Building consistency in explanations: Harmonizing CNN attributions for satellite-based land cover classification

Explainable machine learning has gained substantial attention for its role in enhancing transparency and trust in computer vision applications. Attribution methods like Grad-CAM and occlusion sensitivity analysis are frequently used to identify how features contribute to predictions of neural networks. However, a key challenge is that different attribution methods often produce different outcomes undermining trust in their results. Furthermore, the unique characteristics of remote sensing imagery pose additional challenges for attribution interpretation: it primarily comprises continuous “stuff” classes rather than objects, exhibits fine-grained spatial variability, contains mixed pixels, is often multispectral, and exhibits spatially heterogeneity. To tackle this challenge, we present a novel methodology that harmonizes attributions, resulting in: 1. greater consistency across different attribution methods; 2. more meaningful explanations when validated against known segmentation ground truth; and 3. enhanced transparency and traceability. This is achieved by coherently linking feature representations to attributions derived from analyzing the training data, enabling direct attribution assignment to features in (unseen) images. We evaluate our methodology using two satellite-based land cover classification datasets, three convolutional neural network architectures, and nine attribution methods. Harmonizing attributions increases the Pearson correlation coefficient between different attribution methods by an average of 0.18 across all datasets, models, and methods; and improves the micro F1-score — a measure of accuracy — by 12%. We demonstrate that Grad-CAM attributions are inherently well-aligned with the features, whereas other gradient-based attribution methods exhibit significant noise, mitigated through harmonization. It further enhances the resolution of occlusion-based attribution maps and adjusts misleading explanations. • We link feature representations to training attributions. • Greater consistency across different attribution methods. • More meaningful explanations. • Enhanced transparency and traceability.