An Improved PSPnet Model for Semantic Segmentation of UAV Farmland Images

Abstract

【Background and objective】 Deep Learning is an important method in artificial intelligence but has limited applications in agriculture. This paper aims to fill this technology gap based on farmland images acquired using unmanned aerial vehicle (UAV) and the PSPnet segmentation method. 【Method】 Different dimensional features in the UAV images were concatenated based on the principle of preserving as many detailed image features as possible via the enhanced scene analysis. A lightweight semantic segmentation model was built using the deep separable convolution module. Finally, the activation function was replaced to improve the segmentation effect of the model. 【Result】 Experimental results show that the mean pixel accuracy and the mean intersection over the union of our proposed method are 89.48% and 82.36%, respectively, and their associated segmentation accuracy was improved by 18.12% and 18.93%, respectively. Overall, the segmentation of the proposed method was better than that of Unet and DeeplabV3+. 【Conclusion】 The proposed method can effectively segment the farmland images acquired by UAV.

Keywords

• pspnet
• semantic segmentation
• feature concatenate
• deep separable convolution
• activation function