# A lightweight detector with hybrid pooling and checkerboard attention for solar panel anomalies

## TL;DR

This paper introduces a lightweight model for detecting solar panel anomalies in agricultural IoT systems, achieving high accuracy and real-time performance on edge devices.

## Contribution

A novel lightweight detector with hybrid pooling and checkerboard attention for efficient solar panel anomaly detection on edge devices.

## Key findings

- YOLOv11-HPC achieves an mAP50 of 84.1% and precision of 94.13% on solar panel anomaly detection.
- The model runs at over 55 FPS on an NVIDIA Jetson Orin NX in FP16 format, suitable for edge deployment.

## Abstract

The reliable operation of solar-powered agricultural Internet of Things (IoT) devices heavily depends on the integrity of solar panels. However, monitoring these distributed assets for subtle anomalies such as bird droppings, cracks, and dust accumulation remains challenging under edge computational constraints. This paper presents YOLOv11-HPC, an optimized lightweight detector that incorporates a Hybrid Pooling Spatial Pyramid Pooling Fast module and a Dual-path Multi-scale Checkerboard Attention module. These components collectively improve multi-scale feature representation and introduce sparse attention-guided refinement, enabling accurate identification of small and complex anomalies with low computational overhead. Evaluated on a dedicated solar panel anomaly dataset, YOLOv11-HPC achieves an mAP50 of 84.1% and a precision of 94.13%, surpassing existing YOLO models and classical detectors. When deployed on an NVIDIA Jetson Orin NX, the model sustains real-time inference at over 55 FPS in FP16 format, confirming its practical suitability for edge-based agricultural IoT device monitoring and sustainable agricultural applications.

•Propose a solar panel anomaly detection method for agricultural IoT•Enhance multi-scale small target feature representation•Demonstrate practicality by comprehensive validation on the edge device

Propose a solar panel anomaly detection method for agricultural IoT

Enhance multi-scale small target feature representation

Demonstrate practicality by comprehensive validation on the edge device

Applied computing; Applied computing in engineering