Timely monitoring and control of forest pests is of great significance for maintaining forest health, carbon sequestration and biodiversity. However, due to the influence of environmental factors such as rain, fog and insufficient light, it is often very difficult to track forest pests in the field, which leads to the deterioration of image quality and is difficult to use traditional methods for detection. In this paper, a new deep learning method PCSNet is proposed. In order to solve the problem of forest pest detection under harsh conditions, Chain of Mind Prompt Adaptive (CPA) enhancement module is added to the network, which can perform autonomous super-resolution enhancement for different degradation types of images. Wavelet Transform convolution (WTConv) and wavelet transform Down-sampling module (ADown) were proposed to speed up model convergence and optimize network performance. In addition, for edge devices, we have developed PCSNet-Light, a pruned and refined version that enables smaller model sizes with high accuracy. Experiments show that when using Raspberry PI 5 as the hardware platform, the mAP50 of PCSNet-Light reaches a high level of 96.6%, while maintaining a real-time detection speed of 15.67FPS while greatly reducing the number of parameters by 5 million. This micro-resolution strategy not only enhances the accuracy of pest detection, but also supports efficient and scalable pest monitoring systems that can help achieve more sustainable forest management actions.
