Optimizing Bit Depth for Longan Seedling Identification Using ANN and GLCM

Accurate identification of longan (Dimocarpus longan) seedling varieties is essential for agribusiness to select cultivars meeting market and environmental needs, but manual identification is error-prone due to similar leaf textures. This study optimizes grayscale image bit depth using Artificial Neural Networks (ANN) and Gray Level Co-occurrence Matrix (GLCM) to enhance longan seedling classification accuracy, addressing a gap in texture-based identification efficiency. Leaf images from five longan varieties (Itoh, Pingpong, Merah, Matalada, Diamond River) were captured with a USB digital microscope and converted to grayscale at bit depths of 4 (0–15), 5 (0–31), 6 (0–63), 7 (0–127), and 8 (0–255). Texture features (contrast, correlation, energy, homogeneity, entropy, standard deviation) were extracted using MATLAB. An ANN model, trained with the traingdx algorithm on 800 training and 200 test images, classified the varieties. The 6-bit and 4-bit depths yielded the highest accuracy (84.5%), followed by 7-bit (84.0%), 5-bit (83.5%), and 8-bit (82.0%), with Matalada achieving 90.0% accuracy. The 8-bit depth introduced texture noise, reducing performance. A 6-bit depth is optimal for longan leaf texture classification, though distinguishing similar varieties like Itoh and Pingpong remains challenging. Future research should incorporate color or morphological features to improve agricultural image processing.