CC-SSL: A Self-Supervised Learning Framework for Crop Classification With Few Labeled Samples

Abstract

Labeled samples with real crop types are important for crop classification, but the acquisition of large batches of labeled samples will consume many resources, so it is necessary to study crop classification based on few labeled samples or no labeled samples. To solve the problem of labeling sample dependence in crop classification, this article proposes a self-supervised learning crop classification framework (CC-SSL) that only requires a few labeled samples. The framework includes a new SSL algorithm and adds a tensor transformation and sample processing module to ensure that the framework can be applied in crop classification. Specifically, the tensor transformation module designs a content-rich input tensor that is designed to represent crop growth patterns intuitively and effectively. The sample processing module provides a simple and useful way to maintain sample balance, allowing SSL models to be trained valid. The new SSL algorithm Sim-SCAN can obtain important features from a small number of labeled samples and does not use any labeling information during the training process. Experimental results show that tensors with richer forms can obtain better OA and can more effectively characterize crop growth patterns. The experimental results of sample processing show that keeping the samples balanced through data augmentation can greatly improve the performance of the CC-SSL framework and obtain classification results that exceed supervised learning. The results of the experiments with reduced labeled samples show that the CC-SSL framework using only a few labeled samples can achieve classification performance and robustness comparable to supervised learning.

Keywords

• Crop classification
• data augmentation
• few labeled samples
• sample balance
• self-supervised learning (SSL)