Deep localization model for intra-row crop detection in paddy field

Abstract Automated and precise rice plant localization is crucial for the mechanization of rice production, which can facilitate targeted spraying, site-specific fertilization, and mechanized weeding etc. Existing approaches adopted thus far have mainly focused on inter-row weed detection or rice seedling row detection. Nevertheless, techniques for intra-row individual rice plant positioning continue to face major challenges induced by the specific paddy field environments or complex morphology of rice plant. This paper proposed a new deep localization network for intra-row rice detection at the single plant level in a paddy field. This method designed a two-stage model. The module in stage 1 identified potential locations containing rice plants in the entire image. The module in stage 2 predicted the confidence of rice plant identification and refined the corresponding box bounds. The two-stage processing modules shared a deep backbone network for learning full-image convolutional features and are combined into a unified framework to facilitate an end-to-end training. In addition, we constructed a rice plant detection dataset and proposed a task-oriented evaluation method for performance verification of the algorithm. Experiment results showed the proposed deep model achieved a high localization accuracy of 93.22% and a high testing speed of 15 fps, verifying the effectiveness and efficiency of the method. Using this method, we can develop techniques for finer-level agriculture production, such as spraying and weed control, to achieve healthy and economical rice yields.