Monthly precipitation prediction in mountainous terrain is challenging due to orographic forcing and complex atmospheric interactions. We present a systematic framework evaluating hybrid deep learning architectures for the Colombian Andes (Boyacá, 3,965 grid cells, 145-5490 m elevation, forecast horizons from 1 to 12 months). In a companion study, Convolutional Long Short-Term Memory (ConvLSTM) models achieved R²=0.628 with 79K parameters, while our Graph Neural Network with Temporal Attention (GNN-TAT) matched this accuracy with a compact 98K-parameter architecture through explicit topographic encoding. Late fusion via Ridge regression (R²=0.672) achieves a 6.2% improvement over individual models, while early fusion stacking (R²=0.212) and State Space Models (R²=0.200) fail. Elevation-stratified evaluation reveals 16% performance degradation from valleys to high elevations, concentrated in the 1,500-2,800 m orographic transition band, and variogram analysis confirms Late Fusion produces 29% lower spatial error variance. Sub-cell Digital Elevation Model (DEM) feature engineering (10 elevation deciles, 6 Principal Component Analysis components, 5 statistical descriptors) consistently degrades all architectures by 5.6-20.1%, establishing that categorical elevation encoding outperforms continuous terrain statistics under data scarcity. Non-parametric statistical testing (Friedman, Nemenyi, Mann-Whitney) confirms that the top three individual architectures are statistically indistinguishable (p=0.052), supporting the existence of a predictability ceiling for this region. Fusion strategy matters more than architecture choice; documenting negative results advances understanding as effectively as reporting successes.

Forecasting monthly precipitation in mountainous terrain poses challenges that push conventional deep learning approaches to their limits: convective processes operate locally while orographic effects span entire drainage basins. We compare three architecture families on precipitation prediction across the Colombian Andes: ConvLSTM (convolutional recurrent), FNO-ConvLSTM (spectral–temporal), and GNN-TAT (graph attention LSTM). Using CHIRPS v2.0 and SRTM topography for Boyacá department (61 × 65 grid, 3965 nodes), we evaluate 39 configurations across feature bundles (BASIC, KCE elevation clusters, and PAFC autocorrelation lags) and horizons from 1 to 12 months. GNN-TAT matches ConvLSTM accuracy ($R^2$: 0.628 vs. 0.642; RMSE: 82.29 vs. 79.40 mm) with 95% fewer parameters (∼98K vs. 2.1M). Across configurations, GNN-TAT produces a lower mean RMSE (92.12 vs. 112.02 mm; $p=0.015$) and a 74.7% lower variance. The explicit graph structure, with edges weighted by elevation similarity, appears to reduce sensitivity to hyperparameter choices. Pure FNO struggles with precipitation’s spatial discontinuities ($R^2=0.206$), though adding a ConvLSTM decoder recovers much of the lost skill ($R^2=0.582$). Elevation clustering improves GNN-TAT significantly ($p=0.036$) but not ConvLSTM, suggesting that feature design should match the spatial encoding paradigm. ConvLSTM achieves peak accuracy on local patterns; GNN-TAT provides robust predictions with interpretable spatial reasoning. These complementary strengths motivate stacking ensembles that combine grid-based and graph-based representations.

Monthly precipitation prediction in mountainous terrain is challenging due to orographic forcing and complex atmospheric interactions. We present a systematic framework evaluating hybrid deep learning architectures for the Colombian Andes (Boyacá, 3,965 grid cells, 145-5490 m elevation, forecast horizons from 1 to 12 months). In a companion study, Convolutional Long Short-Term Memory (ConvLSTM) models achieved R²=0.628 with 79K parameters, while our Graph Neural Network with Temporal Attention (GNN-TAT) matched this accuracy with a compact 98K-parameter architecture through explicit topographic encoding. Late fusion via Ridge regression (R²=0.672) achieves a 6.2% improvement over individual models, while early fusion stacking (R²=0.212) and State Space Models (R²=0.200) fail. Elevation-stratified evaluation reveals 16% performance degradation from valleys to high elevations, concentrated in the 1,500-2,800 m orographic transition band, and variogram analysis confirms Late Fusion produces 29% lower spatial error variance. Sub-cell Digital Elevation Model (DEM) feature engineering (10 elevation deciles, 6 Principal Component Analysis components, 5 statistical descriptors) consistently degrades all architectures by 5.6-20.1%, establishing that categorical elevation encoding outperforms continuous terrain statistics under data scarcity. Non-parametric statistical testing (Friedman, Nemenyi, Mann-Whitney) confirms that the top three individual architectures are statistically indistinguishable (p=0.052), supporting the existence of a predictability ceiling for this region. Fusion strategy matters more than architecture choice; documenting negative results advances understanding as effectively as reporting successes.

We evaluate hybrid deep learning architectures and ensemble strategies for monthly precipitation prediction over Boyacá, Colombia (3,965 CHIRPS grid cells, 145–5,490 m elevation, horizons 1–12 months). Three spatial encoding paradigms are compared: convolutional (ConvLSTM), spectral (Fourier Neural Operator hybrids), and graph-based (Graph Neural Network with Temporal Attention, GNN-TAT). GNN-TAT matches ConvLSTM accuracy (R2: 0.628 vs 0.642) with 95% fewer parameters and lower variance, leveraging elevation-weighted edges for interpretable spatial reasoning. Beyond individual models, late fusion via Ridge regression (R2=0.668) improves over all single architectures by exploiting complementary grid-based and graph-based error structures. Conversely, early fusion stacking collapses to R2=0.212, showing that combining predictions preserves inductive biases while merging intermediate representations destroys them. We also report the first evaluation of State Space Models (Mamba) for regional precipitation, which fail to transfer from sequence modeling (R2=0.200). Three operational guidelines emerge: graph-based encoders are efficient alternatives in complex terrain, ensemble gains depend on late-stage combination, and documenting architectural failures narrows the search space for future practitioners. All experiments use standardized CHIRPS/SRTM inputs and fixed random seeds for reproducibility.Keywords: monthly precipitation prediction, deep learning, hybrid architectures, Graph Neural Networks, ConvLSTM, ensemble learning, mountainous terrain, Colombian Andes, State Space ModelsRelated Publications:1. Pérez Reyes, M.R.; Suárez Barón, M.J.; García Cabrejo, Ó.J. Spatiotemporal Prediction of Monthly Precipitation: A Systematic Review of Hybrid Models. Hydrology Research (IWA Publishing), under review — Revision 3.2. Pérez Reyes, M.R.; Suárez Barón, M.J.; García Cabrejo, Ó.J. Hybrid Deep Learning Architectures for Multi-Horizon Precipitation Forecasting in Mountainous Regions: Systematic Comparison of Component-Combination Models in the Colombian Andes. Hydrology (MDPI), accepted.3. Pérez Reyes, M.R.; Suárez Barón, M.J.; García Cabrejo, Ó.J. A Data-Driven Deep Learning Framework for Monthly Precipitation Prediction in Complex Mountainous Terrain: Systematic Evaluation of Hybrid Architectures, Ensemble Strategies, and Emerging Paradigms. Hydrology (MDPI), ready for submission.