Assisted learning for land use classification: The important role of semantic correlation between heterogeneous images
This is the core code of our work published in ISPRS J. :kissing_heart:
This work propose an innovative assisted learning framework that employs a "teacher-student" architecture equipped with local and global distillation schemes for land use classification on heterogeneous data.
It has several advantages as outlined below:
:star: Ability to maintain performance during testing with missing modalities
:star: High interpretability demonstrating knowledge transferability between different modalities
:star: Simplicity and flexibility
Below is the process of the proposed framework:
Step 1: Training teacher models using cross-entropy and dice loss. (Using train_t.py to train the teachr.)
Step 2: Training student models using our framework. (Using train_s.py to train the teachr.)
Step 3: Testing.