Implementasi arsitektur Densenet-Involution dengan Grad-CAM++ untuk deteksi defisiensi nutrisi pada tanaman padi

Sumaryono, Mohammad Rafli (2026) Implementasi arsitektur Densenet-Involution dengan Grad-CAM++ untuk deteksi defisiensi nutrisi pada tanaman padi. Sarjana thesis, UIN Sunan Gunung Djati Bandung.

[img]
Preview
Text (COVER)
1_cover.pdf

Download (63kB) | Preview
[img]
Preview
Text (ABSTRAK)
2_abstrak.pdf

Download (29kB) | Preview
[img]
Preview
Text (LEMBAR PERNYATAAN)
Surat pernyataan Karya sendiri (11).pdf

Download (794kB) | Preview
[img]
Preview
Text (DAFTAR ISI)
3_daftarisi.pdf

Download (48kB) | Preview
[img]
Preview
Text (BAB I)
4_bab1.pdf

Download (83kB) | Preview
[img] Text (BAB II)
5_bab2.pdf
Restricted to Registered users only

Download (446kB) | Request a copy
[img] Text (BAB III)
6_bab3.pdf
Restricted to Registered users only

Download (211kB) | Request a copy
[img] Text (BAB IV)
7_bab4.pdf
Restricted to Registered users only

Download (1MB) | Request a copy
[img] Text (BAB V)
8_bab5.pdf
Restricted to Registered users only

Download (89kB) | Request a copy
[img] Text (DAFTAR PUSTAKA)
9_daftarpustaka.pdf
Restricted to Registered users only

Download (104kB) | Request a copy

Abstract

Defisiensi unsur hara makro Nitrogen (N), Fosfor (P), dan Kalium (K) pada tanaman padi (Oryza sativa L.) dapat menurunkan hasil panen hingga 30%, sementara proses deteksi masih bersifat manual, subjektif, dan tidak konsisten. Penelitian ini mengusulkan model klasifikasi berbasis arsitektur Involution-Infused DenseNet-121 yang dipadukan dengan Explainable AI menggunakan Grad-CAM++ untuk mengidentifikasi empat kondisi nutrisi daun padi: Healthy, defisiensi Nitrogen, Fosfor, dan Kalium. Model memanfaatkan backbone DenseNet-121 pretrained ImageNet yang diintegrasikan dengan dua lapisan Involution melalui residual connection, menghasilkan 9,2 juta parameter dengan overhead komputasi rendah. Set data yang digunakan berjumlah 1.530 citra dari dua sumber publik. Eksperimen dilakukan melalui tiga skenario preprocessing untuk mengevaluasi pengaruh background terhadap performa dan interpretabilitas model. Hasil menunjukkan bahwa skenario S3 Boxcrop memberikan performa terbaik dengan akurasi 97,83%, F1-Weighted 0,9785 dan AUC-ROC 0,9992. Analisis FG Activation Score mengungkap adanya shortcut learning pada skenario baseline yang tidak terdeteksi oleh metrik akurasi, sehingga menegaskan pentingnya evaluasi interpretabilitas secara kuantitatif. Visualisasi Grad-CAM++ menunjukkan bahwa model berfokus pada area biologis yang relevan, seperti klorosis pada defisiensi Nitrogen, pigmentasi pada Fosfor, dan nekrosis tepi daun pada Kalium. Model kemudian diimplementasikan dalam aplikasi mobile berbasis Flutter dengan backend FastAPI. Validasi domain dilakukan melalui konsultasi dengan UPTD Pertanian, mencakup verifikasi pelabelan dataset, konfirmasi gejala agronomis, serta uji coba pada data nyata yang menunjukkan hasil positif. Penelitian ini berkontribusi pada pengembangan sistem deteksi nutrisi padi yang akurat, dan aplikatif untuk pertanian presisi. Macronutrient deficiencies of Nitrogen (N), Phosphorus (P), and Potassium (K) in rice (Oryza sativa L.) can reduce crop yields by up to 30%, while detection methods remain manual, subjective, and inconsistent. This study proposes an Involution-Infused DenseNet-121 architecture integrated with Grad-CAM++ Explainable AI to classify four rice leaf nutrient conditions: Healthy, Nitrogen deficiency, Phosphorus deficiency, and Potassium deficiency. The model utilizes an ImageNet-pretrained DenseNet-121 backbone enhanced with two Involution layers through residual connections, resulting in approximately 9.2 million trainable parameters with minimal computational overhead. A total of 1,530 images were collected from two public datasets. Experiments were conducted under three preprocessing scenarios to assess the impact of background handling on model performance and interpretability. Results show that the S3 Boxcrop scenario achieved the best performance, with 97.83% accuracy, 0.9785 F1-Weighted and 0.9992 AUC-ROC. FG Activation Score analysis revealed the presence of shortcut learning in the baseline scenario that was not captured by accuracy metrics alone, highlighting the importance of quantitative interpretability evaluation. Grad-CAM++ visualizations consistently focused on biologically relevant regions, including chlorosis for Nitrogen deficiency, pigmentation changes for Phosphorus, and leaf edge necrosis for Potassium. The best-performing model was deployed in a Flutter-based mobile application with a FastAPI backend. Domain validation was conducted through consultation with agricultural experts, including dataset label verification, confirmation of agronomic symptoms, and testing on real-world data, which produced positive feedback. This study contributes an accurate, and practical nutrient deficiency detection system to support precision agriculture in Indonesia.

Item Type: Thesis (Sarjana)
Uncontrolled Keywords: DenseNet-121; Involution; Grad-CAM++; defisiensi nutrisi padi; explainable AI; klasifikasi citra daun
Subjects: Special Computer Methods > Artificial Intelligence
Special Computer Methods > Computer Vision
Divisions: Fakultas Sains dan Teknologi > Program Studi Teknik Informatika
Depositing User: Mohammad Rafli Sumaryono
Date Deposited: 12 May 2026 06:57
Last Modified: 12 May 2026 06:57
URI: https://digilib.uinsgd.ac.id/id/eprint/131297

Actions (login required)

View Item View Item