Interpretable land cover classification with modal decision trees

ABSTRACT

Land cover classification (LCC) refers to the task of classifying each pixel in satellite/aerial imagery by predicting a label carrying information about its nature. Despite the importance of having transparent, symbolic decision models, in the recent literature, LCC has been mainly approached with black-box functional models, that are able to leverage the spatial dimensions within the data. In this article, we argue that standard symbolic decision models can be extended to perform a form of spatial reasoning that is adequate for LCC. We propose a generalization of a classical decision tree learning model, based on replacing propositional logic with a modal spatial logic, and provide a CART-like learning algorithm for it. We evaluate its performance at five different LCC tasks, showing that this technique leads to classification models whose performances are superior to those of their propositional counterpart, and at least comparable with those of non-symbolic ones. Ultimately, we show that spatial decision trees and random forests are able to extract complex, but interpretable spatial patterns.

KEYWORDS:

• Interpretable machine learning
• modal logic
• decision tree learning
• hyperspectral image classification
• land use
• land cover

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available from the corresponding author, G. Pagliarini, upon reasonable request.

Notes

¹ https://github.com/aclai-lab/ModalDecisionTrees.jl.

² https://github.com/aclai-lab/Sole.jl.

Additional information

Funding

We acknowledge the support of the INDAM-GNCS project Symbolic and Numerical Analysis of Cyberphysical Systems (code CUP-E53C22001930001), funded by INDAM, and of the FIRD project Symbolic Geometric Learning, funded by the University of Ferrara.