Confident Naturalness Explanation (CNE): A Framework to Explain and Assess Patterns Forming Naturalness

TL;DR

The study addresses explaining and quantifying naturalness in protected areas from satellite imagery, tackling the lack of quantitative, confidence-aware pattern metrics in prior work. It introduces the Confident Naturalness Explanation (CNE) framework, which combines a segmentation model with an interpretable logistic regression and Monte Carlo Dropout-based uncertainty to produce pattern-level contributions. A core contribution is a quantitative metric defined as $CNE_c = \frac{\alpha_{c^{+}}}{u_c}$, with $\alpha_{c^{+}} = \max(\alpha_c, 0)$ and $u_c = \sum_{h,w} S_c$, where $S_c = \sqrt{\frac{1}{J} \sum_{j=1}^{J} (Y_{c,j} - A_c)^2}$, summarizing the confident contribution of each pattern. Empirical results on Fennoscandia using the AnthroProtect and CORINE data show wetlands achieving CNE values in the range $[0.8,1]$, while glaciers, grasslands, and water bodies cluster around $0.2$, with uncertainty-aware segmentation masks accompanying the results. The framework provides a scalable, objective tool for protected-area monitoring and supports conservation decision-making.

Abstract

Protected natural areas are regions that have been minimally affected by human activities such as urbanization, agriculture, and other human interventions. To better understand and map the naturalness of these areas, machine learning models can be used to analyze satellite imagery. Specifically, explainable machine learning methods show promise in uncovering patterns that contribute to the concept of naturalness within these protected environments. Additionally, addressing the uncertainty inherent in machine learning models is crucial for a comprehensive understanding of this concept. However, existing approaches have limitations. They either fail to provide explanations that are both valid and objective or struggle to offer a quantitative metric that accurately measures the contribution of specific patterns to naturalness, along with the associated confidence. In this paper, we propose a novel framework called the Confident Naturalness Explanation (CNE) framework. This framework combines explainable machine learning and uncertainty quantification to assess and explain naturalness. We introduce a new quantitative metric that describes the confident contribution of patterns to the concept of naturalness. Furthermore, we generate an uncertainty-aware segmentation mask for each input sample, highlighting areas where the model lacks knowledge. To demonstrate the effectiveness of our framework, we apply it to a study site in Fennoscandia using two open-source satellite datasets.

Figures (3)

• Figure 1: An Illustration for (CNE) framework. In the explainability part, the input images are fed to the segmentation model, resulting in predicted segmentation masks; they are fed to logistic regression with ground truth labels of the input images to vectorize patterns and classify the input into naturalness and anthropogenic areas. After training the logistic regression, we use only the positive coefficients to calculate the CNE metric. In the uncertainty, the MC-Dropout resembles multiple sampled models used to quantify the uncertainty of each pattern in the input image. In the lower right corner, the knowledge gained from parts 1 and 2 is combined to calculate the CNE metric in part 3 and assign a quantifiable metric value to each pattern, reflecting its confident contribution to the concept of naturalness. The uncertainty part is shown in detail in \ref{['fig2:output']}
• Figure 2: Illustrative diagram for the tensor generated by the segmentation model. The input images are passed to the model, and the middlebox includes the tensor ${\hbox{\fontencoding{T1}\sffamily\slshape{Y}}}$ where $b$ is the image index at a single batch, and $j$ is the number of sampled models. On the right, we have two outputs. The upper image shows the output after taking the average over dimension $J$ and assigning each pixel to the class with the highest probability, and at the bottom, an uncertainty aware segmentation mask is generated by getting the standard deviation across MC runs where high-intensity pixels -white- represent high uncertainty and vice-versa.
• Figure 3: Qualitative results.Demonstrating RGB Sentinel 2 images, predicted segmentation masks, and uncertainty-aware segmentation maps for two examples. The greyscale bar indicates pixel uncertainty in the segmentation maps. Each color in the segmentation masks represents a different pattern