FAWINSTARNet: A Lightweight MobileNetV2 Model for Early Instar Fall Armyworm Detection in Maize

The Fall Armyworm (Spodoptera frugiperda) has emerged as a major constraint on maize cultivation throughout warm-climate agricultural zones.Management practices are most effective during the earliest larval stages, making precise recognition of first-and second-instar caterpillars essential for minimizing crop damage and limiting indiscriminate pesticide application.In response to this requirement, the present work proposes FAWINSTARNet, a computationally efficient deep-learning framework derived from the MobileNetV2 family and tailored for six-category instar discrimination.An initial image repository containing 12,169 samples validated by entomological experts was systematically enlarged to 187,152 images through controlled augmentation to enhance feature variability.A group of ten pretrained convolutional neural networks was evaluated to determine an appropriate trade-off between predictive performance and resource demand.The selected FAWINSTARNet configuration attained an accuracy near 97% and was sufficiently lightweight for execution on mobile hardware, thereby supporting on-site pest surveillance for growers.The study offers a full account of dataset development, experimental procedures, architectural design, and comparative assessment of competing models.