CPD: A Lightweight Image Processing Framework for Cloud Presence Detection in UAV Imagery

Cloud presence in UAV imagery introduces radiometric distortions that degrade visual quality and interfere with downstream image analysis tasks. This paper presents CPD, a lightweight image processing framework for detecting cloud presence in UAV-acquired images. The proposed method leverages a combination of dark channel prior and local contrast measurement to generate a binary haze map for image-level classification. Unlike deep learning-based cloud detectors, CPD requires no training data and performs inference in a fast, scalable manner. To further improve runtime efficiency, we implement an optimized GPU-enhanced version using convolution-based approximations of morphological and statistical operators. Experimental results on three datasets, including public satellite and haze datasets as well as a UAV-based dataset, demonstrate the method’s accuracy and runtime advantage. CPD achieves over 99% detection accuracy with up to 5× runtime improvement on mid-range consumer hardware, validating its applicability in real-time UAV image processing pipelines.