Document Type : Original Research Paper
Authors
1 Dept. of Electrical and Computer Engineering, University of Science and Technology of Mazandaran, Behshahr, Iran
2 Dept. of Geomatics Engineering, Babol Noshirvani University of Technology, Babol, Iran
Abstract
Background and Objectives: With the advancement of technology and the emergence of multifunctional satellites, a significant amount of real time information is now transmitted from the Earth's surface. These satellites are equipped with sensors that obtain critical information by sending signals at various frequencies to the Earth's surface. The data received from these satellites are utilized in various scientific and military applications, including aviation, geographical studies, meteorology, agriculture, and other research fields. The agricultural sector and crop monitoring have especially benefited from remote sensing methods compared to traditional methods, becoming a primary tool for collecting environmental information for area monitoring applications, attracting researchers' attention. One such application is regional monitoring to examine agricultural products over cultivated areas. The use of remote sensing tools and satellite images is highly efficient due to their wide regional coverage. To automatically analyze these images, classify, and segment cultivated areas, machine learning methods are currently employed. Among these methods, deep learning offers superior performance and higher speed compared to other learning methods, such as manual or semi-automatic methods.
Methods: This paper utilizes deep learning models suitable for segmenting agricultural areas. Generally, these models produce an output of equivalent dimensions for each input. Therefore, for working with satellite images, an improved U-Net model is proposed in this study. The proposed model is developed using Vision Transformers (ViT) in the model's bottleneck for classifying and segmenting four types of agricultural products: rice, wheat, canola, and corn. Compared to convolutional layers, ViT is more efficient in terms of conceptual and algorithmic implementation and requires less computational power. This model addresses the problems and weaknesses of the base U-Net model that arise with complex datasets, diverse in shape, size, and texture, enabling more accurate and reliable segmentation results. Additionally, the proposed improvements enhance the model's robustness and adaptability to various agricultural scenarios.
Findings: In the numerous experiments conducted, the proposed method achieved an accuracy of 83.84 and Precision of 70.69%, providing a better classification of the five target products compared to other methods. The qualitative outputs also indicate better segmentation of the input images when applying the proposed method. Alongside the accuracy metric, other metrics such as focal loss, recall, precision, and MIoU were examined, with the proposed method reaching acceptable levels in most cases. Notably, as the target area was in Iran, data collection and labeling were also carried out in this research, providing a suitable dataset for training other models
Conclusion: This research presents an end-to-end model for learning features related to the segmentation of satellite images. The results indicate that the proposed method can be effectively used for segmenting satellite images received from Sentinel-2 for the target products, such as various crops. Therefore, the results obtained can play a crucial role in water consumption management, planting structure adjustment, loss estimation, and agricultural performance evaluation, providing significant insights for stakeholders. By utilizing these methods, it is possible to achieve improved efficiency and accuracy in agricultural management and optimize resource use in this area, contributing to sustainable agricultural practices and better decision-making processes.
Keywords
Main Subjects
COPYRIGHTS
© 2024 The Author(s). This is an open-access article distributed under the terms and conditions of the Creative Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) (https://creativecommons.org/licenses/by-nc/4.0/)
)