Implementing CNN and LSTM for Cotton Seed Quality Assessment Using FTIR Spectroscopy Data

Cotton seed quality is one of the major determinants of agricultural productivity, crop yield, and sustainability. The assessment of seed germination potential is essential; however, traditional methods for the same usually involve time-consuming and destructive techniques. FTIR spectroscopy has emerged as a non-destructive alternative that probes the biochemical composition of seeds. In general, such complex spectral data requires robust methods for analysis. It performs the implementation and comparison of two advanced deep learning models, CNNs and LSTMs, on cotton seed germination using FTIR spectroscopy data for prediction. For this study, 10 cotton seed samples were used, which were labeled as positive for germinating and negative for nongerminating seeds after the seeds were subjected to viability tests in the laboratory. After measurement, the recorded FTIR spectra were in the wavenumber range from 4000 to $800 \mathrm{~cm}^{-1}$, and preprocessing normalization and noise reduction of values were performed. The CNN, for feature extraction on spectral patterns, obtained an accuracy of 85%. On the other hand, the LSTM model, capturing sequential dependencies in spectral data, had performed better compared to CNN, obtaining an accuracy of 92%. Both models are trained and tested by using an 80 %-20 % split and performance is calculated with respect to accuracy and loss. The results highlight both the strengths of CNN in identifying local spectral features and the advantage of LSTM in modeling the inherent sequential relationship of spectral data. The higher performance of LSTM will indeed show that it can exploit the pattern present in the spectral data to predict germination potential. Future work is necessary to develop an improved version with more expanded datasets, integrating ensemble methods, and the use of transfer learning that would further improve model performance and generalization.