Soil texture prediction using optical and radar remote sensing data in a random forest approach

Herausgeber: 8th International Workshop on Retrieval of Bio- & Geo-physical Parameters from SAR Data for Land Applications
Urheber: European Space Agency (ESA)
Jahr: 2023

Einordung:
Institut: Professur Geodäsie und Geoinformatik

Abstract:
The application of precision agriculture (PA) helps reduce ecosystem damage from agricultural emissions without compromising food security. One problem farmers face in implementing PA applications is the lack of high spatial resolution soil information.
For this study, the research project "pH-BB: Precision Liming in Brandenburg" provided data from 1091 soil samples collected and analyzed in the laboratory for clay (C), silt (U), and sand (S) textures from 3 farms in Brandenburg, Germany. Google Earth Engine was used to process earth observation data (EO) from 415 Sentinel-1 (S1) SAR scenes of one orbit and 41 cloud-free Sentinel-2 (S2) scenes that fully covered the study areas from March 01, 2016 - December 14, 2022. Vegetation and ground indices were calculated using the S2 optical data; radar backscatter in VV and VH polarization was extracted from the S1 data. Long-term patterns in the EO data were identified using statistical parameters ( coefficient of variation, standard deviation, mean and maximum pixel values) along the temporal axis. Together with the calculated terrain attributes, 61 covariate grids were finally available for model building. The reference samples (RS) were randomly divided into a training dataset (70%,) and a validation dataset (30%). A random forest machine learning algorithm was used to train two individual models for the alr-transformed target variables C and U. The developed models were then applied to the grids of covariates to predict the alr-transformed target variables. The final maps (2053 km2 at 10*10m resolution) for all 3 texture fractions were calculated by back transformation. In the derived models, the EO data had the greatest importance. In the validation, the spatial distribution of the C, U, and S fractions could be predicted with an RMSE of 6.8, 7.2, and 11.3 mass %, respectively. Errors were lower in the sand-dominated soil classes, while they increased in the clay-dominated soil classes.
With a mean error of 7-11 mass-%, the approach shows good potential for evaluating topsoil texture of agricultural land at high spatial resolution (10m), for global applications or as preliminary information for high-resolution soil geophysical mapping.

Ansprechpartner

Weitere Information im WWW

Zur Publikation: Mitarbeiter, Projekte
Weitere Publikationen im Forschungsteilschwerpunkt