Addressing Missing-Modality Data Quality Issues in Multimodal Remote Sensing via Automated Encoder Architecture Search

Multimodal remote sensing has become an essential paradigm for land-cover classification by exploiting complementary information from heterogeneous sensors. However, in practical applications, one or more modalities are often unavailable at inference time due to sensor failure, acquisition constraints, or environmental conditions. Such missing-modality scenarios introduce a systematic form of data quality degradation, which fundamentally undermines the reliability of conventional multimodal models and limits the effectiveness of manually designed compensation strategies. To address this challenge, this paper proposes PE-MMRSS, an automated encoder architecture design framework based on evolutionary neural architecture search for multimodal remote sensing classification under missing-modality conditions. Instead of relying on handcrafted hallucination architectures, PE-MMRSS constructs a unified and generic search space that jointly accommodates encoders associated with both available and missing modalities. A dual-part encoding scheme is designed to represent coupled teacher-student architectures, enabling effective evolutionary optimization. To further improve search efficiency, a lightweight LSTM-based Structural Compatibility Predictor is introduced to estimate the relative effectiveness of candidate architectures using early-stop validation signals, thereby significantly reducing the computational cost of architecture evaluation. Experimental evaluations across a diverse set of multimodal remote sensing benchmarks (Huston2013 and Augsburg) show that the proposed method achieves superior performance compared with the manually designed baselines in a wide range of missing-modality scenarios. PE-MMRSS achieves improvements of up to 0.5%-1.6% in overall accuracy and Kappa coefficient compared with the strongest handcrafted baseline, while reducing the architecture search cost to approximately 1.4 GPU-days. These results indicate that automated neural architecture search provides an effective and efficient solution for mitigating data quality degradation in multimodal remote sensing systems.