Researchers from Zhejiang University have developed a new non-linear memory-based learning (N-MBL) model that enhances the prediction accuracy of soil properties using visible near-infrared (vis-NIR) spectroscopy. By comparing N-MBL with traditional machine learning and local modeling methods, the study reveals its superior performance, particularly in predicting soil organic matter and total nitrogen.
In a world where over 700 million people suffer from hunger, according to the Food and Agriculture Organization (FAO), the need for efficient agricultural productivity and soil quality monitoring has never been more urgent. Traditional methods of analyzing soil properties through laboratory testing are often expensive and time-consuming, making them impractical for large-scale applications. Visible near-infrared (vis-NIR) spectroscopy has emerged as a rapid and non-destructive alternative, offering the potential to revolutionize soil analysis (1,2). However, to fully harness this technology, robust and accurate predictive models are essential. In a new study, researchers from Zhejiang University in China have developed a non-linear memory-based learning (N-MBL) model that significantly improves the prediction of soil properties from vis-NIR spectral data (1).
Read More: NIR in Soil Analysis
Details of the Research and Findings:
Visible near-infrared (vis-NIR) spectroscopy measures the reflected electromagnetic spectrum from soil materials within a wavelength range of 350 to 2500 nm. This technique has gained widespread recognition for its ability to predict soil properties quickly and without the need for destructive sampling. However, the accuracy of these predictions relies heavily on the models used to interpret the spectral data. Traditionally, linear models such as partial least squares regression (PLSR) have been employed, but these models often fall short in capturing the complex, non-linear relationships between soil properties and spectral data (1).
Memory-based learning (MBL) has emerged as a powerful local modeling technique, particularly when applied to large soil spectral libraries. MBL models rely on forming associations between new soil samples and similar patterns found in existing spectral libraries. This approach allows MBL to effectively predict various soil properties based on local relationships within the data. However, the conventional MBL models typically use linear approaches, which may not fully capture the intricate, non-linear interactions between soil properties and vis-NIR spectra (1).
To address this limitation, the research team at Zhejiang University, led by Zheng Wang, Songchao Chen, Rui Lu, Xianglin Zhang, Yuxin Ma, and Zhou Shi, developed a non-linear version of MBL, termed N-MBL. This new algorithm was evaluated using the Lateritic Red soil spectral library (LRSSL) from Guangdong province in China. The LRSSL comprises 742 samples, with detailed information on soil properties such as pH, soil organic matter (SOM), total nitrogen (TN), total phosphorus (TP), and total potassium (TK) (1).
The study compared the performance of N-MBL with several established machine learning (ML) models, including PLSR, cubist, random forest (RF), support vector machine (SVM), and convolutional neural network (CNN). Additionally, the N-MBL was tested against the traditional linear MBL model. The results were compelling: local models, including both MBL and N-MBL, generally outperformed the more generalized ML models, especially when applied to large spectral libraries (1).
One of the key findings was the stability and accuracy of N-MBL in predicting certain soil properties as the number of selected nearest neighbors (k) varied. As variable k increased, the performance of MBL showed significant fluctuations, whereas N-MBL demonstrated a more consistent improvement. For instance, the R² values for SOM and TN predictions using N-MBL surpassed those of MBL when k exceeded 70 and 90, respectively. However, N-MBL showed lower performance for pH and TK prediction compared to MBL when k was above certain thresholds. Notably, N-MBL outperformed MBL in predicting TP across a wide range of k values (1).
Conclusion:
The development of the N-MBL model marks a significant advancement in soil property prediction using vis-NIR spectroscopy. By integrating non-linear modeling techniques into the MBL framework, researchers have successfully enhanced the predictive capabilities of local models, offering a more accurate and reliable method for soil analysis. This has the potential to improve soil monitoring and management practices, ultimately contributing to better agricultural productivity and food security. The study not only highlights the importance of non-linear approaches in spectral modeling but also paves the way for future innovations in soil science (1).
References
(1) Wang, Z.; Chen, S.; Lu, R.; Zhang, X.; Ma, Y.; Shi, Z. Non-linear memory-based learning for predicting soil properties using a regional vis-NIR spectral library. Geoderma 2024, 441, 116752. DOI: 10.1016/j.geoderma.2023.116752
(2) Piccini, C.; Metzger, K.; Debaene, G.; Stenberg, B.; Götzinger, S.; Borůvka, L.; Sandén, T.; Bragazza, L.; and Liebisch, F. In‐field soil spectroscopy in Vis–NIR range for fast and reliable soil analysis: A review. Eur. J. Soil Sci. 2024, 75 (2), e13481. https://doi.org/10.1111/ejss.13481
Machine Learning Used for Meteorite Classification to Unlock Asteroid Composition Mysteries
September 24th 2024A team of researchers has developed a new machine learning (ML) method to classify asteroid spectra by analyzing meteorite spectroscopic data. Using logistic regression, the model accurately grouped meteorites into eight categories, helping to better understand the distribution of asteroid compositions in the asteroid belt. The study, published in Icarus, opens new avenues for predicting asteroid composition using spectroscopy.
OSIRIS-APEX: Unveiling the Secrets of Asteroid Apophis Flyby Post-Earth Encounter
September 23rd 2024NASA’s OSIRIS-REx spacecraft embarks on a new mission—OSIRIS-APEX—to study asteroid Apophis after its anticipated close flyby of Earth on Friday, April 13, 2029. Utilizing advanced spectroscopy and other instruments, the mission will provide unprecedented insights into how planetary encounters reshape small bodies in our solar system.