<![CDATA[Bird pests significantly threaten rice production in Indonesia, particularly during the generative stage, yet most existing studies focus on general bird detection without distinguishing pest species. This study developed a species-aware classification model to identify five classes (four granivorous pest species and one non-pest class) in rice-field environments. A lightweight convolutional neural network based on EfficientNet-B0 with transfer learning was trained on 2,999 granivorous and 635 non-pest images, which were filtered through duplicate removal, quality screening, and class balancing to produce a curated dataset of 2,112 images. The dataset was split into 70% training and 30% validation sets, and five-fold cross-validation was conducted within the training subset during model development to assess robustness. The model achieved a validation accuracy of 83.38%, with an average cross-validation accuracy of 84.7% ± 1.52% and a macro-average AUC of 0.9724, indicating strong class discrimination. Most misclassifications occurred among morphologically similar Lonchura species rather than random class confusion. An additional 18 in situ field images were used for preliminary external evaluation under natural field conditions, where prediction confidence decreased due to environmental complexity. Overall, the results demonstrate that a lightweight EfficientNet-B0 backbone can effectively distinguish visually similar granivorous pest species from non-pest birds while maintaining computational efficiency suitable for precision agriculture.]]>
