A Comparative Study of Recognition Models Based on Fine-Tuning 3D CNNs for Farming Behaviors

Highlights Proposed application of 3D CNNS for recognition of farming behavior. Transfer learning was used to speed up training and improve model accuracy. A farming behavior dataset was constructed, expanded and compared with previous studies. An object detection network was used for data preprocessing rather than using traditional methods. Abstract. The quality and quantity of crop yields in agriculture primarily depend on the timing and precision of various implemented farming behaviors. Basins and hills dominate southwest China. Due to topographical constraints, the rate of agricultural mechanization in the region remains low, and agriculture remains primarily non-mechanized. The acquisition and recognition of information on farming behaviors play an important role in crop production. In this article, transfer learning was used in a current state-of-the-art 3DCNN-based behavior recognition model for farming behavior recognition and classification tasks. The focus was on fine-tuning and evaluating state-of-the-art 3D convolutional neural networks for farming behavior recognition. The evaluated architectures included Res3D, MC3, and R2+1D. The six common farming behaviors recognized include weeding, planting, harvesting, transplanting, fertilizing, and spraying. The accuracy of all models pretrained on Kinetics-400 after fine-tuning exceeded 90%, where MC3 had the best performance, with an accuracy of 0.9628, precision of 0.9647, sensitivity of 0.963, and specificity of 0.9925, which was slightly greater than the other two. MC3 was also the most lightweight of all models; its parameters were only 32.6% of Res3D and 36.7% of R2+1D. The experimental results demonstrated that the fine-tuned MC3 model offers high classification accuracy and effective recognition and classification of farming behaviors, which lays a good foundation for improved crop production. Keywords: Deep learning, Farming behavior recognition, Farm management, Fine-tuning, Precision agriculture, 3D convolutional neural networks, Transfer learning.