In this paper, we investigate the application of neural network architectures for semantic segmentation of rodent burrows for monitoring their population in agricultural fields. In particular, three models for semantic segmentation are considered: convolutional autoencoder (CAE), SegNet, and U-Net. These models are applied to analyze images obtained from unmanned aerial vehicles (UAVs) and ground robotic means, which allows for automatic burrow detection, minimizing the need for labor costs in processing large amounts of data. A sample of 247 RGB images containing 1098 labeled burrows was prepared for training and testing the models. The quality indicators of semantic segmentation were assessed using the Jaccard metric (IoU), which resulted in the following values: 0.511 for CAE, 0.548 for SegNet, and 0.529 for U-Net. An assessment of the computational resources required to implement these models in on-board computing units (OCUs) of mobile robotic means was conducted. Two criteria were considered: the number of floating-point operations (GFLOPS) and the number of model parameters. The results showed that SegNet requires 2.23 GFLOPS and has 0.76 million parameters, which is 2.58 and 2.33 times less than SAE and U-Net, respectively. The number of floating-point operations for SegNet was also 2.43 and 1.88 times lower than that of SAE and U-Net, respectively. As a result, SegNet outperformed SAE and U-Net in both segmentation efficiency and required computational resources. This work was carried out as part of the implementation of a computer vision system for an agricultural robotic means.
