Recognition Method of Flat and Ridged Crop Types in Dry Fields Based on Propriety Sensing Signals of Agricultural Robot

Abstract

Dryland agricultural cultivation modes include flat cropping and ridge cropping, and the terrain undulation of different cultivation modes varies greatly, so accurate crop row cultivation mode recognition is of great significance to the stability of robot travelling. A methodology for identifying the terrain of crop rows and ridges utilizing appropriate sensor signals was introduced. Initially, the inertial measurement unit ( IMU) signals were collected from a quadrupedal robot navigating through the crop rows of a corn field. The velocity data from the robot’s left front leg served as supplementary information to compile a comprehensive signal dataset, encompassing the robot’s movement in both flat cropping and row cropping modes, with two distinct row heights. Subsequently, spatial information features were extracted from the signals by using convolutional neural networks (CNN), while time series features were derived through bidirectional long short-term memory (BiLSTM) networks. Additionally, self-attention (SA) was employed to capture the attention scores of the output feature information from both CNN and BiLSTM. Ultimately, the efficacy of the proposed model in distinguishing between flat and ridge crop types was validated through model comparisons and field experiments. The results indicated that the F1 score of proposed CNN BiLSTM SA model reached 92% , marking an improvement of 10.17, 3.51, 2.57 and 1.27 percentage points over that of the CNN, CNN LSTM, CNN LSTM SA, and CNN BiLSTM models, respectively. When the recognition model was embedded in the field robot, it achieved a 90% accuracy rate in identifying the current crop row tillage type within 1.4 s, and met the classification criteria for flat and ridge categories within 4.8 s. This performance satisfied the robot’s requirements for rapid and accurate recognition across various tillage terrains. The algorithm can provide the robot with ability to recognize crop rows under typical tillage patterns in dry fields, and the results can provide technical support for improving the field stability of quadrupedal robots in autonomous operations.

Keywords: agricultural robots, inertial measurement units, proprioceptive signals, terrain recognition, long and short-term memory networks