LiDAR was one of the basic sensors for agricultural robot navigation in forests. However, due to the interference of the outdoor environment, obvious noise appeared in the LiDAR data, which reduced the navigation performance. To solve the problem that point cloud details are easily lost in point cloud denoising, an denoising algorithm was proposed based on dynamic filter radiu，and the denoising parameters were automatically determined. Besides, a convolutional neural network classifier was proposed, which was used to identify the planting pattern. By way of preset denoising parameters, it avoided the cumbersome parameter adjustment process and could be directly applied to dense planting and sparse planting scenarios. These approaches reduced the impact of point cloud density differences on noise removal, thereby achieving efficient denoising in large scenes. The denoising experiments in apple plantations, poplar forests and dry willow forests were completed. The results showed that the proposed method effectively removed multi-scale point cloud noise, and significantly reduced sparse outliers, dense noise, and noise around the target. It took 43.2ms to remove the noise of a single frame point cloud (6400 points). After denoising by the method, the accuracy rate of density clustering was 94.3%, and the recall rate was 78.9%. Compared with the original data, they were improved by 40.4% and 33.9%, respectively. The method had high real-time, versatility and robustness, and significantly improved the clustering effect.