Utilization of AI for Digital Image Based Fruit Quality and Weight Classification

Advances in artificial intelligence (AI) have revolutionized modern agriculture, particularly in the process of digital image-based fruit quality classification and weight estimation. Market demand for fruit with uniform quality, no defects, and consistent weight drives the need for faster and more accurate non-destructive inspection systems. Through computer vision and deep learning technologies, AI can analyze visual characteristics of fruit such as color, shape, texture, and size without damaging the sample, while also precisely estimating fruit weight. Various types of image sensors have been used, ranging from low-cost RGB cameras to multispectral and hyperspectral cameras capable of detecting internal attributes such as sugar and water content. Studies by Chuquimarca (2024) and Kalantar et al. (2020) show that the combination of image data and deep learning models significantly improves the accuracy of external quality classification and fruit weight estimation. In technical applications, architectures such as Convolutional Neural Networks (CNN), EfficientNet, DenseNet, and the YOLO (You Only Look Once) model are widely used for automatic fruit detection and classification. The latest versions, such as YOLOv11, excel due to their high speed and accuracy in detecting fruit in complex environments. After fruit objects are detected, an image segmentation process is performed to calculate the dimensions and textural features used in predictive regression of fruit weight.

Weight estimation approaches are carried out using two main methods: direct weight estimation using a CNN that predicts weight from the image, and indirect weight estimation using geometric analysis such as diameter or surface area. Research using AppleV (Barda et al., 2024) and the FruitBox dataset shows that a top-view projection area-based approach can achieve errors below 10% after calibration with actual mass. AI systems for fruit quality and weight classification generally go through five main stages: image acquisition, detection, segmentation, feature extraction, classification or regression, and evaluation of the results. These stages involve preprocessing such as normalization, data augmentation, and lighting correction to improve the model's robustness to environmental variations. Model performance evaluation typically uses metrics such as accuracy, F1-score, mean Average Precision (mAP), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE). Recent studies report that image-based AI technology can achieve accuracy above 95% on controlled datasets. However, its performance declines in field conditions due to unstable lighting and complex backgrounds. Researchers such as Rojas (2025) and Knott (2023) emphasize the importance of domain adaptation and real-world data augmentation to maintain model adaptability. The use of hyperspectral imagery also improves internal quality detection, although the cost remains high.

Key challenges in implementing AI in precision agriculture include lighting variations, fruit overlap, and the limited availability of annotated datasets with actual weights. Models trained on one variety often do not readily apply to other varieties without transfer learning. Suggested solutions include combining multi-view imagery and additional spectral data, such as near-infrared (NIR), to improve weight estimation accuracy. Overall, the application of AI in digital image-based fruit quality classification and weight estimation has a significant impact on the efficiency of the horticultural supply chain. This technology accelerates the grading process, reduces human subjectivity, and opens up opportunities for application in MSMEs and the agricultural industry through the integration of low-cost sensors and lightweight edge computing-based models. Future research focuses on multimodal models that combine visual, spectral, and textural data to produce more reliable predictions in real-world conditions.

Pemanfaatan AI untuk Klasifikasi Mutu dan Bobot Buah Berbasis Citra Digital

Kemajuan kecerdasan buatan (AI) telah merevolusi dunia pertanian modern, khususnya dalam proses klasifikasi mutu dan estimasi bobot buah berbasis citra digital. Permintaan pasar terhadap buah dengan kualitas seragam, bebas cacat, dan bobot konsisten mendorong kebutuhan akan sistem inspeksi non-destruktif yang lebih cepat dan akurat. Melalui teknologi computer vision dan deep learning, AI mampu menganalisis karakteristik visual buah seperti warna, bentuk, tekstur, dan ukuran tanpa merusak sampel, sekaligus memperkirakan bobot buah secara presisi. Berbagai jenis sensor citra telah digunakan, mulai dari kamera RGB berbiaya rendah hingga kamera multispektral dan hiperspektral yang mampu mendeteksi atribut internal seperti kadar gula dan kadar air. Studi oleh Chuquimarca (2024) dan Kalantar et al. (2020) menunjukkan bahwa kombinasi data citra dan model deep learning meningkatkan akurasi klasifikasi mutu eksternal serta estimasi bobot buah secara signifikan. Dalam penerapan teknisnya, arsitektur seperti Convolutional Neural Networks (CNN), EfficientNet, DenseNet, hingga model YOLO (You Only Look Once) banyak digunakan untuk deteksi dan klasifikasi buah otomatis. Versi terbaru seperti YOLOv11 unggul karena kecepatan dan akurasi tinggi dalam mendeteksi buah di lingkungan kompleks. Setelah objek buah terdeteksi, proses segmentasi citra dilakukan untuk menghitung dimensi dan fitur tekstural yang digunakan dalam regresi prediktif bobot buah.

Pendekatan estimasi bobot dilakukan melalui dua metode utama yaitu langsung dengan CNN yang memprediksi bobot dari citra, dan tidak langsung melalui analisis geometris seperti diameter atau luas permukaan. Penelitian AppleV (Barda et al., 2024) dan dataset FruitBox menunjukkan bahwa pendekatan berbasis area proyeksi top-view dapat mencapai error di bawah 10% setelah kalibrasi dengan massa aktual. Sistem AI untuk klasifikasi mutu dan bobot buah umumnya melalui lima tahap utama yaitu akuisisi citra, deteksi segmentasi, ekstraksi fitur, klasifikasi atau regresi, dan evaluasi hasil. Tahapan ini melibatkan prapemrosesan seperti normalisasi, augmentasi data, serta koreksi pencahayaan untuk meningkatkan ketahanan model terhadap variasi lingkungan. Evaluasi kinerja model biasanya menggunakan metrik seperti akurasi, F1-score, mean Average Precision (mAP), Mean Absolute Error (MAE), dan Root Mean Square Error (RMSE). Berbagai studi terkini melaporkan bahwa teknologi AI berbasis citra dapat mencapai akurasi di atas 95% pada dataset terkontrol. Namun, performanya menurun di kondisi lapangan akibat pencahayaan yang tidak stabil dan latar belakang kompleks. Peneliti seperti Rojas (2025) dan Knott (2023) menekankan pentingnya domain adaptation dan data augmentation berbasis kondisi nyata agar model tetap adaptif. Penggunaan citra hiperspektral juga meningkatkan deteksi mutu internal, meski biayanya masih tinggi.

Tantangan utama dalam penerapan AI di sektor pertanian presisi meliputi variasi pencahayaan, tumpang tindih buah, serta keterbatasan dataset beranotasi bobot aktual. Model yang dilatih pada satu varietas sering kali tidak langsung cocok untuk varietas lain tanpa transfer learning. Solusi yang disarankan mencakup penggabungan citra multi-view dan data spektral tambahan seperti Near-Infrared (NIR) untuk memperkuat akurasi estimasi bobot. Secara keseluruhan, penerapan AI dalam klasifikasi mutu dan estimasi bobot buah berbasis citra digital membawa dampak besar terhadap efisiensi rantai pasok hortikultura. Teknologi ini mempercepat proses grading, mengurangi subjektivitas manusia, serta membuka peluang penerapan pada UMKM dan industri pertanian melalui integrasi sensor murah dan model ringan berbasis edge computing. Arah penelitian ke depan difokuskan pada model multimodal yang menggabungkan data visual, spektral, dan tekstural untuk menghasilkan prediksi yang lebih andal di kondisi nyata.

Reference

• Barda, S., Rahimi, A., & Patel, D. (2024). AppleV: A dataset for apple fruit volume estimation. Proceedings of the ACM International Conference on Multimedia.

• Chuquimarca, L. (2024). A review of external quality inspection for fruit grading using CNN and computer vision. Computers in Agriculture, 3(2), 45–62.

• Kalantar, A., Mahmoudi, S., & Esmaeili, M. (2020). A deep learning system for single and overall weight estimation of melons from top-view images. Computers and Electronics in Agriculture, 178, 105750. 

• Knott, M. (2023). Facilitated machine learning for image-based fruit quality. Journal of Food Engineering, 354, 111558.

• Rojas Santelices, I. (2025). Artificial vision systems for fruit inspection and classification. Sensors, 25(5), 1524.

PT. Precision Agriculutre Indonesia adalah ekosistem digital pertanian Indonesia yang mengintegrasikan agrotech, pertanian presisi, pertanian cerdas, dan pertanian pintar melalui pemanfaatan teknologi seperti sensor pertanian, Internet of Things (IoT), kecerdasan buatan, sistem irigasi otomatis, pemupukan cerdas, dan pemantauan tanaman berbasis data real-time, serta menghadirkan layanan edukasi petani modern, digitalisasi agribisnis, pasar produk pertanian online, penguatan rantai pasok, inovasi teknologi tepat guna, dan solusi pertanian ramah lingkungan yang mendukung pertanian modern, berkelanjutan, dan berdaya saing tinggi di era Revolusi Industri 4.0.  Pertanian Presisi Indonesia