Recent technological advances in machine learning, especially deep learning models, have transformed data analysis and prediction tasks remarkably. Models like Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNNs), and their hybrids, such as CNN-LSTM and Bidirectional LSTM (BiLSTM), have exhibited notable proficiency in handling time series data for regression tasks. In this study, BiLSTM model is employed to forecast Land Surface Temperature (LST) using time series data derived from MODIS satellite images between 2000 and 2024. The BiLSTM model's hyperparameters, including the number of neurons, epochs, and batch size, are fine-tuned to optimize its performance on MODIS LST dataset. To identify the optimal hyperparameters for accurate LST forecasting in different land cover types, we conduct examinations on four distinct areas in Morocco: forest, build-up, road-soil, and sand areas. By systematically testing different hyperparameter configurations, we determine the combination that yields the highest accuracy for each land cover type. These optimal hyperparameter values are then utilized to forecast LST for the respective regions. In conclusion, the findings of our study support our efficient approach to identifying optimal hyperparameters for BiLSTM model using LST data derived from MODIS satellite imagery. The determined configuration of 120 neurons, 100 epochs, and a batch size of 32 demonstrates superior performance in accurately forecasting LST variations across diverse land cover types.