Compact model improves Earth observation from multiple sensors and data types
MEOX: Compact Multimodal Mixture-of-Experts for Earth Observation
Computer Vision and Pattern Recognition
Summary
Earth observation often uses different sensors that can sometimes have missing data, making it hard to analyze. The authors created MEOX, a compact computer program that combines information from various sensors effectively without getting too large. Their method uses special parts that handle specific sensors before mixing the information together, allowing it to learn from a lot of satellite images. They tested MEOX on many tasks related to land and crop detection and found it performs better than some previous models, all while staying small and efficient. This shows it’s possible to build smart Earth observation tools that work well even with limited computer resources.
Earth observationmultimodal learningmasked autoencodermixture-of-expertssensor fusiontransfer learningrotary attentionparameter efficiencysegmentationremote sensing
Authors
Mohanad Albughdadi
Abstract
Recent advances in Earth Observation representation learning accommodate heterogeneous sensors and missing observations, often through larger architectures. We present MEOX (Multimodal Earth Observation with eXperts), a multimodal masked autoencoder with a 2.939 million-parameter encoder and 3.115 million parameters in total. Sensor-specific adapters, explicit validity signals, and a shared sparse-expert block preserve modality-dependent processing before a learned patch-wise fusion. Four metadata tokens then accompany a single spatial sequence through fourteen further encoder blocks. Shared expert projections with private low-rank residuals constrain parameter growth, while rotary attention supports downstream spatial grids different from pretraining. The model is pretrained on 1.228 million MMEarth64 samples using modality-balanced masked reconstruction and structured sensor dropout. Frozen transfer is evaluated on six GEO-Bench tasks at both 64 and 224 pixels. The model reaches 64.42% mean intersection-over-union on cashew segmentation at 64 pixels and 90.56% average accuracy on EuroSAT at 224 pixels, exceeding the corresponding reported CSMoE results. BigEarthNet finetuning reaches 72.95% micro-average precision. Routing diagnostics distinguish expert participation, spatial dependence, modality association, and functional contribution. A held-out WorldCover probe measures a 0.64-percentage-point benefit from metadata, while retrieval separates same-sensor semantics from cross-sensor alignment. These results demonstrate sensor-flexible representation learning and strong task transfer using a compact parameter budget.