An adaptive image restoration and object detection model for UAVs in hazy and sandstorm environments

Sandstorm and haze conditions significantly degrade the quality of remote sensing images captured by unmanned aerial vehicles (UAVs), leading to reduced object detection accuracy and compromised model robustness. To address this challenge, this paper proposes an adaptive dehazing and dedusting object detection model, termed CA-YOLO. The model integrates an Adaptive Image Recovery Module (AIRM) and an object detection module, which are jointly trained using an optimization strategy that facilitates feature sharing across tasks. Within the AIRM, we introduce the Chain-of-Thought Prompted Module (COT) and the Adaptive Context Enhancement Module (ACEM). The COT module dynamically selects the appropriate restoration task (dehazing or dedusting) based on the image characteristics, while the ACEM module enhances contextual information to improve content understanding and facilitate accurate task selection. The image restoration process leverages the Sand-All-in-One Dedusting and Dehazing Network (SAND-AODNet) to ensure high-quality image recovery. For object detection, an enhanced YOLOv11 model is employed, incorporating the Frequency-Aware Feature Fusion-BiFPN (Freq Fusion-BiFPN) module to mitigate intra-class inconsistency and boundary ambiguity caused by dense and overlapping small objects in UAV imagery. The proposed model is evaluated on the Annotated Real-world Task-driven Testing Set (RTTS) dataset and multiple types of synthetic sandstorm datasets, demonstrating its superior performance in complex environmental conditions.