Coastal soil salinity is a major land-degradation concern, while field-based monitoring can be costly and spatially limited. This study evaluates a temporal deep-learning pipeline that forecasts six Landsat 8 surface-reflectance bands (B2--B7) at 500 fixed sampling locations in the Chirala coastal region of Andhra Pradesh, India, using annual observations from 2015 to 2025. A four-year input window is used to predict the next year's six-band vector. A residual Temporal Convolutional Network (TCN) is compared with LSTM and GRU baselines using chronological training, validation, and test periods. On the 2024--2025 test sequences, the TCN achieves a global $R^2$ of 0.5048, RMSE of 0.6624, and MAE of 0.5027 on standardized reflectance values. The trained model is recursively applied to forecast bands for 2026--2035; NDSI, NDVI, and nine additional study-specific spectral indices are then calculated from the forecasts and spatially interpolated. These outputs are spectral-indicator projections, not direct predictions of field-measured soil electrical conductivity (EC).

deep learning,temporal and spatial distribution,landsat,Time-series prediction,soil-salinity