As food security gains increasing attention, automated pest monitoring is crucial for agricultural early warning systems. However, in practical light-trap capturing sensors, the extremely small scale of pests and complex background interference, such as unexpected reflection and occlusions, severely undermine the performance of existing models, resulting in frequent missed and false detections. To deal with these challenges, this study proposes EP-YOLO, an enhanced lightweight detection architecture based on YOLOv8n. Specifically, to retain the spatial pixels of micro-targets during downsampling and isolate pest features while eliminating background noise without compromising channel information, the Spatial-to-Depth Convolution (SPD) module and the Efficient Multi-Scale Attention (EMA) module are introduced. We evaluate our model through experiments on Pest24, a dataset consisting of 24 tiny pest categories. The results demonstrate that EP-YOLO achieves a mAP@50 and mAP@50:95 of 70.5% and 47.3%, respectively, improving upon the baseline by 1.1% and 1.9%. Furthermore, EP-YOLO achieves a significant improvement in detecting certain extremely small pests. For example, Rice planthopper and Plutella xylostella show improvements of 8.4% and 3.1%, respectively, compared to the baseline. In conclusion, the physical limitations of detecting tiny pests are successfully overcome by EP-YOLO, providing a robust and deployable design for real-time agricultural monitoring systems.

Larch casebearer (Coleophora laricella) poses a serious threat to forest ecological security, and remote sensing object detection is essential for early warning and targeted intervention. However, existing detectors are hindered by insufficient feature extraction, limited self-attention inductive bias, and weak spatial localization, which collectively constrain their practical deployment. To address these issues, we propose DROMAL-Net, which integrates global feature enhancement and positional dynamic clustering for forest pest detection. Built upon the YOLOv11n baseline, our model incorporates three key innovations. First, we design the MALAPSA module by embedding a magnitude-aware linear attention (MALA) mechanism into the feature extraction pipeline, achieving linear computational complexity while preserving global contextual modeling. Second, to address the inductive bias deficiency of C2PSA and better preserve spatial topology, we introduce C3kDR (DCCC3k), a positional dynamic clustering module that combines dynamic clustering convolution with complex-domain positional compensation to enhance multi-scale localization accuracy. Third, we integrate 2D Rotary Position Embedding (2D-RoPE) to further mitigate spatial ambiguity in dynamic clustering.Extensive experiments on the public SLFPD dataset validate the effectiveness of each component. Under the standard random partition, DROMAL-Net achieves 77.98% mAP@0.5, outperforming YOLOv11n by 2.46 percentage points. Furthermore, under a strict leave-one-block-out protocol to prevent spatial leakage, DROMAL-Net consistently outperforms YOLOv11n across all five held-out blocks, achieving a macro-average mAP@0.5 of 69.12% vs. 65.94%, corresponding to a 3.18 percentage point improvement, which confirms its strong generalization capability to unseen geographic regions. Results across five random seeds yield consistently low standard deviations, confirming the stability of these gains. We further evaluate generalization on the TreeFinder dataset, which encompasses diverse tree species and complex canopy structures, where DROMAL-Net maintains robust performance, underscoring the effectiveness of MALAPSA and C3kDR for global feature enhancement and accurate localization in cross-regional applications.