Evaluating Explainable Artificial Intelligence Methods for Multi-label Deep Learning Classification Tasks in Remote Sensing

Ioannis Kakogeorgiou, Konstantinos Karantzalos

Abstract

Although deep neural networks hold the state-of-the-art in several remote sensing tasks, their black-box operation hinders the understanding of their decisions, concealing any bias and other shortcomings in datasets and model performance. To this end, we have applied explainable artificial intelligence (XAI) methods in remote sensing multi-label classification tasks towards producing human-interpretable explanations and improve transparency. In particular, we utilized and trained deep learning models with state-of-the-art performance in the benchmark BigEarthNet and SEN12MS datasets. Ten XAI methods were employed towards understanding and interpreting models' predictions, along with quantitative metrics to assess and compare their performance. Numerous experiments were performed to assess the overall performance of XAI methods for straightforward prediction cases, competing multiple labels, as well as misclassification cases. According to our findings, Occlusion, Grad-CAM and Lime were the most interpretable and reliable XAI methods. However, none delivers high-resolution outputs, while apart from Grad-CAM, both Lime and Occlusion are computationally expensive. We also highlight different aspects of XAI performance and elaborate with insights on black-box decisions in order to improve transparency, understand their behavior and reveal, as well, datasets' particularities.

Introduction

Deep learning has revolutionized remote sensing, enabling the mapping of complex land cover from satellite imagery with unprecedented accuracy. However, these models typically operate as "black boxes," providing predictions without explaining the underlying reasoning. In Earth observation, where decisions can impact environmental policy and disaster response, this lack of transparency is a significant hurdle. Researchers must verify if a model is identifying relevant features, such as vegetation patterns, or merely picking up on sensor artifacts or cloud shadows. This challenge is particularly acute in multi-label classification, where a single image patch may contain overlapping classes like urban areas, rivers, and forests.

Methodology

This study evaluates ten XAI methods across three categories:

  1. Gradient-based: (e.g., Saliency, Integrated Gradients) Analyzing how pixel-level changes affect model output.
  2. Perturbation-based: (e.g., Occlusion, Lime) Systematically blocking image regions to observe changes in prediction confidence.
  3. CAM-based: (e.g., Grad-CAM) Utilizing internal feature maps to generate importance heatmaps.

To provide an objective comparison, the researchers employed four quantitative metrics: Max-Sensitivity (stability against noise), AUC-MoRF (faithfulness to the model's logic), File Size (conciseness), and Computation Time (efficiency).

Model Performance and Datasets

The evaluation was conducted using two massive benchmark datasets: BigEarthNet (590,000+ Sentinel-2 patches) and SEN12MS (180,000+ multi-sensor triples). State-of-the-art architectures, including DenseNet-121 and ResNet-50, were trained to provide a robust baseline for the XAI tests.

Table 1: Performance comparison of various CNN architectures on the BigEarthNet dataset. DenseNet121 achieves the highest F1-score of 82.22%.

As shown in Table 1, DenseNet-121 emerged as the top-performing model on the BigEarthNet dataset, achieving an F1-score of 82.22%. This high baseline ensures that the XAI methods are interpreting a highly optimized and accurate decision-making process.

XAI Evaluation Results

The quantitative analysis revealed a clear hierarchy among the ten XAI methods. Occlusion, Grad-CAM, and Lime were identified as the most reliable and interpretable methods, consistently achieving the lowest Max-Sensitivity scores (often under 0.20) and superior AUC-MoRF results.

Table 2: Quantitative comparison of XAI methods across two datasets. Grad-CAM, Occlusion, and Lime emerge as top performers in terms of sensitivity and faithfulness.

While these three methods are the most faithful to the model's logic, they differ significantly in efficiency. As detailed in Table 2, Grad-CAM is the clear winner for practical, large-scale applications, requiring only 0.03 seconds to generate an explanation. In contrast, perturbation-based methods like Occlusion and Lime, while highly accurate, are computationally expensive, taking up to 30 seconds per image on larger datasets.

Key Insights

  • Reliability vs. Resolution: High-resolution gradient methods often act as simple edge detectors and are prone to noise. Low-resolution methods like Grad-CAM provide coarser "blobs" of importance but are significantly more robust and class-discriminative.
  • The Role of SmoothGrad: Averaging explanations over multiple noisy versions of an input (SmoothGrad) consistently improved the stability of gradient-based techniques.
  • XAI for Dataset Debugging: XAI revealed systematic labeling errors in the datasets. For instance, models often failed to predict "Beaches" because those regions were frequently mislabeled as "Marine Water" in the ground truth.

Conclusion

For remote sensing practitioners, Grad-CAM offers the best balance of reliability and computational speed. While Occlusion and Lime provide high-quality explanations, their computational cost limits their use in real-time or global-scale pipelines. Future work should focus on developing methods that combine the reliability of CAM-based approaches with higher spatial resolution to provide pixel-perfect transparency.

Summary updated

Abstract

Although deep neural networks hold the state-of-the-art in several remote sensing tasks, their black-box operation hinders the understanding of their decisions, concealing any bias and other shortcomings in datasets and model performance. To this end, we have applied explainable artificial intelligence (XAI) methods in remote sensing multi-label classification tasks towards producing human-interpretable explanations and improve transparency. In particular, we utilized and trained deep learning models with state-of-the-art performance in the benchmark BigEarthNet and SEN12MS datasets. Ten XAI methods were employed towards understanding and interpreting models' predictions, along with quantitative metrics to assess and compare their performance. Numerous experiments were performed to assess the overall performance of XAI methods for straightforward prediction cases, competing multiple labels, as well as misclassification cases. According to our findings, Occlusion, Grad-CAM and Lime were the most interpretable and reliable XAI methods. However, none delivers high-resolution outputs, while apart from Grad-CAM, both Lime and Occlusion are computationally expensive. We also highlight different aspects of XAI performance and elaborate with insights on black-box decisions in order to improve transparency, understand their behavior and reveal, as well, datasets' particularities.
BibTeX
@article{kakogeorgiou2021evaluating,
  author = {Ioannis Kakogeorgiou and Konstantinos Karantzalos},
  title = {{Evaluating Explainable Artificial Intelligence Methods for Multi-label Deep Learning Classification Tasks in Remote Sensing}},
  journal = {Int. J. Appl. Earth Obs. Geoinformation},
  year = {2021},
  doi = {10.1016/j.jag.2021.102520},
}
Figure