Vegetation masking of remote sensing data aids machine learning for soil fertility prediction

Authors: WINZELER, HANS - University Of Texas At Arlington; MANCINI, MARCELO - University Of Arkansas; Blackstock, Joshua; Libohova, Zamir; Owens, Phillip; Ashworth, Amanda; MILLER, DAVID - University Of Arkansas; SILVA, H - Federal University Of Lavras

Submitted to: Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/31/2024
Publication Date: 9/5/2024
Citation: Winzeler, H.E., Mancini, M., Blackstock, J.M., Libohova, Z., Owens, P.R., Ashworth, A.J., Miller, D., Silva, H.G. 2024. Vegetation masking of remote sensing data aids machine learning for soil fertility prediction. Remote Sensing. https://doi.org/10.3390/rs16173297.
DOI: https://doi.org/10.3390/rs16173297

Interpretive Summary: Soil fertility is important for crops but varies within farm fields and changes over time. Detailed soil fertility maps are desirable for effective soil management and precision agriculture applications but are challenging to produce due to high spatial and temporal variability. Detailed maps of soil fertility can be produced from intensive and high-density soil sampling. However, collecting soil samples in the field and analyzing them is time consuming and expensive. Data collected from Sentinel-2 (S2) satellite missions with high spatial and temporal resolution and combined with field sampling were utilized to create detailed maps of soil fertility for an intensively managed farm field. A routine was developed in the Google-Earth-Engine (GEE) platform to remove the vegetation from the satellite images that allows for using bare soil images to improve the soil fertility maps and determine the appropriate number of soil samples. The newly developed method would save farmers funds in field sampling and management by supporting precision management practices while saving the environment.

Technical Abstract: Soil nutrient content varies spatially across agricultural fields in hard-to-predict ways, particularly in floodplains with complex fluvial depositional history. Satellite-reflectance-data from Sentinel-2 (S2) mission provides spatially continuous land-reflectance data that can aid model development when used with point observations of nutrients. Reflectance from vegetation is assumed to obstruct land-reflectance of bare soil, such that researchers have masked vegetation in models. We devel-oped a routine for masking vegetation within Google-Earth-Engine (GEE) using Random Forest classification for iterative application to libraries of S2-images. Using gradient boosting we then developed soil nutrient models for surface soils at a 250-ha agricultural site using S2 images. Soils were sampled at 2,145 point-locations to 23-cm depth and analyzed for Ca, K, Mg, P, pH, S, Zn. Results showed that masking vegetation improved model performance for models from subsets of the data (80% of samples used for model development, 20% validation), but that full sets of data did not require masking to achieve accuracy. Models of Ca, K, Mg, and S were successful (validation R2 > 0.60 to 0.96), but models for pH, P, and Zn failed. Bare soil composite images from S2-data are helpful in predicting some soil fertility characteristics in low-relief floodplains.