Laser-induced breakdown spectroscopy combined with a convolutional neural network: A promising methodology for geochemical sample identification in Tianwen-1 Mars mission

Transfer learning based on dynamic time warping algorithms to improve qualitative analysis and quantitative prediction of rocks over multiple LIBS instruments

“…With this in mind, the aim was to eliminate the effects of data fluctuations during the experiment. 33–35 In the end, there were 49 spectra recorded for 49 samples. Fig.…”

Atomic spectrometry update – a review of advances in environmental analysis

“…The development of the LIBS technique has benefitted greatly from its successful deployment in the SuperCam instrument on Mars. Compensation for spectral differences caused by varying distances between sample and sensor usually involve conventional spectral data processing but a new chemometrics model with powerful learning ability has been constructed 237 for this correction. The performance of the convolutional neural network designed in this project surpassed those of four alternative chemometric approaches, making it a promising methodology for geochemical sample identification in future space missions.…”

Rapid classification of copper concentrate by portable laser-induced breakdown spectroscopy combined with transfer learning and deep convolutional neural network

“…As a branch of machine learning, deep learning can automatically extract the features in one pass, which greatly simplifies the workflow of machine learning [17] . In addition, deep learning is considered adept at correcting the interference of multiple factors [18] . Therefore, it is currently attracting more and more researchers' interest.…”

“…Li et al achieved a high-accuracy classification of geological samples by employing a CNN with five convolutional layers and two pooling layers. Their spectra were collected by MarSCoDe during preflight testing [18] . Zhao et al developed a 1D-CNN model to classify iron ore, and for the first time interpreted the effectiveness of the CNN model by the t-distributed symmetric neighbor embedding algorithm (t-SNE) [19] .…”

“…In general, increasing the depth of the CNN model is an effective approach toward increasing the performance of the model [20] . At present, most of the existing studies on CNN and LIBS designed their own convolutional structures with a small number of convolutional layers (typically 2-5 layers, due to the limited sample number), which are unable to compare to the deep CNN (typically more than 100 layers, trained with millions of data) [17][18][19] . Moreover, the design of the CNN model structure is usually based on experience and is time-consuming and laborious.…”