Uncovering Long-Term Agricultural Landscape Change Through Customized Segmentation of Historic and Modern Remote Sensing Imagery

Understanding long-term changes in agricultural landscape structure is essential for assessing biodiversity patterns, ecosystem condition and the impacts of land-use policy. Landscape metrics such as field size, edge length and spatial connectedness serve as important biodiversity-relevant indicators in cropland mosaics, yet their derivation over large areas and long time periods remains challenging as it requires consistent high-resolution remote sensing data and a robust method for the accurate delineation of field patches.

Here we exploit the strengths of the Segment Anything Model (SAM) for the automatic segmentation of farmland parcels using two distinct sources of remote sensing imagery: historic panchromatic CORONA data from 1965 and modern digital orthophotos from 2021/2022. Our study focuses on the German state of Saxony, which experienced severe transitions in agricultural landscape structure throughout the 1960s due to substantial policy changes, such as collectivization, with major consequences for field size and associated farmland biodiversity. We selected several study regions across the state for which scenes from both data sources were pre-processed, harmonized and masked. We then employed a two-step SAM algorithm combined with customized post-processing steps and evaluated segmentation performance using the Intersection over Union (IoU) between predicted patches and user-derived validation patches. Input parameters such as compactness and minimum size were tuned to favor “field-shaped” segments. Across most test regions, the method achieved high accuracy with median IoU values of around 0.7 for historic and over 0.9 for modern images, with some limitations in mountainous areas and those with low image quality. Based on these segmentation results we then calculated landscape metrics and evaluated their long-term changes, confirming and quantifying the substantial regional increase in field size in Saxony.

The developed approach holds considerable potential for long-term biodiversity monitoring frameworks for agricultural landscapes, particularly where historic imagery is available, by providing a path to obtaining consistent time series of landscape-structure indicators aligned with essential biodiversity variables. As a next step, we plan to scale the method to a larger area and more time steps, enabling a more holistic examination of agricultural landscape change and ultimately supporting future conservation planning in line with emerging global targets.