Existing methods always neglect the value of channel features in processing multi-modal agricultural aerial images with Near-Infrared (NIR) characteristics, as well as ignore the importance of channel features in fusion features that contain both downsampled low-level features and upsampled high-level features. Two modules are proposed in this paper to address this issue. For fusion features composed of spatial information, semantic information, and multi-modal information, a transformer-based channel feature enhancement module is first constructed to facilitate the recognition of fusion features located in different channels. The second module is Dual Cross-Entropy Dice (Dual-CE-Dice) loss, which can improve the phenomenon of class imbalance while helping the model to better learn channel features. Extensive experiments have been conducted on the Agriculture-Vision-2021 dataset and the Tianchi suichang-round1 dataset, proving that the proposed method Channel Transformer (CFormer) is superior to the previous methods.

Knowledge representation learning (KRL) aims to encode entities and relationships in various knowledge graphs into low-dimensional continuous vectors. It is popularly used in knowledge graph completion (or link prediction) tasks. Translation-based knowledge representation learning methods perform well in knowledge graph completion (KGC). However, the translation principles adopted by these methods are too strict and cannot model complex entities and relationships (i.e., N-1, 1-N, and N-N) well. Besides, these traditional translation principles are primarily used in static knowledge graphs and overlook the temporal properties of triplet facts. Therefore, we propose a temporal knowledge graph embedding model based on variable translation (TKGE-VT). The model proposes a new variable translation principle, which enables flexible transformation between entities and relationship embedding. Meanwhile, this paper considers the temporal properties of both entities and relationships and applies the proposed principle of variable translation to temporal knowledge graphs. We conduct link prediction and triplet classification experiments on four benchmark datasets: WN11, WN18, FB13, and FB15K. Our model outperforms baseline models on multiple evaluation metrics according to the experimental results.