•  
  •  
 

Turkish Journal of Electrical Engineering and Computer Sciences

Author ORCID Identifier

JUTANG SWER: 0009-0006-0118-9476

DEBDATTA KANDAR: 0000-0002-3409-5189

GOUTAM SAHA: 0000-0003-0679-5855

Abstract

Grasshoppers can seriously harm crops, but it can be challenging to identify them in field images because they might appear in different lighting settings and blend in with the background. A U-Net-based grasshopper segmentation model is presented in this study. MobileNetV2 is used as the encoder to reduce computational cost, while squeeze-and-excitation blocks are combined with the decoder to improve feature selection during segmentation. An exclusive dataset of 4743 manually annotated RGB images taken from rice fields in Meghalaya, in India's Ri Bhoi District, was used to test the model. Two separate training runs with different random seeds were carried out using the same fixed training-validation split. A mean Dice score of 0.9052 ± 0.0022 and a mean intersection over union value of 0.8278 ± 0.0032 were attained using the proposed model. Approximately 11.57 million parameters and 56.988 billion floating point operations are needed for the model. It was also contrasted with DeepLabV3+, the conventional U-Net, and other U-Net variations. Further experiments were conducted using partial occlusion, blur, and brightness adjustments. The results show that the proposed model achieves good segmentation performance with reduced computational cost compared with the standard U-Net, although its performance decreases under strong occlusion.

DOI

10.55730/1300-0632.4289

Keywords

Grasshopper segmentation, MobileNetV2, grasshopper detection, squeeze-and-excitation blocks, U-Net architecture

First Page

1054

Last Page

1068

Publisher

The Scientific and Technological Research Council of Türkiye (TÜBİTAK)

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Share

COinS