery can be utilized for vegetation health classification and vegetation index computation applications with moderate accuracy levels. This approach potentially serves as an alternative for agricultural practitioners with limited access to multispectral sensors.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	agroforestri kakao, generative adversarial network, multispektral, penginderaan jauh,translasi citra. cocoa agroforestry, generative adversarial network, multispectral, remote sensing, image translation
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Dalila Anwar Dini Rasyidah
Date Deposited:	29 Jan 2026 08:56
Last Modified:	29 Jan 2026 08:56
URI:	http://repository.its.ac.id/id/eprint/131199

Actions (login required)

View Item