Reinforcement learning-driven small droplet detection for precision agricultural spraying

Spray systems play an important role in precision agriculture. To analyze the efficiency of a spray system, droplet characteristics are often considered. While small droplets tend to cover a larger surface area, they are more vulnerable to evaporation and wind drift. Large droplets, on the other hand, have greater mass and downward velocity but cover less surface area than smaller ones. In addition, small droplets pose a significant challenge to detect and track due to their size and blending with complex backgrounds. To address this problem, we propose a deep learning framework for small droplet detection in agricultural spraying, integrating an efficient multi-scale feature extraction architecture, an attention-based zooming mechanism, and reinforcement learning (RL) optimization. Our approach employs a convolutional backbone to extract robust features, a multi-scale feature pyramid to handle diverse droplet sizes, and an attention-driven zooming strategy to enhance small droplet detection during inference. RL is used to train a policy network that dynamically selects optimal focus regions to improve detection performance. The droplet detector is trained on a dataset of 1,489 droplet images, augmented to 1,904 for training. Through extensive experiments, we demonstrate that our framework achieves strong performance, outperforming baseline methods with an [email protected] of 0.990 on the test set, along with a precision of 0.961 and a recall of 0.986. Our qualitative analysis further highlights the framework's ability to detect small droplets in dense clusters and challenging lighting conditions (overly dark or bright scenes), even identifying droplets missed in the ground truth annotations. With an inference time of just 18.1 ms, the framework provides a practical solution for precision spraying and enhances agricultural efficiency. To the best of our knowledge, this work is the first to automatically capture characteristics of small droplet analysis using reinforcement learning, thereby paving the way for future advancements in real-time detection and broader agricultural applications.