Generasi Citra Multispektral dari RGB menggunakan Model UNet-ResNet untuk Estimasi Karbon, Prediksi Tingkat Stres, dan Produktivitas pada Agroforestri Kakao

Simanihuruk, Laurensia (2026) Generasi Citra Multispektral dari RGB menggunakan Model UNet-ResNet untuk Estimasi Karbon, Prediksi Tingkat Stres, dan Produktivitas pada Agroforestri Kakao. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241050-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Citra multispektral memiliki peran penting dalam monitoring agroforestri karena kemampuannya menangkap informasi spektral Red-Edge dan Near-Infrared (NIR) yang krusial untuk analisis kesehatan tanaman, estimasi biomassa, dan deteksi stres vegetasi. Namun, kamera multispektral umumnya mahal dan sulit diakses oleh petani kecil serta peneliti di negara berkembang. Penelitian ini mengembangkan Model Triple Cascaded Decoder UNet-ResNet50 untuk menghasilkan citra multispektral sintetis 5-band (Blue, Green, Red, Red-Edge, NIR) dari input RGB. Model mengintegrasikan tiga decoder bertingkat dengan correlation-adaptive loss weighting, Atrous Spatial Pyramid Pooling (ASPP), dan attention mechanism untuk menangani perbedaan korelasi spektral antar band, dengan pelatihan menggunakan dataset dari UAV pada lahan agroforestri kakao di Côte d'Ivoire. Evaluasi komparatif terhadap tiga metode state-of-the-art menunjukkan keunggulan signifikan metode yang diusulkan. Untuk band Blue, TC-SA-ResUNet mencapai SSIM 0,9515±0,0136 dan PSNR 40,46±2,39 dB, mengungguli T-GAN (Zhang et al., 2022) sebagai metode terbaik kedua dengan SSIM 0,9052±0,0227. Peningkatan paling signifikan terlihat pada band NIR dimana metode yang diusulkan mencapai SSIM 0,7209±0,0661, meningkat 33% dibanding UNet baseline (Zeng et al., 2021) dengan SSIM 0,5417±0,0267 dan 12% dibanding TGAN dengan SSIM 0,6412±0,0278. Keunggulan ini diatribusikan pada integrasi ASPP-SE di bottleneck untuk pemahaman konteks multi-skala, penerapan PhysicsGuided Spectral Loss (RedEdge-NIR Correlation dan NDVI Consistency Loss) yang memaksa jaringan mematuhi prinsip biofisik vegetasi, serta strategi Correlation-Guided Adaptive Weighting yang memberikan fokus optimasi intensif pada spektrum sulit diprediksi. Model mencapai composite score 0,8543 dengan SSIM band Blue 0,9515, Green 0,8796, Red 0,9182, Red-Edge 0,8014, dan NIR 0,7209. Rekonstruksi indeks vegetasi menunjukkan performa tinggi dengan NDVI mencapai MAE 0,0388, R² 0,7607, dan korelasi 0,9067. Validasi aplikasi menunjukkan Overall Accuracy 64,8% untuk prediksi tingkat stres, overall agreement 47,9% dengan deteksi area produktivitas tinggi 94,3%, serta estimasi stok karbon dengan R² 0,5468 dan efektivitas relatif 79% dari ground truth. Penelitian ini memvalidasi bahwa citra multispektral sintetis dapat digunakan sebagai alternatif hemat biaya untuk pemantauan agroforestri kakao tanpa memerlukan sensor multispektral mahal.
==================================================================================================================================
Multispectral imagery plays a crucial role in agroforestry monitoring due to its capability to capture Red-Edge and Near-Infrared (NIR) spectral information essential for plant health analysis, biomass estimation, and vegetation stress detection. However, multispectral cameras are generally expensive and difficult to access for smallholder farmers and researchers in developing countries. This study develops a Triple Cascaded Decoder UNet-ResNet50 Model to generate synthetic 5-band multispectral imagery (Blue, Green, Red, Red-Edge, NIR) from RGB input. The model integrates three cascaded decoders with correlation-adaptive loss weighting, Atrous Spatial Pyramid Pooling (ASPP), and attention mechanism to handle spectral correlation differences among bands, trained using UAV datasets from cocoa agroforestry plantations in Côte d'Ivoire. Comparative evaluation against three state-of-the-art methods demonstrates significant superiority of the proposed method. For the Blue band, TC-SA-ResUNet achieves SSIM of 0.9515±0.0136 and PSNR of 40.46±2.39 dB, outperforming TGAN (Zhang et al., 2022) as the second-best method with SSIM of 0.9052±0.0227. The most significant improvement is observed in the NIR band where the proposed method achieves SSIM of 0.7209±0.0661, showing 33% improvement over UNet baseline (Zeng et al., 2021) with SSIM of 0.5417±0.0267 and 12% over T-GAN with SSIM of 0.6412±0.0278. This uperiority is attributed to the integration of ASPP-SE in the bottleneck for multi-scale context understanding, implementation of Physics-Guided Spectral Loss (RedEdge-NIR Correlation and NDVI Consistency Loss) that enforces biophysical principles of vegetation, and Correlation-Guided Adaptive Weighting strategy that provides intensive optimization focus on difficult-to-predict spectra. The model achieves a composite score of 0.8543 with SSIM for Blue band 0.9515, Green 0.8796, Red 0.9182, Red-Edge 0.8014, and NIR 0.7209. Vegetation index reconstruction demonstrates high performance with NDVI achieving MAE of 0.0388, R² of 0.7607, and correlation of 0.9067. Application validation shows Overall Accuracy of 64.8% for stress level prediction, overall agreement of 47.9% with high productivity area detection of 94.3%, and carbon stock estimation with R² of 0.5468 and relative effectiveness of 79% from ground truth. This study validates that synthetic multispectral imagery can be used as a cost-effective alternative for cocoa agroforestry monitoring without requiring expensive multispectral sensors.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Kata Kunci: agroforestri kakao, deep learning, estimasi stok karbon, sintesis citra multispektral, UNet-ResNet50 Keywords: cocoa agroforestry, deep learning, carbon stock estimation, multispectral image synthesis, UNet-ResNet50
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:	Laurensia Simanihuruk
Date Deposited:	30 Jan 2026 02:28
Last Modified:	30 Jan 2026 02:28
URI:	http://repository.its.ac.id/id/eprint/131166

Actions (login required)

View Item