Interoperability in Earth Observation deep learning: A cross-language and cross-framework evaluation

The Open Neural Network Exchange (ONNX) is widely promoted as a framework-independent solution for deep learning model portability, but systematic evidence of its reliability across languages, frameworks, and deployment environments remains limited for Earth Observation (EO) workflows. This paper introduces an eight-step evaluation protocol for assessing ONNX inference portability of EO deep learning models, demonstrated through two case studies under documented software conditions. For a TempCNN crop-type classifier, native-to-ONNX deviation was within 1 0 − 6 RMSE for both PyTorch and Flux exports, and cross-runtime outputs across Python, R, and Julia were bitwise-identical under the CUDAExecutionProvider . For a ConvLSTM precipitation nowcasting model, three conversion pathways were tested: the Keras 3-on-PyTorch pathway (KPT) passed all eight protocol steps, including a directly computed RMSE port = 2.81 × 1 0 − 7 ; the other two pathways required workarounds or were blocked from native re-import. ONNX portability is architecture-dependent in the tested configurations. Static convolutional models built from common operators port cleanly, while recurrent models with explicit temporal loops expose distinct failure modes that require toolchain-specific validation. The protocol’s demonstrated scope covers these two architectures and software conditions; its behavior on other model families should be confirmed through fresh application.