Space-ground laser communication links are highly susceptible to interruptions
caused by cloud cover. Reliable dynamic scheduling of ground stations and
seamless link handover, therefore, requires accurate short-term cloud-layer
forecasts. However, cloud evolution is inherently chaotic, and atmospheric
conditions exhibit evident seasonal non-stationarity. These factors weaken the
prediction stability of existing models and make it difficult to maintain
consistent accuracy over extended forecasting horizons. To address these
challenges, this paper proposes KFMNet, an adaptive cloud prediction network
that integrates Koopman theory with frequency-domain Mamba modeling. Unlike
conventional frameworks that mainly process cloud features in the spatial
domain, KFMNet embeds a linear state-space modeling mechanism into the
frequency-domain feature evolution process. Specifically, an adaptive spectral
gating mechanism is used to capture global cloud distribution patterns in the
frequency domain, thereby reducing the computational burden associated with
high-resolution image sequences. Meanwhile, a Koopman seasonal gating module
introduces monthly prior knowledge into the network, enabling the model to adapt
to atmospheric drift across seasonal transitions. Within the encoder-decoder
backbone, KFMNet jointly incorporates temporal modeling, frequency-domain
representation, and physical consistency constraints, providing a more stable
framework for short-term cloud prediction. We evaluate KFMNet on multi-season
cloud images collected over Nanjing and Beijing. The climate profiles of these
two cities are markedly different. Under these conditions, KFMNet achieves
competitive accuracy in cloud segmentation and shows a better ability to handle
seasonal and regional shifts than several reference models. In essence, our
method adapts well to changing atmospheric conditions and provides a new cloud
prediction approach for space-ground laser link assurance. This method combines
physical interpretability with the actual needs of real-world deployment.