Emphasizing Data Quality for the Identification of Chili Varieties in the Context of Smart Agriculture

Aim/Purpose Background Methodology Contribution Findings Recommendations for Practitioners This research aims to evaluate models from meta-learning techniques, such as Riemannian Model Agnostic Meta-Learning (RMAML), Model-Agnostic Meta Learning (MAML), and Reptile meta-learning, to obtain high-quality metadata. The goal is to utilize this metadata to increase accuracy and efficiency in identi fying chili varieties in smart agriculture. The identification of chili varieties in smart agriculture is a complex process that requires a multi-faceted approach. One challenge in chili variety identification is the lack of a large and diverse dataset. This can be addressed using meta-learn ing techniques, which allow the model to leverage knowledge learned from other related tasks or artificially expand the dataset by applying transformations to existing data. Another challenge is the variation in growing conditions, which can affect the appearance of chili varieties. Meta-learning techniques can help address this challenge by allowing the model to adapt to variations in growing conditions with task-specific embeddings and optimizations. With the help of meta-learning techniques, such as data augmentation, data characterization, selection of datasets, and performance estimation, quality metadata for accurate identification of chili varieties can be achieved even in the presence of limited data and variations in growing conditions. Furthermore, the use of meta-learning techniques in chili variety identification can also assist in addressing challenges related to the computational complexity of the task. The research approach employed is quantitative, specifically comparing three models from meta-learning techniques to determine which model is most suita ble for our dataset. Data was collected from the variety assembly garden in the form of images of chili leaves using a mobile device. The research successfully gathered 1,974 images of chili leaves, with 697 images of large red chilies, 649 images of curly red chilies, and 628 images of cayenne peppers. These chili leaf images were then processed using augmentation techniques. The results of im age data augmentation were categorized based on leaf characteristics (such as oval, lancet, elliptical, serrated leaf edges, and flat leaf edges). Subsequently, training and validation utilized three models from meta-learning techniques. The final stage involved model evaluation using 2-way and 3-way classification, as well as 5-shot and 10-shot learning scenarios to select the dataset with the best performance. Improving classification accuracy, with a focus on ensuring high-quality data, al lows for more precise identification and classification of chili varieties. Enhanc ing model training through an emphasis on data quality ensures that the models receive reliable and representative input, leading to improved generalization and performance in identifying chili varieties. With small collections of datasets, the authors have used data augmentation and meta-learning techniques to overcome the challenges of limited data and varia tions in growing conditions. By leveraging the knowledge and adaptability gained from meta-learning, accu rate identification of chili varieties can be achieved even with limited data and variations in growing conditions. The use of meta-learning techniques in chili variety identification can greatly improve the accuracy and reliability of the iden tification process. 2 Suwarningsih, Kirana, Khotimah, Rozie, Riswantini, Nugraheni, Munandar, Arisal, & Ahmadi Recommendations for Researchers Impact on Society Future Research Keywords Using meta-learning techniques, such as transfer learning and parameter optimi zation, researchers can overcome challenges related to limited data and varia tions in growing conditions in chili variety identification. The findings from this research can help identify superior chili seeds, thereby motivating farmers to cultivate high-quality chilies and achieve bountiful har vests. We intend to verify our approach on a more extensive array of datasets and ex plore the implementation of more resilient regularization techniques, going be yond image augmentation, within the meta-learning techniques. Furthermore, our goal is to expand our research to encompass the automatic learning of pa rameters during training and tackle issues associated with noisy labels. Building on the insights gained from our observed outcomes, a future objective is to en hance the refinement of model-agnostic meta-learning techniques that can ef fectively adapt to intricate task distributions with substantial domain gaps be tween tasks. To realize this aim, our proposal involves devising model-agnostic meta-learning techniques specifically designed for multi-modal scenarios.