- author = "Joao Eduardo Silva Pombinho Batista",
- title = "Evolving robust and readable land cover models for improved forest monitoring",
- school = "Departamento de Informatica, Faculty of Science, Universidade de Lisboa",
- year = "2024",
- address = "Portugal",
- keywords = "genetic algorithms, genetic programming, Evolutionary Computation, Machine Learning, Interpretability, Environmental Science, Remote Sensing, MOGP, XAI",
-
URL = "
https://biblius.ciencias.ulisboa.pt/publications/54376",
-
URL = "https://researchportal.ulisboa.pt/en/studentTheses/evolving-robust-and-readable-land-cover-models-for-improved-fores/",
-
URL = "http://hdl.handle.net/10451/64825",
-
URL = "https://www.cienciavitae.pt/7B1C-1220-49FF",
-
URL = "https://repositorio.ulisboa.pt/entities/publication/79165236-fc8a-4fc2-8b4b-8958c796205c",
- size = "187 pages",
- abstract = "The importance of forests worldwide is well established throughout society, and it is well known that deforestation has serious implications for biodiversity, for rural communities dependent on forests for food and income, and for greenhouse gas emissions driving the global climate. Mechanisms such as REDD+, developed by the UNFCCC, help developing countries by rewarding them for avoiding deforestation and forest degradation while promoting forest conservation, sustainable management and enhancement of forest carbon stocking. However, for those mechanisms to be effective, they require the ability to monitor forest development efficiently. This work contributes to the scientific fields of computer science and environmental science by applying evolutionary machine learning to tackle this issue, addressing two important objectives in forest monitoring: mapping land cover changes and detecting forest degradation. It also contributes to the machine learning field on topics such as automatic feature construction methods for creating robust and interpretable models that can be read, understood and corrected by the experts of the application fields. Several of our applications successfully increase the robustness and transferability of machine learning models, improving the reliability of forest monitoring models. The induced models can be validated by identifying whether the features being used make sense in the context of the problem, and also by reducing the dimensionality of the datasets and visually identifying issues with the data. We also study the relationship between the complexity of feature engineering models and their interpretability. Model interpretability and explainability have become hot topics over the last few years. This motivates future studies on using evolutionary machine learning to force the evolution of more easily interpretable models via specific fitness functions.",
-
notes = "in english
Supervisor: Sara Guilherme Oliveira da Silva Vanneschi",
