Predicting Acid Sulphate Soils in Finland’s Coastal Areas Using Deep Learning Fusion of Remote Sensing Map Tile Data

Abstract

Acid Sulfate (AS) soils represent a significant ecological risk and are a major environmental concern, particularly in Finland, where they are recognized as one of the country’s most pressing environmental challenges. Traditional mapping methods for AS soils are labor-intensive and time-consuming, resulting in limited mapping efforts despite increasing awareness and efforts over the past decade. However, recent technological advancements have revolutionized soil data generation, providing new opportunities for AS soil classification through deep learning (DL) and machine learning (ML) techniques. This master’s thesis endeavors to evaluate various DL methods for AS soil classification across a substantial portion of the Finnish landscape, with a specific focus on identifying effective strategies for mapping AS soils. Notably, convolutional neural networks (CNNs) are comprehensively examined in this regard. The study area covers the north-western coast of Finland, with a dedicated dataset tailored for DL analysis. Meanwhile, ML techniques consider only central pixel values. Obtained results highlight that scalar-input models and Random Forest (RF) emerge as highly effective methods for AS soil classification, demonstrating promising improvements in accuracy and efficiency throughout the conducted studies.