Abstract
Abstract The Food and Agriculture Organization (FAO) estimates that pests cause up to 40% of the world's crop production to be lost each year. Therefore, timely crop pest detection is essential for sustainable agriculture, yet it is still difficult in large, noisy field settings. This study presents a hybrid DL model called AudioNet-Pest, which enables the monitoring of pests on a large-scale agriculture field. The pest sounds are segregated and coded into potent tabular features through statistical feature extraction using the InsectSet32 dataset and processed through a hybrid preprocessing pipeline which comprehensively cleans the audio to give reliable inputs to the detection steps. This research proposes a hybrid backbone with: (i) YAMNet model, which contributes each representation to a 96 × 64 patch and learns multi-scale spectro-temporal representations with depth wise-separable MobileNetV1, and (ii) TabNet model, which predicts sparse feature at each step using sparsemax masks; gated linear units to obtain interpretable decision embeddings. Following a channel-aligning and fusing of the pyramids with learnable weights a multi-stage head processes them and finally a lightweight detection head is produced. Stages are semantic enrichment of HLSFEB, bidirectional scale fusion of Bi-FPNB, fine detail of LLFEB, channel-spatial attention of ConvBAP and a DSB of receptive-field reweighting. The experimental outcomes are more favorable to compare to the benchmark models, and achieve higher accountability of 99.90% accuracy, 100% specificity, 99.90% sensitivity and recall, 99.90% precision, and 99.90% F1-Score. This research has also helped in the development of a scalable and automated model that helps in pests’ identification and could have decent prospects of application in industrial farming as a method of enhancement of sustainability and yield.