A quality-aware GeoAI pipeline for monthly Landsat NDVI forecasting: benchmarking LSTM against persistence in metropolitan San Luis Potosí
Vegetation-index time series derived from Earth observation provide a consistent basis for monitoring environmental dynamics, but their predictive use is constrained by clouds, incomplete temporal coverage, spatial aggregation choices, and strong short-term persistence. Moreover, increasingly complex deep-learning models are not necessarily useful unless they are evaluated against competitive baselines and under temporally honest validation schemes. This study develops and critically evaluates a reproducible GeoAI pipeline for one-month-ahead forecasting of the Normalized Difference Vegetation Index (NDVI) at grid-cell level over metropolitan San Luis Potosí, Mexico. The objective was to determine whether a Long Short-Term Memory (LSTM) network can extract predictive information beyond the most recent NDVI observation and to identify when and where such added value occurs.
Monthly Landsat Collection 2 Level-2 surface-reflectance observations covering January 2017 through December 2025 were processed using the red and near-infrared bands together with QA_PIXEL information for quality filtering. Monthly NDVI surfaces were aggregated to a regular 1 km × 1 km grid, and each cell was assigned a persistent cell_id linking raster observations, tabular records, temporal predictors, predictions, and spatial diagnostics. Cell-month observations with less than 70% valid-NDVI coverage were excluded. The reference predictor set comprised current mean NDVI, three autoregressive lags, three- and six-month rolling means, and cyclical month encodings. Six consecutive monthly predictor vectors were used to predict NDVI one month ahead, producing 2,335 valid sequences with eight predictors per time step.
To prevent temporal leakage, samples were partitioned chronologically rather than randomly. In the fixed reference experiment, 2,056 sequences were used for training, 149 for validation with target date November 2025, and 130 for final testing with target date December 2025. Continuous predictors were scaled using parameters estimated exclusively from the training partition. The reference model consisted of an LSTM layer with 32 units, 20% dropout, a 16-unit ReLU dense layer, and a linear output neuron. Training used mean squared error, the Adam optimizer, early stopping, and adaptive learning-rate reduction. Persistence, defined as carrying the latest available cell-level mean NDVI forward one month, served as the primary benchmark. Complementary experiments evaluated a multilayer perceptron and an LSTM augmented with monthly CHIRPS precipitation. Performance was assessed using mean absolute error (MAE), root mean squared error (RMSE), and the coefficient of determination (R²), together with independent target-month evaluations, seven balanced two-month temporal test windows, and cell-level maps of absolute-error differences.
In the fixed December 2025 test split, persistence achieved MAE = 0.008298, RMSE = 0.009362, and R² = 0.907587. The NDVI-history LSTM achieved MAE = 0.008072, RMSE = 0.009236, and R² = 0.910054. Although the aggregate improvement was modest, it was obtained against a strong baseline for a highly autocorrelated monthly variable. Across independent target-month evaluations, the LSTM performed better in five of seven months, while persistence was superior in August and September. In the balanced temporal-window analysis, the LSTM outperformed persistence in six of seven two-month test windows, with August–September as the only interval favoring persistence. Spatial diagnostics further showed that better aggregate metrics did not imply spatial dominance: in the December reference split, persistence produced lower absolute error in 75 of 130 evaluated cells and the LSTM in 55, indicating that the LSTM's global advantage resulted from larger error reductions in a smaller subset of locations.
Greater model complexity did not automatically improve forecasting. The precipitation-enhanced LSTM produced MAE = 0.010943, RMSE = 0.012560, and R² = 0.833682, whereas the multivariate MLP yielded MAE = 0.029637, RMSE = 0.032344, and R² = −0.102948. These results do not imply that precipitation is physically irrelevant to vegetation dynamics; rather, the specific monthly CHIRPS representation tested here did not provide sufficient additional predictive information beyond recent NDVI history.
The principal contribution is therefore methodological rather than architectural. The study integrates quality-aware Landsat preprocessing, spatially consistent aggregation, leakage-aware chronological sequence construction, benchmarking against a demanding persistence reference, temporal robustness analysis, and spatially explicit error diagnostics within a single GeoAI workflow. The results demonstrate that sequential learning can provide measurable predictive skill beyond persistence, but that this advantage is temporally and spatially conditional. They also show that additional covariates or architectural complexity require explicit empirical justification. Because the current model does not explicitly represent inter-cell adjacency, future work will investigate spatial-dependence models, multiple training seeds, improved treatment of temporal gaps, and land-cover-dependent environmental covariates.