Pengembangan Aplikasi Web untuk Pendugaan Kadar Nitrogen Relatif Daun Dewa (Gynura segetum) Menggunakan Machine Learning Berbasis Citra Digital

Abstract

Pendugaan kadar nitrogen relatif tanaman daun dewa (Gynura segetum) umumnya dilakukan menggunakan SPAD secara langsung sehingga kurang efisien dan berpotensi merusak jaringan daun. Penelitian ini bertujuan mengembangkan aplikasi web pendugaan nitrogen relatif berbasis citra digital menggunakan algoritma machine learning, membandingkan performa model, dan menganalisis pengaruh cekaman ultraviolet terhadap nitrogen relatif tanaman. Sebanyak 336 sampel citra daun dikumpulkan dari tanaman yang diberi perlakuan cekaman ultraviolet dalam plant factory. Tujuh fitur citra diekstraksi melalui segmentasi berbasis ruang warna HSV, kemudian digunakan untuk melatih model ANN, Random Forest, dan SVR yang dievaluasi menggunakan K-Fold Cross Validation dengan k=5. SVR menghasilkan performa terbaik dengan R²=0,8274, MAE=0,6775, dan MAPE=4,10%, dan diimplementasikan dalam aplikasi web berbasis Streamlit yang dilengkapi kategorisasi nitrogen relatif ke dalam tiga kelas. Seluruh fungsi aplikasi berjalan sesuai rancangan berdasarkan pengujian black box testing. Analisis terhadap 504 sampel menunjukkan cekaman ultraviolet memberikan pengaruh biphasic terhadap nitrogen relatif, dengan rata-rata kelompok perlakuan (16,71%) lebih tinggi dibandingkan kontrol (15,76%). Aplikasi yang dikembangkan dapat digunakan sebagai sarana pemantauan kondisi fisiologis tanaman daun dewa secara non-destruktif berbasis citra digital.

Relative nitrogen content in daun dewa (Gynura segetum) is commonly measured using SPAD directly, which is inefficient and potentially damaging to leaf tissue. This study aimed to develop a digital image-based web application for non-destructive relative nitrogen estimation using machine learning algorithms, compare model performance, and analyze the effect of ultraviolet stress on relative nitrogen content. A total of 336 leaf image samples were collected from plants subjected to ultraviolet stress in a plant factory. Seven image features were extracted through HSV color space-based segmentation and used to train ANN, Random Forest, and SVR models evaluated using K-Fold Cross Validation with k=5. SVR achieved the best performance with R²=0.8274, MAE=0.6775, and MAPE=4.10%, and was implemented in a Streamlit-based web application with relative nitrogen categorization into three classes. All application functions performed as designed based on black box testing. Analysis of 504 samples revealed that ultraviolet stress exerted a biphasic effect on relative nitrogen content, with the treatment group average (16.71%) exceeding the control group (15.76%). The developed application provides a non-destructive, digital image-based tool for monitoring the physiological condition of daun dewa plants.