The Development of Lateral Flow Immunoassay for Hepatitis B Detection in Saliva

Rice cultivation is fundamental to Indonesia’s agricultural sector, with nitrogen management being crucial for optimal yield. While traditional nitrogen detection methods are time-consuming and costly, computer vision Al offers promising alternatives. However, the real-world reliability of AI-based detection systems remains inadequately validated. This study evaluates the accuracy of a pre-trained EfficientNet_B4 model through field testing and comparison with expert assessments and instrumental measurements. Research was conducted across four rice fields in Yogyakarta, Indonesia, using a structured sampling methodology covering 200 m² per field. Multiple data collection methods were employed: Al image analysis, SPAD-502 chlorophyll meter readings, and agronomist evaluations. Each field was divided into 4 quadrants with random sampling points, generating 144 leaf samples per field. Statistical analysis revealed that the Al model achieved an overall accuracy of 78%, with varying performance across nitrogen level classifications. While the model shows promise for nitrogen deficiency detection, improvements are needed for operational use, as small gaps between lab output and AI can impact farmer decisions on fertilizer purchases, affecting cost and yield. These findings highlight the need for robust calibration methods, standardized image capture protocols, and improved model architecture to enhance real-world applicability in agricultural settings.