Abstract
B21L-01 Earth Observation and Machine Learning to Improve the Estimation of Changing Land Cover and Management in Soil Erosion Assessment presented at the American Geophysical Union Annual Meeting '23 (AGU23)
Soil erosion has been identified as the greatest threat to ecosystems and humans in several regions, leading to loss of soil fertility, ecosystem services, and safety concerns. Rainfall-induced soil erosion is a complex process due to the combined effect of precipitation features, soil properties, and surface conditions. Therefore, a reliable estimation of changes in land cover of natural lands (e.g., deforestation, fire) and management of arable lands (e.g., tillage and crop rotation) is crucial for its assessment.
The land cover and management factor (i.e., the C-factor in the well-established Revised Universal Soil Loss Equation, RUSLE) depends on vegetation cover, species composition, and farming practices. Usually, its calculation relies on lookup tables relating its value to land cover/use maps. Yet, this method neglects intra-annual cover variations due to vegetation dynamics (e.g., phenology) and crop rotation. Recently, dynamic estimation has been proposed by regressing relationships between the C-factor and Earth Observation (EO) data. Nevertheless, the coarse spatial resolution of EO sensors, 100 m up to 1 km for most of the used satellite sensors (e.g., PROBA-V, Sentinel-3 OLCI, and MODIS), undermines the reliability of these relationships.
This work aims to improve C-factor estimation by combining information at 10 m of ESA Sentinel-2 data with machine learning (ML) and accounting for the vegetation seasonal dynamics. The proposed method (i) uses a self-supervised classification of EO images to determine land cover, (ii) regresses a new relationship between spectral properties at the pixel scale and static C-factor values, and (iii) dynamically corrects C-factor values by analyzing spectral variations in arable lands and applying ML probabilistic semi-automatic recognition of cover changes in non-arable lands. The use of high-resolution EO data provides a massive number of points, increasing the robustness of the found relationships. Moreover, combining EO and ML allows for detaching from corrective sub-factors while enabling dynamic monitoring. Since the C-factor is the most manageable soil-erosion factor, such monitoring has the potential to support decision-making and policy implementation in the perspective of slowing soil erosion, especially in the most vulnerable areas.