Modified self‐adaptive model for improving the prediction accuracy of soil organic matter by laser‐induced breakdown spectroscopy

Abstract

Soil organic matter (SOM) significantly influences soil fertility, biology, and the global C cycle. Laser‐induced breakdown spectroscopy (LIBS) is attractive, as it can measure SOM accurately and quickly with minimal sample preparation and effective cost. Our earlier self‐adaptive (SA) model based on individual distances between LIBS soil spectra coupled with partial least squares (PLS) was favorable for SOM prediction. To optimize this model by including soil identification via the use of a hybrid distance between the spectra, two hybrid distance SA models to enhance SOM predictions from LIBS spectra in 250 Chinese topsoils were built and studied. The models used LIBS soil spectra and SOM content measured by the traditional potassium dichromate oxidation method for calibration. The performance of these models for the validation set was compared with the conventional no‐identification PLS model and SA‐PLS models. The modified models that used (a) a hybrid of Euclidean distance (ED) and angle cosine (SAxc–PLS), and (b) a hybrid of ED and the distance of response y (SAxy–PLS) improved SOM predictions , with a R2 of 0.908, a RMSE of 5.65 g kg−1, and a residual prediction deviation (RPD) of 2.91 in the validation set for SAxc–PLS; and a R2 of 0.900, a RMSE of 5.49 g kg−1, and a RPD of 2.99 in the validation set for SAxy–PLS. The improved performance mainly resulted from simultaneous consideration of diversity and correlation of the spectra. Therefore, modified SA models hold great potential for SOM prediction, especially for different soils.

Sensing of Soil Organic Matter Using Laser-Induced Breakdown Spectroscopy Coupled with Optimized Self-Adaptive Calibration Strategy

“…For example, the soil samples in Study Area 1 were mainly meadow solonchak soil, and the frequency statistics of the selected samples (Figure 7) suggested the samples in Study Area 1 were selected with higher probability for calibration than the unknown samples from the study, even if the SD and CV in Study Area 2 were larger than those in Study Area 1 (Table 1). In terms of larger sample selection for calibration at a national scale, 250 samples from all over the country were used with SAM-PLSR and LIBS for predicting SOM [22]; the results showed R 2 , RPD and RMSE values of 0.89, 2.82 and 5.84 g kg −1 , respectively (Figure 9), which showed lower model performance compared with that in Figure 8. For a national database, the SD and CV of samples were 16.77 and 63.39%, respectively, which were larger than those of the samples in Study Area 1.…”

“…Compared with the conventional PLSR method, the values of R 2 and RPD were increased by 1.60 and 1.96 times, respectively, and the RMSE value was reduced by almost 50%. The SAM-PLSR model combined with LIBS spectra was also used for predicting SOM, and the prediction accuracy was significantly improved compared with the conventional PLSR method [22]. Moreover, different spectral similarity calculation methods, such as correlation coefficients (CC), ED, Mahalanobis distance (MD), angle cosine (AC) and k-medoids (KM), have been used on various types of soil samples to determine the soil property prediction performance of the SAM-PLSR algorithm, and the SAM-PLSR based on ED and CC had the optimal model accuracy and robustness in the prediction of SOM; furthermore, a modified SAM-PLSR with LIBS spectra, based on individual distance and hybrid distance, was introduced [22], and the study also showed that the ED-based SAM-PLSR method showed good prediction performance.…”

“…Besides the algorithm in modelling, the selection of the calibration dataset played a very important role in the prediction model, which depended on the database for calibration [23,24]. However, the predicted results varied due to different calibration databases [21,22]. Calibration datasets selected from one soil type at the county scale [21] and different types at the national scale [22] were investigated.…”

“…However, the predicted results varied due to different calibration databases [21,22]. Calibration datasets selected from one soil type at the county scale [21] and different types at the national scale [22] were investigated. When the variability of the soil samples was too small, the selected calibration dataset was not representative, which led to lower prediction accuracy; however, when the soil variability was too large, the increasing disturbance reduced the prediction accuracy.…”

High-Throughput Recognition of Tumor Cells Using Label-Free Elemental Characteristics Based on Interpretable Deep Learning

“…Nevertheless, cell recognition with LIBS remains unreported, and simultaneous classification for four types of cancer cells is a challenge. Although classification models such as SVM, RF, , and SOM achieved excellent performance coupled with LIBS, the algorithm’s inner mechanism was hidden to the user (called the “black box”) and lacked interpretability and extensibility . Consequently, we adopted a deep learning model: a one-dimensional convolution neural network (1D-CNN) with gradient-weighted class activation mapping (Grad-CAM) to build a “transparent” model and provide interpretability of results …”