Abstract:
Accurate and efficient crop classification using multispectral remotely sensed data is
essential for crop yield estimation and agricultural management. However, one of the
major challenges in this task is the limited availability of labeled data, which hinders
the ability to achieve good classification results. In our study, we propose a two-step
approach to address this challenge and improve crop classification accuracy. Firstly,
we employ a self-supervised pre-training step using data extracted from Sentinel Hub.
This pre-training step utilizes unlabeled data to initialize the model and capture valuable
information about crop growth patterns. By leveraging the abundant unlabeled data, the
model learns to extract meaningful features and understand the contextual relationships
within the data. This enhances the model’s ability to classify crops accurately. In the
second step, we perform transfer learning for supervised classification using labeled data.
The weights obtained from the pre-training step serve as the starting point, and the
model is further optimized using the labeled data to improve its classification accuracy.
Our experiments demonstrate that incorporating self-supervised pre-training leads to
faster convergence and better results compared to training without pre-training. The
pre-training phase enables the model to acquire prior knowledge about crop growth
patterns, which facilitates more efficient learning and better generalization to unseen
data during the supervised classification step. Moreover, by utilizing multispectral data
instead of the traditional 4-channel data, our approach captures more comprehensive
and discriminative information, further enhancing the classification performance.
