Model Prediksi Tingkat Keparahan Penyakit Daun Pestalotiopsis sp. pada Tanaman Karet Berbasis Indeks Vegetasi dan Deep Learning

Abstract

Penyakit gugur daun yang berkaitan dengan Pestalotiopsis sp. merupakan kendala penting pada perkebunan karet karena menurunkan kualitas tajuk, mengganggu proses fotosintesis, dan berpotensi mengurangi produktivitas lateks. Penelitian ini bertujuan mengembangkan model klasifikasi empat tingkat keparahan penyakit daun Pestalotiopsis sp. pada tanaman karet berbasis indeks vegetasi multivariat (Multivariate Vegetation Indices, MVIs) yang diekstraksi dari citra multispektral pesawat udara nirawak (Unmanned Aerial Vehicle, UAV). Data dikumpulkan di Pusat Penelitian Karet Sembawa, Sumatera Selatan, pada klon BPM 24, GT 1, dan RRIC 100. Dataset penelitian terdiri atas 144 titik GT berlabel dan 1760 titik kandidat tidak berlabel. Tahapan penelitian mencakup prapemrosesan citra, penilaian tingkat keparahan penyakit, perhitungan indeks vegetasi, seleksi fitur menggunakan Random Forest (RF), penanganan ketidakseimbangan kelas dengan Synthetic Minority Oversampling Technique (SMOTE), pembangunan model Convolutional Neural Network satu dimensi (CNN 1-D), validasi silang lima lipatan, serta prediksi awal pada data tidak berlabel. Seleksi fitur menghasilkan lima variabel paling informatif, yaitu NDRE, LCI, CI, NDVI_NDRE_Interaction, dan GCI_Ratio. Kelima fitur tersebut merepresentasikan respons red-edge, kandungan klorofil, tingkat kehijauan tajuk, serta interaksi spektral yang berkaitan dengan perubahan fisiologis tanaman akibat serangan penyakit. Model CNN 1-D memperoleh rata-rata akurasi validasi silang sebesar 82,22% dan rata-rata F1-score sebesar 0,8218. Penerapan model pada data tidak berlabel menghasilkan 105 prediksi berkepercayaan tinggi. Verifikasi lapangan terbatas menunjukkan bahwa 10 dari 16 sampel yang diperiksa sesuai dengan kondisi aktual. Hasil ini menunjukkan bahwa kombinasi fitur spektral terpilih, penanganan ketidakseimbangan kelas, dan CNN 1-D dapat mendukung klasifikasi tingkat keparahan penyakit secara objektif, terukur, dan berbasis tingkat keyakinan. Penelitian ini memberikan kontribusi representasional melalui penggunaan MVIs terpilih yang ringkas dan relevan secara fisiologis, kontribusi metodologis melalui integrasi RF, SMOTE, dan CNN 1-D, serta kontribusi praktis bagi pengembangan pemantauan kesehatan tanaman karet berbasis UAV. Meskipun demikian, generalisasi model masih terbatas oleh lokasi, waktu pengamatan, klon, sensor, dan jumlah data berlabel. Oleh karena itu, penerapan operasional yang lebih luas memerlukan validasi lintas lokasi, musim, klon, sensor, dan dataset. Validasi tersebut juga diperlukan untuk menilai kestabilan performa, kalibrasi probabilitas, sensitivitas terhadap variasi lingkungan, serta kemampuan model dalam mendukung prioritas inspeksi dan pengendalian penyakit pada skala perkebunan secara konsisten dalam praktik lapangan.

Leaf fall disease associated with Pestalotiopsis sp. is an important constraint in rubber plantations because it reduces canopy quality, disrupts photosynthesis, and may decrease latex productivity. This study aimed to develop a four-level classification model for the severity of Pestalotiopsis sp. leaf disease in rubber plants using multivariate vegetation indices derived from multispectral Unmanned Aerial Vehicle imagery. Data were collected at the Sembawa Rubber Research Centre, South Sumatra, from the BPM 24, GT 1, and RRIC 100 clones. The dataset comprised 144 labelled ground-truth points and 1760 unlabelled candidate points. The research workflow included image preprocessing, disease-severity assessment, vegetation-index calculation, feature selection using Random Forest, class-imbalance handling using the Synthetic Minority Oversampling Technique, development of a one-dimensional Convolutional Neural Network, five-fold cross-validation, and preliminary prediction on unlabelled data. Feature selection identified five variables as the most informative: NDRE, LCI, CI, NDVI_NDRE_Interaction, and GCI_Ratio. These variables represented red-edge response, chlorophyll content, canopy greenness, and spectral interactions associated with physiological changes caused by disease infection. The CNN 1-D model achieved a mean cross-validation accuracy of 82.22% and a mean F1-score of 0.8218. Application of the model to the unlabelled data produced 105 high-confidence predictions. Limited field verification showed that 10 of the 16 examined samples agreed with actual field conditions. These results indicate that the combination of selected spectral features, class-imbalance handling, and CNN 1-D can support objective, measurable, and confidence-based classification of disease severity. This study contributes representationally through the use of a compact set of physiologically relevant multivariate vegetation indices, methodologically through the integration of Random Forest, SMOTE, and CNN 1-D, and practically through the development of UAV-based rubber-plant health monitoring. Nevertheless, model generalisation remains limited by the study location, observation periods, clones, sensor, and the number of labelled samples. Therefore, broader operational implementation requires validation across locations, seasons, clones, sensors, and datasets. Such validation is also needed to evaluate performance stability, probability calibration, sensitivity to environmental variability, and the model's ability to support consistent prioritisation of field inspection and disease management at plantation scale under diverse operational conditions before routine deployment in large commercial rubber plantations.