Location: Cropping Systems and Water Quality Research
Title: Vis-NIR spectroscopy for on-the-go soil organic matter estimation in agricultural fieldsAuthor
![]() |
Ransom, Curtis |
![]() |
CONWAY, LANCE - John Deere & Company |
![]() |
Sudduth, Kenneth |
Submitted to: Meeting Proceedings
Publication Type: Proceedings Publication Acceptance Date: 10/14/2024 Publication Date: 10/14/2024 Citation: Ransom, C.J., Conway, L., Sudduth, K.A. 2024. Vis-NIR spectroscopy for on-the-go soil organic matter estimation in agricultural fields. Proceedings of The Sixth Global Proximal Soil Sensing Workshop, October 14-17, 2024. Ghent, Belgium. p. 33-38. Interpretive Summary: The research evaluates the effectiveness of deep learning and ensemble models in estimating soil organic matter using full-spectrum data. It compares on-the-go methods using discrete and full-spectrum approaches for estimating field-scale organic matter levels. Three sources of data were used, and the results showed that deep learning and ensemble approaches improved accuracy for estimating organic matter. However, these improvements did not translate well to in-field on-the-go estimations. Commercially available sensors using discrete wavebands performed better. However, additional steps need to be investigated to improve research based sensors to account for the noise introduced with on-the-go scanning of soils. Improving these sensors will increase farmer’s ability to measure and track surface soil organic matter, which will lead to improved management of inputs (e.g., fertiliser and pesticides) for site-specific applications. Technical Abstract: Optical proximal sensing technologies use visible and near-infrared spectroscopy to map soil properties. However, the accuracy of products available to producers is limited due to reliance on specific wavelengths. Previous studies have shown that using preprocessing and machine learning techniques on full-spectrum data improves accuracy for estimating soil organic matter. This research aims to evaluate the effectiveness of deep learning and ensemble models in estimating soil organic matter using full-spectrum data. It also aims to compare on-the-go methods using discrete and full-spectrum approaches for estimating field-scale organic matter levels. Three sources of data were used: Missouri and Illinois, USA surface soils (n=90) scanned using a bench-top spectrometer at two water content levels, additional Missouri and Illinois surface samples (n=15) scanned with a research spectrometer with five different water content levels, and in-situ (n=20) surface scans collected within a single field with a research spectrometer (full-spectrum) and a commercial sensor (discrete wavebands). Results showed that deep learning and ensemble approaches had similar or slightly improved accuracy for estimating organic matter over commonly used methods (e.g., SVM,PLS) (RMSE all between 0.51 to 0.83 g kg-1). However, deep learning and ensemble-based learning resulted in less overfitting on the training datasets. Still these slight improvements did not translate well to in-field on-the-go estimations (RMSE >1.24 g kg^-1). In contrast, while commercially available sensors using discrete wavebands overestimated organic matter levels, they performed better (RMSE = 1.04 g kg^-1). Results might be improved by taking into account the noise introduced with on-the-go scanning of soils resulting from poor soil contact. This could be done by automating the cleaning of outlier scans, and taking additional scans under varying in-field conditions (i.e., more moisture conditions, wider range of organic matter, or using unground samples). |