Rice plant counting in paddy field based on UAV and density map regression

Rice is one of the most important food crops, and the stability of its yield is crucial to national food security. Rice plant counting becomes critical in fine agriculture research and is closely related to rice yield. In this paper, we proposed a new multi-scale fusion based rice counting method (RPCNet), which consists of one feature extractor frontend and two feature decoder modules namely Density Map Estimator (DME) and Plant Object Recognizer (POR). Fineloss is introduced into DME to improve the network's ability to separate adherent plants. To verify the validity of our method, we conducted experiments on a high-throughput rice plant image dataset. Experiment results show that the MAE and RMSE of the proposed RPCNet are 2.6 and 3.4 respectively, which outperforms states-of-the-art methods. Results suggest that RPCNet can accurately and efficiently estimate the number of rice plants and replace traditional manual counting.