Impact of Varied Conv-LSTM Model Parameters on Prediction Accuracy for NDSI-Based Salinity Index Analysis

In recent years, machine learning models have emerged as potent tools for prediction studies, particularly when dealing with sequential data. This research delves into the impact of Convolutional Long Short-Term Memory (Conv-LSTM) model parameters on the accuracy of predictions using the Normalized Difference Salinity Index (NDSI) time series. The Conv-LSTM model architecture constitutes an evolution of the Long Short-Term Memory (LSTM) model, engineered to capture both spatial and temporal dependencies in sequences. Unlike traditional LSTMs, Conv-LSTMs embrace the spatial structure of input data by coupling LSTM units with convolutional layers. This synergistic design allows Conv-LSTMs to extract and integrate features across both time and space, making them adept at analyzing complex time series data with spatial correlations, such as satellite images or environmental measurements. The NDSI is an indicator of salinity in water bodies and coastal regions, making it a valuable tool for environmental monitoring. Utilizing Landsat-8 data of Tangier city of Morocco collected between 2015 and 2022, the NDSI time series dataset was constructed, forming the foundation for evaluating prediction accuracy. In this study, Conv-LSTM model was configured with three pivotal parameters: i) the number of “filters” in the main layer, ii) the number of “neurons” in the fully connected layer, and iii) the number of training “epochs”. The NDSI time series data were employed to train and evaluate the model, with prediction accuracy assessed using the coefficient of determination (R²) metric. The results uncover substantial insights into the relationship between Conv-LSTM model parameters and prediction accuracy for NDSI analysis. When considering a high number of epochs (i.e., epochs=100), the prediction accuracy remained relatively consistent at 97% across varying values of filters and neurons. This suggests that rendering the number of epochs beyond a certain point less influential on accuracy improvements. In the context of medium number of epochs (i.e., epochs=50), the observed accuracy variations were more pronounced. Notably, the accuracy was influenced by the number of filters in the main layer. Specifically, when filters numbered between 10 and 100, accuracy remained below 60%. However, with a rise in filter count, accuracy exhibited an upward trend, ultimately plateauing at 96%. In contrast, for a low epoch count (i.e., epochs=10), the initial accuracy was negative. However, this was addressed by introducing an extensive number of filters in the main layer, reaching up to 10,000. The infusion of this high filter count yielded positive accuracy outcomes reaching more than 60%, indicating that a substantial filter count compensated for the limited training epochs.