Segmentasi Kuantitatif Gejala Penyakit Xanthomonas pada Bibit Eucalyptus pellita Berbasis Deep Learning dengan Dynamic Loss

Abstract

Eucalyptus pellita menjadi salah satu spesies utama pada Hutan Tanaman Industri (HTI) di Indonesia karena pertumbuhannya cepat, adaptasi ekologinya luas, dan bernilai ekonomi tinggi sebagai bahan baku industri pulp dan kertas. Namun, pada fase pembibitan, bibit E. pellita rentan mengalami hawar daun bakteri. Tim Plant Protection mengidentifikasi gejala lapang tersebut sebagai gejala yang konsisten dengan serangan Xanthomonas sp. Gejala ini dapat menurunkan kualitas bibit, mengganggu pertumbuhan, dan menimbulkan kerugian operasional. Pemantauan di area pembibitan sebagian besar masih dilakukan secara manual, sehingga memerlukan waktu, tenaga, dan biaya yang besar. Kondisi ini menegaskan perlunya sistem pemantauan berbasis computer vision dan deep learning yang mampu bekerja pada lingkungan nursery terbuka dengan variasi pencahayaan, latar belakang kompleks, dan tumpang tindih daun. Tujuan penelitian ini adalah mengembangkan model segmentasi daun bibit E. pellita, merancang model segmentasi area nekrosis gejala Xanthomonas pada kondisi ketidakseimbangan piksel ekstrem, serta membangun indikator keparahan berbasis rasio luas nekrosis terhadap luas daun pada tingkat tray. Penelitian mengembangkan alur kerja pemantauan berbasis segmentasi dua tahap yang diikuti kuantifikasi keparahan berbasis rasio piksel. Tahap pertama menggunakan Modified U-Net dengan encoder ResNet-50 pralatih dan regularisasi L2 untuk menghasilkan masker daun sebagai area acuan pengukuran. Tahap kedua menggunakan Attention U-Net berbasis encoder ResNet-50 untuk segmentasi area nekrosis. Kontribusi metodologis penelitian ini adalah penyesuaian Dynamic Areaaware Tversky Loss (DTL) pada alur kerja segmentasi berbasis deep learning untuk mendeteksi lesi nekrosis kecil. DTL dikalibrasi dengan a = 0,7, ß = 0,3, ? = 602, t = 0,00083, dan e = 1 × 10?7; rasio r dihitung secara dinamis dari masker acuan setiap citra. Tahap ketiga membentuk indikator keparahan turunan pada tingkat tray. Citra diperoleh di pembibitan terbuka PT XYZ, Riau, menggunakan sistem akuisisi berbasis Internet of Things dengan kamera RGB 1080p pada boom sprayer yang terhubung ke mini PC Intel NUC sebagai unit komputasi tepi. Hasil segmentasi daun menunjukkan bahwa Modified U-Net memberikan kinerja terbaik dan paling stabil pada seluruh kondisi pencahayaan. Median Dice mencapai 0,872 pada kondisi mendung, 0,841 pada kondisi cerah, dan 0,854 pada kondisi terik. Median IoU masing-masing adalah 0,773, 0,725, dan 0,745. Uji Kruskal–Wallis pada metrik per citra menunjukkan bahwa perbedaan kinerja antarkondisi pencahayaan tidak signifikan untuk Modified U-Net (H = 4,012; p = 0,1345; e² = 0,008), yang menunjukkan bahwa model cukup stabil terhadap variasi iluminasi alami. Dibandingkan SegNet dan DeepLabv3+, Modified U-Net juga menunjukkan keseimbangan yang lebih baik antara Precision dan Recall, sehinggaii masker daun dapat digunakan sebagai dasar normalisasi area pada tahap estimasi keparahan. Pada segmentasi nekrosis, konfigurasi terbaik diperoleh pada model A2, yaitu Attention U-Net dengan DTL. Pada data uji, model ini mencapai median Dice 0,812, IoU 0,680, Precision 0,788, dan Recall 0,838. Kinerja tersebut lebih baik daripada baseline A1 yang menggunakan Tversky loss standar dengan median Dice 0,714 dan IoU 0,548. Uji Wilcoxon signed-rank berpasangan dengan koreksi Holm menunjukkan bahwa peningkatan A1 ke A2 signifikan secara statistik dengan ? median Dice sebesar 0,098, p < 0,0001, dan ukuran efek besar r = 0,904. Hasil ini menunjukkan bahwa pada konfigurasi model dan dataset penelitian ini, DTL efektif menurunkan negatif palsu pada lesi mikro tanpa meningkatkan positif palsu secara berlebihan. Peningkatan juga terlihat pada pasangan model tanpa attention; pada data uji penelitian ini, selisih peningkatan terbesar diperoleh pada konfigurasi yang menggabungkan DTL dan attention gate. Integrasi hasil segmentasi daun dan segmentasi nekrosis menghasilkan indikator keparahan berbasis rasio piksel yang konsisten secara komputasional dan dapat diaudit sebagai indikator turunan berbasis citra. Pada tingkat tray, keparahan dihitung sebagai rasio luas nekrosis terhadap luas daun. Untuk kebutuhan operasional, kamera dipicu otomatis melalui skrip Python sesuai interval akuisisi. Citra disimpan secara lokal. Setelah satu lintasan selesai, mini PC menjalankan inferensi untuk menghitung severity_tray pada tray sampel dan mengirimkan hasilnya melalui koneksi internet ke dasbor, sehingga alur akuisisi–inferensi– pelaporan berlangsung tanpa pemindahan berkas manual. Dengan kecepatan boom sprayer terukur 0,1308 m s?¹, interval pemicu akuisisi yang sesuai adalah 7,83 detik, dengan estimasi waktu operasional sekitar 55 menit per jalur. Hasil ini menunjukkan bahwa sistem yang dikembangkan menyediakan kerangka pemantauan operasional berbasis edge-to-cloud yang terotomasi, konsisten, dan terukur di lingkungan HTI. Validasi berpasangan dilakukan pada tray yang sama di tiga baris sampling dalam satu jalur dan menghasilkan 238 pasangan valid. Estimasi sistem menunjukkan kesesuaian yang kuat dengan penilaian pengawas, dengan koefisien korelasi intrakelas AI–manual 0,79, Spearman ? 0,84, galat absolut rerata 0,25 poin persentase, kesesuaian kategori tepat 71,8%, dan kesesuaian dalam toleransi satu tingkat 100%.

Eucalyptus pellita is one of the main species in industrial plantation forests (HTI) in Indonesia because of its rapid growth, broad ecological adaptation, and high economic value as raw material for the pulp and paper industry. However, during the nursery stage, E. pellita seedlings are susceptible to bacterial leaf blight. The Plant Protection team identified the observed field symptoms as being consistent with Xanthomonas sp. infection. These symptoms may reduce seedling quality, disrupt growth, and cause operational losses. Monitoring in nursery areas is still largely conducted manually, requiring substantial time, labor, and cost. This condition emphasizes the need for a computer vision- and deep learning-based monitoring system capable of operating in open nursery environments with varying illumination, complex backgrounds, and overlapping leaves. The objectives of this study were to develop a leaf segmentation model for E. pellita seedlings, design a necrotic area segmentation model for Xanthomonas symptoms under extreme pixel imbalance, and construct a severity indicator based on the ratio of necrotic area to leaf area at the tray level. To achieve these objectives, this study developed a monitoring pipeline based on two-stage segmentation followed by pixel-ratio-based severity quantification. The first stage included leaf segmentation using Modified U-Net with a pretrained ResNet-50 encoder and L2 regularization to generate leaf masks as the reference area for measurement. The second stage included necrotic area segmentation using Attention U-Net with a ResNet-50 encoder. The methodological contribution of this study was the adaptation of Dynamic Area-aware Tversky Loss in a deep learning-based segmentation pipeline to detect small necrotic lesions of Xanthomonas symptoms in E. pellita seedlings in an open nursery environment. In this study, DTL was used and calibrated to strengthen learning on images with a very small necrotic ratio so that the model became more sensitive to microlesions. DTL used a = 0.7 to weight false negatives, ß = 0.3 to weight false positives, ? = 602 as the linear amplification coefficient, t = 0.00083 as the small-lesion activation threshold, and e = 1 × 10?7 as the numerical stabilizing constant. The ratio r was dynamically calculated for each image based on the ground truth mask. The third stage included the development of a derived tray-level severity indicator for operational monitoring needs. All images were acquired in the open nursery area of PT XYZ, Riau, using an Internet of Things-based acquisition system with a 1080p RGB camera mounted on a boom sprayer and directly connected to an Intel NUC mini PC as the edge computing unit. The leaf segmentation results showed that Modified U-Net produced the best and most stable performance across all illumination conditions. The median Dice reached 0.872 under cloudy conditions, 0.841 under sunny conditions, and 0.854 under high-intensity sunlight conditions. The corresponding median IoU values were 0.773, 0.725, and 0.745. The Kruskal–Wallis test on per-image metricsiv showed that performance differences among illumination conditions were not significant for Modified U-Net (H = 4.012; p = 0.1345; e² = 0.008), indicating that the model was sufficiently stable against natural illumination variation. Compared with SegNet and DeepLabv3+, Modified U-Net also showed a better balance between Precision and Recall, so the leaf mask could be used as the basis for area normalization in the severity estimation stage. For necrosis segmentation, the best configuration was obtained by model A2, namely Attention U-Net with DTL. On the test data, this model achieved a median Dice of 0.812, IoU of 0.680, Precision of 0.788, and Recall of 0.838. This performance was better than the A1 baseline using standard Tversky loss, which achieved a median Dice of 0.714 and IoU of 0.548. The paired Wilcoxon signedrank test with Holm correction showed that the improvement from A1 to A2 was statistically significant, with a median Dice increase of 0.098, p < 0.0001, and a large effect size of r = 0.904. These results indicate that, under the model configuration and dataset used in this study, DTL effectively reduced false negatives in microlesions without excessively increasing false positives. The improvement effect was also observed in the pair without attention, but the largest improvement occurred when DTL was combined with the attention gate. The integration of leaf segmentation and necrosis segmentation produced a pixel-ratio-based severity indicator that was computationally consistent and auditable as an image-derived indicator. At the tray level, severity was calculated as the ratio of necrotic area to leaf area. For operational monitoring, a Python script automatically triggered the camera at a predefined acquisition interval, and the images were stored locally. After each pass, the mini PC performed inference and calculated severity_tray for each sampled tray as the ratio of necrotic pixels to leaf pixels. The values were then sent through an internet connection to the dashboard so that the acquisition–inference–reporting workflow proceeded without manual file transfer. With a measured boom speed of 0.1308 m s?¹, the appropriate acquisition trigger interval was 7.83 seconds, with an estimated operational time of approximately 55 minutes per line. These results show that the developed system provides an automated, consistent, and measurable edge-to-cloud operational monitoring framework in the HTI environment. Paired validation on the same trays, conducted across three sampling rows along one lane, comprising 238 valid pairs, showed strong agreement between system estimates and supervisor assessments, both in percentage scores and operational categories, with AI–manual intraclass correlation coefficient = 0.79, Spearman ? = 0.84, mean absolute error = 0.25 percentage points, exact category agreement = 71.8%, and agreement within onelevel tolerance = 100%