A grid-level segmentation model based on encoder-decoder structure with multi-source features for crop lodging detection

Crop lodging assessment plays a critical role in acquiring accurate information regarding the location and area of lodging, which is essential for loss assessment and adjustments of harvester parameters. In this paper, we proposed LDVO (Lodging Detection with Visible-image Only), a comprehensive grid-to-grid semantic segmentation method for timely and accurate identification of crop lodging. The LDVO model uses Inception block and dense connection to construct a lightweight feature extraction network, and complements the texture feature and crop vegetation indices as reference features for semantic segmentation. Besides, the model meshes the aerial images according to the operation characteristics of the harvester, and the accuracy of the segmentation task is reduced from the pixel level to the grid level, which minimizes the network scale and computing cost under the premise of meeting the accuracy requirements of lodging detection. Experimental results demonstrate the superiority of the proposed LDVO model over mainstream semantic segmentation networks in terms of processing speed and model parameters. Remarkably, the LDVO model achieves the highest prediction accuracy of 94.86% by leveraging a combination of RGB semantic features, VIs and texture features. Therefore, the proposed LDVO model provides a fast, feasible and low-cost reference for monitoring crop lodging status in complex field environments. It also offers a universal idea for the improvement of semantic segmentation network in special application scenarios.