Automatic Farm Insects Detection Using Individual/Fused EfficientNet Features

Deep learning (DL)-based data assessment techniques are widely adopted in agriculture to endorse sustainable farming practices. During this process, the necessary images are collected from the farm using a chosen imaging scheme and is then examined using a chosen DL-scheme. This research aims to propose a DL-tool for monitoring the common pest which provides the damage to the crops. This research proposes a methodology to process the digital images to accurately detect the beetle and grasshopper. The phases in the proposed DL-tool encompass: (i) data acquisition, resizing, and augmentation; (ii) extraction of deep-features utilizing chosen model, feature reduction, and serial fusion to generate a new feature vector, followed by classification and three-fold cross-validation. This study examined the pretrained EfficientNet (EN) to investigate the performance of the DL-tool, which is validated using individual-features (IF) and fused-features (FF) using a SoftMax classifier. The results of this study demonstrate that the IF-based detection achieves accuracy > 92%, while the FF-based technique attains an accuracy > 98% on the selected insect database. This confirms that the implemented technique works well in detecting the Beetle/Grasshopper from the chosen digital images.