Agricultural Field Boundaries Extraction from Images with Deep Learning and Convolutional Neural Network

نوع: Type: Thesis

مقطع: Segment: masters

عنوان: Title: Agricultural Field Boundaries Extraction from Images with Deep Learning and Convolutional Neural Network

ارائه دهنده: Provider: Sajjad Yavari

اساتید راهنما: Supervisors: DR Morteza Heidari Mozzaffar

اساتید مشاور: Advisory Professors:

اساتید ممتحن یا داور: Examining professors or referees: - Dr Hasan Khotanlou - Dr Hossein Torabzadeh

زمان و تاریخ ارائه: Time and date of presentation: 2024

مکان ارائه: Place of presentation: 44

چکیده: Abstract: delineation of agricultural field boundaries is necessary to statistical examination of agricultural products, creation of cadastral maps and database of agricultural lands. With this information, it will be possible to estimate the water consumption and determine the harvest amount. It can also be used to automate the steering of agricultural machinery. Due to the fact that the field survey of field boundaries is time-consuming and expensive, remote sensing methods are a reasonable choice for this task. But it is still difficult to manually draw the boundary of agricultural land from remote sensing images. Therefore, using automatic methods is a good solution. In this research, first, the technical and natural factors influencing the process of extracting field boundaries have been investigated. Then, different types of boundary detection algorithms as a linear feature have been discussed. Among them, the convolutional neural network algorithm, which is a type of image hierarchical boundary extraction algorithm, has been chosen to continue the work. Boundary extraction is a kind of image segmentation. Therefore, different types of convolutional neural network architectures about image segmentation were investigated. One of the problems of using convolutional neural networks is the lack of a training dataset with suitable size. In this research, solving the problem of Suitable Training dataset has been done by using transfer learning and fine-tuning techniques. In order to train convolutional neural networks, the open access dataset of France was used along with three data sets of Hamadan city, Bahar city and Khersan region. These three datasets were prepared in order to investigate the effect of transfer learning as well as the generalizability of the models. Eight experimental scenarios were considered for training, fine-tuning and testing. Five of these tests were performed with fine tuning and three others without fine tuning. Also, five different architectural modes of the U-Net network were implemented along with different backbones. To evaluate the performance of the method, Dice Score, IoU, Accuracy, Recall and F1-Score metrics were calculated. finally, it was found that transfer learning and fine-tuning are a way to compensate for the lack of training data and increase the accuracy of the model. Using fine-tuning, the performance accuracy of the best test scenario in the IoU metric went from 73% to 87%, an improvement of 14%. Also, the Attention mechanism in combination with neural network architectures improves the accuracy of boundary extraction. Another parameter is the spatial resolution of the images, which has a direct relationship with the accuracy of the model. The results show that the existing architectures have not yet reached a suitable generalization capability for different regions.

فایل: ّFile: Download فایل