Loading...
Search
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
Digital Object List
Bookmark
Review List
53 viewed
تقسیم بندی نمونه ی زمین های کشاورزی توسط یادگیری عمیق
شمشیرگرها، محمد رضا Shamshirgarha, Mohammad Reza
Deep Learning for Instance Segmentation of Agricultural Fields
Shamshirgarha, Mohammad Reza | 2022
296
Viewed
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 55969 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Manzuri Shalmani, Mohammad Taghi
- Abstract:
- Geographical data, agricultural field boundaries and their segmentation are essential for many agricultural applications. For example, monitoring of field parcel for resource management. Since manual delineation of land parcels with the help of a real person requires a lot of time and special tools, the need for repeatable automation of this work is felt. Traditional approaches of image segmentation do not have enough generalizability and can be used only for specific areas; so we turned to deep learning, which has proven to be successful in computer vision tasks. Instance segmentation is the most advanced deep learning-based method in object recognition and has numerous applications in remote sensing data that can produce significant results. The goal of this research is to process satellite images and detect the agricultural fields in them. For this purpose, we used the Mask RCNN model, which is one of the leading models presented for instance segmentation in deep learning. In this work, 3 main tasks were done: (a) Designing the system to convert polygons of ground truth data and satellite images into COCO annotation format. (b) Modifying the code of the detectron2 framework to be used in areas such as Iran where the farming lands have irregular shapes and different textures, as well as the compatibility of the model with images with different numbers of input channels. (c) Providing a solution for semantic segmentation based on instance segmentation outputs. To test the proposed method, we prepared several datasets that are located in 3 locations in Denmark, the Netherlands, and Iran. For evaluation, we used COCO instance segmentation metrics and standard metrics of semantic segmentation. In the baseline experiment of Denmark dataset, where we used Sentinel-2 RGB images with a resolution of 10m, we achieved the overall average precision (AP) of 54.4% for instance segmentation as well as an F1 score of 90.3% and an overall accuracy of 89.8% in the semantic segmentation
- Keywords:
- Deep Learning ; Neural Network ; Sentinel Satellite ; Instance Segmentation ; Agricultural Field Segmentation ; Mask RCNN Model
-
محتواي کتاب
- view