MOSSNet : Multiscale and Oriented Sorghum Spike Detection and Counting in UAV images

Jianqing Zhao¹Zhiyin Jiao²Jinping Wang²Zhifang Wang²Yongchao Guo²Ying Zhou¹Shiyi Chen²Wenjie Wu¹Yannan Shi^2*Peng LV^2*

• ¹Jiangsu Second Normal University, Nanjing, China
• ²Institute of Millet Crops, Hebei Academy of Agriculture and Forestry Sciences, Shijiazhuang, Hebei Province, China

Accurate sorghum spike detection is critical for monitoring growth conditions, accurately predicting yield, and ensuring food security. Deep learning models have improved the accuracy of spike detection thanks to advances in artificial intelligence. However, the dense distribution of sorghum spikes, variable sizes and complex background information in UAV images make detection and counting difficult. Existing horizontal detection methods often results in overlapping detection boxes during detection and counting, which includes large amounts of background information. This study proposes a multiscale and oriented sorghum spike detection and counting model in UAV images. The model creates a Deformable Convolution Spatial Attention (DCSA) module to improve the network's ability to capture small sorghum spike features. It also integrates circular smooth labels to enable the network to extract directional information and effectively represent morphological features. The model also employs a Wise IoU-based localization loss function to improve network loss for bounding box alignment under complex field conditions. To further refine detection results, we remove redundant oriented detection boxes, and provide detailed output on spike size, position, angle, and class information. The results show that MOSSNet achieves a mean average precision (mAP) of 90.3% for sorghum spike detection, with a RMSE and MAE of 9.3 and 8.1, respectively. These results demonstrate the effectiveness of MOSSNet in complex field environments with significant variability and severe occlusion, providing a valuable technical reference for sorghum phenotyping.