Extended Abstract
Introduction
The purpose of extracting borders is to create a database of agricultural lands. With this information, it is possible to estimate the amount of water consumption and crop harvest and automate agricultural machinery's guidance. The field survey of land borders is time-consuming and expensive. Therefore, remote sensing methods can be used for this. Manually delineating agricultural land boundaries from remote sensing images is still difficult. Thus, using automatic algorithms is a good solution. These algorithms are divided into Traditional algorithms and algorithms based on deep learning. Traditional algorithms have problems such as incorrect positioning and poor detection accuracy. Therefore, researchers have proposed a variety of edge detection algorithms based on deep learning. Convolutional neural networks are among the algorithms used in this research.
 
Materials & Methods
To train convolutional neural networks, the open-access dataset of France and three datasets of Hamadan, Bahar, and Khersan were used. To conduct this research, Global mapper & Adobe Photoshop software was used to prepare images, draw label polygons, and unify the format of labels belonging to different datasets. Python programming language, Anaconda environment, and Cross and Tensor Flow libraries have been used to implement deep learning algorithms. Eight scenarios were tested with different datasets. Five of them were performed without the use of transfer learning and fine-tuning, and the other three were performed using fine-tuning. By combining U-Net architecture and pre-trained neural networks, five hybrid architecture modes were constructed. The attention mechanism was also used to improve the performance of the models. To evaluate the performance, Dice Score, IoU, Accuracy, Recall, and F1-Score metrics were calculated.
 
Results & Discussion
In the first scenario, the training and testing of the model was done on the French dataset. The sameness of these datasets in this scenario helps to compare different backbones. In the second scenario, the training and testing were done on the Hamadan dataset. In this scenario, due to the same training and test datasets, the performance of different backbones was compared. In the third scenario, training was done on the French dataset and testing on the Hamadan dataset. The purpose of designing this scenario is to evaluate the performance of the models on the images of different regions, with different spatial resolutions and sensors. In the fourth scenario, training was done on the French dataset and testing on the Bahar dataset. This scenario examines the possibility of using images with a spatial resolution of 10 meters for training and testing models. In the fifth scenario, training was done on the French dataset and testing was done on the Khersan dataset. In this scenario, the generalizability of the model was investigated. In the sixth scenario, training on the French dataset, fine-tuning on the Hamedan dataset, and testing on the Hamedan dataset were done. In this scenario, the effect of transfer learning was investigated. Also, the result of using images of different regions on the performance of the model was evaluated. In the seventh scenario, training on the French dataset, fine-tuning on the Hamadan dataset, and testing on the Khersan dataset were performed. One of the goals of this scenario is to evaluate the generalizability of the model. In the eighth scenario, the model was trained on the French dataset and fine-tuned on the Hamedan dataset. Finally, the test was performed on the Bahar dataset. The purpose of this scenario is to evaluate the effect of the spatial resolution of the test dataset images on the performance of the model.
 
Conclusion
In the first scenario, the accuracy of detection in the first, fourth, and fifth architectural states is similar and equal to 0.71. In the IoU metric. The recognition quality of the model is not favorable for small plots. One of the reasons is the low spatial resolution of images. In the second scenario, compared to the first scenario, the first, fourth, and fifth architectural states have suffered a drop in accuracy. Reducing the number of training datasets can be one of the reasons. In the third scenario, the number of Training images increased compared to the second scenario. As a result, the accuracy of the model's performance in the fourth architectural mode increased by 0.03 in the IOU metric compared to the second scenario. However, due to the difference in the resolution of the training and test datasets, the performance of the model was not satisfactory. The accuracy of the model in the fourth scenario decreased by 0.57 compared to the third scenario. The main difference between the third and fourth scenarios is in the spatial resolution of the test images. This shows that to achieve the desired results, the test images should have sufficient spatial resolution. In the fifth scenario, the accuracy of the first architectural mode decreased by 0.10 and the fourth architectural mode decreased by 0.18 in the IOU metric. This decrease in accuracy is due to the difference in topography and the size and shape of agricultural lands in the regions of France and Khersan. It can be concluded that the generalizability of the model is not suitable. The sixth scenario is the same as the third scenario, only fine-tuning on the Hamadan dataset has been added. The detection accuracy of the sixth scenario in the fourth architectural mode reached 0.87 and compared to the third scenario in the same architectural mode, it increased by 0.14. This shows that fine-tuning is effective in increasing model accuracy. The 7th scenario compared to the 5th scenario in the 4th architectural mode had a 0.04 increase in accuracy in the IOU metric due to the use of fine tuning. But still, the detection accuracy in this scenario was not desired. One of the reasons for this problem is the difference between agricultural lands in France and Khersan regions. It can be concluded that the existing models for use in different regions do not have proper generalization capabilities.