Rice seedling row detection based on morphological anchor points of rice stems

Field management in rice seedling paddies through vision navigation assistance is an effective measure for rice production automation and yield promotion. Rice row detection lays the foundation for vision navigation while also presenting challenges. One crucial issue hindering further improvement in rice row detection performance is the irregular growth morphology of rice plants. Rice leaves' divergent growth postures and uneven orientations make it challenging to accurately determine the actual growth site of rice seedlings. In this study, rice stems, which maintain a more stable growth posture than the leaves, were used as the primary recognition objects to eliminate the interference caused by rice leaves, thereby promoting rice row detection accuracy. Transformer-based semantic segmentation models were adopted to identify triangular morphological masks on the stem of individual rice seedlings. The anchor points representing the ground-breaking position of rice seedlings were then calculated from the predicted stem masks. A dynamic search direction-based clustering algorithm was presented to group the sparsely distributed anchor points into rows and complete the row fitting simultaneously. The proposed stem-recognition-based method achieved an excellent rice row detection performance with up to 92.93% detection accuracy under complex natural conditions, which was far superior to the 16.36% obtained by the leaf-recognition-based approach.(c) 2022 IAgrE. Published by Elsevier Ltd. All rights reserved.

Automated summary

Field management in rice seedling paddies can be improved through vision navigation assistance, with a focus on rice row detection. The irregular growth morphology of rice plants poses a challenge to accurately determine the growth site of rice seedlings. This study proposes a stem-recognition-based method that achieves excellent rice row detection performance, outperforming the leaf-recognition-based approach.