DDPM-based feature learner enables label-efficient transfer in land cover classification

Denoising diffusion probabilistic models (DDPMs) have demonstrated remarkable capacity in generative learning by capturing rich, transferable representations. In this study, we harness a DDPM-based feature learner for label-efficient land cover classification, a critical task in ecosystem monitoring and environmental assessment. Accurate and timely land cover mapping supports biodiversity protection, land use regulation, and climate impact analysis. However, traditional approaches remain heavily reliant on costly expert-annotated data. We propose an unsupervised generative pretraining framework using large-scale remote sensing imagery to learn spatial and semantic patterns without labels. The learned representations from the diffusion model’s decoder are transferred to downstream land cover segmentation tasks through a lightweight MLP segmentor. This design enables efficient adaptation to diverse ecological scenes with minimal supervision. Experimental results show that our method achieves competitive performance on land cover classification benchmarks, while requiring significantly fewer labeled samples. These findings highlight the promise of diffusion-based generative learning in scalable, label-efficient ecosystem observation.