Wheat Lodging Segmentation Based on Lstm_PSPNet Deep Learning Network

Abstract

Lodging is one of the major issues that seriously affects wheat quality and yield. To obtain timely and accurate wheat lodging information and identify the potential factors leading to lodged wheat in wheat breeding programs, we proposed a lodging-detecting model coupled with unmanned aerial vehicle (UAV) image features of wheat at multiple plant growth stages. The UAV was used to collect canopy images and ground lodging area information at five wheat growth stages. The PSPNet model was improved by combining the convolutional LSTM (ConvLSTM) timing model, inserting the convolutional attention module (CBAM) and the Tversky loss function. The effect of the improved PSPNet network model in monitoring wheat lodging under different image sizes and different growth stages was investigated. The experimental results show that (1) the improved Lstm_PSPNet model was more effective in lodging prediction, and the precision reached 0.952; (2) choosing an appropriate image size could improve the segmentation accuracy, with the optimal image size in this study being 468 × 468; and (3) the model of Lstm_PSPNet improved its segmentation accuracy sequentially from early flowering to late maturity, and the three evaluation metrics increased sequentially from 0.932 to 0.952 for precision, from 0.912 to 0.940 for recall, and from 0.922 to 0.950 for F1-Score, with good extraction at mid and late reproductive stages. Therefore, the lodging information extraction model proposed in this study can make full use of temporal sequence features to improve image segmentation accuracy and effectively extract lodging areas at different growth stages. The model can provide more comprehensive reference and technical support for monitoring the lodging of wheat crops at different growth stages.

Keywords

• unmanned aerial vehicle
• deep learning
• Lstm_PSPNet
• lodging
• wheat