Segmentasi Semantik Gulma dan Tanaman Jagung Menggunakan Recurrent Residual Attention U-Net dengan Backbone EfficientNet-B3 pada Citra Udara Multispektral

Rahman, Irvan Abdul (2026) Segmentasi Semantik Gulma dan Tanaman Jagung Menggunakan Recurrent Residual Attention U-Net dengan Backbone EfficientNet-B3 pada Citra Udara Multispektral. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221044-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Gulma adalah tanaman yang tumbuh disekitar tanaman budidaya. Keberadaan gulma dapat merugikan tanaman dan menurunkan hasil produksi bagi petani hingga 60%. Pengendalian gulma dengan penyiangan manual dan penyemprotan herbisida dapat membutuhkan waktu yang lama, biaya yang relatif mahal, dan resiko resistensi gulma. Pengendalian gulma berbasis lokasi spesifik (Site-Specific Weed Management) sangat efektif diterapkan, terutama dalam skala besar secara otomatis. Salah satu tugas dalam sistem tersebut adalah segmentasi. Penelitian ini mengimplementasikan model Recurrent Residual Attention U-Net (R2AU-Net) dengan backbone EfficientNet-B3 dalam tugas segmentasi semantik gulma dan tanaman jagung menggunakan citra udara multispektral. Backbone EfficientNet-B3 dipilih karena kemampuannya dalam menangkap konteks spasial multi-skala secara efektif. Model dibangun dengan dataset WeedsGalore yang terdiri dari kanal spektral, yaitu Red, Green, Blue, Red-Edge dan Near-Infrared. Ukuran citra 600 × 600 piksel dengan jumlah citra beranotasi sebanyak 156 gambar. Dataset diproses melalui tahapan resize, normalisasi, dan augmentasi pada data latih yang mencakup transformasi geometri, transformasi intensitas warna, injeksi noise, dan penerapan kernel filter. Model menghasilkan segmentasi lima kelas yang terdiri dari kelas background, maize, amaranth, barnyard grass, quickweed dan other weeds. Penelitian ini melakukan tiga percobaan berupa konfigurasi masukan tiga kanal, empat kanal, dan lima kanal untuk menemukan kombinasi masukan spektrum pada model terbaik. Hasil percobaan menunjukkan bahwa performa terbaik diperoleh pada model dengan inputan tiga kanal dengan konfigurasi G, NIR, RE yang menghasilkan mIoU sebesar 0.5177, mean dice sebesar 0.5924, dan pixel accuracy sebesar 0.9763.
=================================================================================================================================
Weeds are unwanted plants that grow in association with cultivated crops and compete for essential resources such as nutrients, water, light, and space. Their presence can significantly reduce crop productivity, causing yield losses of up to 60%. Conventional weed control methods, including manual weeding and herbicide application, are often labor-intensive, time-consuming, costly, and may contribute to the development of herbicide-resistant weed populations. In this context, Site-Specific Weed Management (SSWM) has emerged as an effective and sustainable approach, particularly for large-scale automated agricultural systems. A critical component of SSWM is the accurate segmentation of crops and weeds from field imagery. This study implements the Recurrent Residual Attention U-Net (R2AU-Net) model with an EfficientNet-B3 backbone for the task of semantic segmentation of weeds and corn plants using multispectral aerial imagery. The EfficientNet-B3 backbone was chosen for its ability to effectively capture multiscale spatial context. The model was developed using the WeedsGalore dataset, which contains five spectral bands: Red (R), Green (G), Blue (B), Red-Edge (RE), and Near-Infrared (NIR). The dataset comprises 156 annotated images with a spatial resolution of 600 × 600 pixels. The dataset was processed through the stages of resizing, normalization, and augmentation of the training data, which included geometric transformations, color intensity transformations, noise injection, and the application of kernel filters. The proposed model performs six-class semantic segmentation, consisting of the classes background, maize, amaranth, barnyard grass, quickweed, and other weeds. To identify the most effective spectral combination, three experimental configurations were evaluated using three-channel, four-channel, and five-channel input settings. Experimental results demonstrate that the best performance was achieved using a three-channel input configuration composed of Green (G), Near-Infrared (NIR), and Red-Edge (RE) bands. This configuration attained a mean Intersection over Union (mIoU) of 0.5177, a mean Dice coefficient of 0.5924, and a pixel accuracy of 0.9763.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Citra Multispektral, EfficientNet-B3, Recurrent Residual Attention U-Net, Segmentasi Semantik, Multispectral Imaging, EfficientNet-B3, Recurrent Residual Attention U-Net, Semantic Segmentation
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) S Agriculture > S Agriculture (General) T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Irvan Abdul Rahman
Date Deposited:	30 Jul 2026 00:59
Last Modified:	30 Jul 2026 00:59
URI:	http://repository.its.ac.id/id/eprint/139083

Actions (login required)

View Item