UAV-IRE : Pengembangan Superresolusi pada Citra UAV untuk Segmentasi Gulma Lahan Persawahan

Huda, Muhamad Syaiful (2026) UAV-IRE : Pengembangan Superresolusi pada Citra UAV untuk Segmentasi Gulma Lahan Persawahan. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6002241012-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (1MB) | Request a copy

Abstract

Gulma merupakan faktor yang menurunkan produktivitas tanaman padi yang berdampak langsung pada ketahanan pangan nasional. Penerapan pertanian cerdas dengan dukungan citra Unmanned Aerial Vehicle (UAV) dan kecerdasan buatan menjadi solusi untuk pengendalian gulma secara efisien dan ramah lingkungan. Namun, kualitas citra UAV yang menurun akibat kondisi operasional drone menjadi tantangan utama bagi akurasi segmentasi gulma. Penelitian ini mengembangkan model super-resolusi berbasis Improved Real-ESRGAN (IRE) yang diadaptasi secara khusus untuk karakteristik citra UAV, selanjutnya disebut UAV-IRE, melalui integrasi UAV-Specific Degradation Pipeline serta penambahan empat modul arsitektural, yaitu Noise-aware Residual Denoising Block (NRDB), Motion-Blur Compensation Module (MBCM), Edge-Guided Attention (EGA), dan Vegetation Similarity Discriminator (VSD). Model dilatih dan diuji menggunakan dataset WeedyRice-RGBMS-DB. Hasil pengujian menunjukkan model UAV-IRE menghasilkan kualitas citra dengan PSNR sebesar 27,2512 dB dan SSIM sebesar 0,9217 pada data uji. Selanjutnya, pengaruh super-resolusi terhadap performa segmentasi gulma diuji menggunakan model U-Net dengan backbone EfficientNet-B5 yang dilatih secara terpisah pada citra hasil super-resolusi (SR) dan citra resolusi tinggi asli (HR). Hasil pengujian menunjukkan model yang dilatih pada citra SR memperoleh IoU Gulma sebesar 0,7468, lebih tinggi dibandingkan model yang dilatih pada citra HR asli sebesar 0,7264, atau meningkat sekitar 2,81%. Hasil ini mengindikasikan bahwa metode UAV-IRE yang diusulkan mampu mempertahankan, bahkan meningkatkan, informasi tekstur yang relevan bagi tugas segmentasi gulma dibandingkan citra resolusi tinggi aslinya, sehingga berpotensi menjadi tahap preprocessing yang layak diterapkan dalam sistem segmentasi gulma berbasis citra UAV pada lahan persawahan.
====================================================================================================================================
Weeds are a factor that reduces rice plant productivity, directly impacting national food security. The implementation of smart agriculture supported by Unmanned Aerial Vehicle (UAV) imagery and artificial intelligence has emerged as a solution for efficient and environmentally friendly weed control. However, the degraded quality of UAV imagery caused by drone operational conditions poses a major challenge to weed segmentation accuracy. This research develops a super-resolution model based on Improved Real-ESRGAN (IRE) that is specifically adapted to the characteristics of UAV imagery, hereafter referred to as UAV-IRE, through the integration of a UAV-Specific Degradation Pipeline along with the addition of four architectural modules, namely the Noise-aware Residual Denoising Block (NRDB), Motion-Blur Compensation Module (MBCM), Edge-Guided Attention (EGA), and Vegetation Similarity Discriminator (VSD). The model was trained and tested using the WeedyRice-RGBMS-DB dataset. The test results show that the UAV-IRE model produces image quality with a PSNR of 27.2512 dB and SSIM of 0.9217 on the test data. Furthermore, the effect of super-resolution on weed segmentation performance was evaluated using a U-Net model with an EfficientNet-B5 backbone, trained separately on super-resolution (SR) images and original high-resolution (HR) images. The test results show that the model trained on SR images achieved a Weed IoU of 0.7468, higher than the model trained on the original HR images at 0.7264, representing an increase of approximately 2.81%. These results indicate that the proposed UAV-IRE method is able to preserve, and even enhance, texture information relevant to the weed segmentation task compared to the original high-resolution images, suggesting its potential as a viable preprocessing stage for UAV-image-based weed segmentation systems in rice paddy fields.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Super-resolusi citra, UAV, Improved Real-ESRGAN, segmentasi gulma, smart farming, Image super-resolution, weed segmentation
Subjects:	Q Science > QA Mathematics Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44101-(S2) Master Thesis
Depositing User:	Muhamad Syaiful Huda
Date Deposited:	01 Aug 2026 05:39
Last Modified:	01 Aug 2026 05:39
URI:	http://repository.its.ac.id/id/eprint/138625

Actions (login required)

View Item