Evaluating explainable artificial intelligence methods for multi-label deep learning classification tasks in remote sensing

No single XAI method is perfect: The study tested ten explainable AI methods and found that while Occlusion, Grad-CAM, and LIME were most reliable, each had significant trade-offs—particularly that none produced high-resolution outputs, and two of the three top performers were computationally expensive.

XAI can reveal hidden dataset problems: By applying explainability methods to deep learning models trained on remote sensing data, researchers were able to uncover biases and shortcomings in datasets that would otherwise remain hidden behind the "black box" of neural networks.

Explainability performance varies by situation: The effectiveness of XAI methods differed depending on the context—straightforward predictions, competing multiple labels, and misclassification cases each presented different challenges, suggesting that organizations may need multiple tools rather than relying on one method.

Transparency requires practical trade-offs: The findings highlight a key challenge in responsible AI: the most interpretable methods often come with computational costs or technical limitations (like low-resolution outputs), meaning organizations must balance the desire for transparency against practical constraints.

XAI is essential for high-stakes applications: In remote sensing—where AI decisions can affect land use planning, environmental monitoring, and resource management—the ability to understand and verify model decisions through explainability methods is critical for building trust and accountability.