Automatic Farmland Recognition of Remote Sensing Images Using Transfer Deep Learning

Abstract

Abstract Aiming at the feature extraction of farmland from remote sensing images relying on greatly manual interpretation, which consumes a lot of manpower, financial resources, and is inefficient. This paper proposes a remote sensing image farmland recognition method based on a transfer deep learning model. We also propose a simple but effective method to overcome the problems of unclear edge segmentation and partial field of view of the convolutional network. The experimental results show that the PA of the U-Net network model has reached PA 0.9124, mPA 0.7757, mIoU 0.6832, Recall 0.9586, Precision 0.94215, F1-Score 0.9503. We have obtained a competitive result and the U-Net neural network has an excellent capacity for remote sensing image farmland recognition, robustness, and practicability.