Advanced Groundwater Level Forecasting using QSO-based Vision Transformer Model for Sustainable Water Resource Management

The reliable predictions of groundwater levels are crucial for long-term management of water resources, and they are an excellent source for human well-being. While conventional ML approaches work well for low-dimensional data, they are optimized for hyperparameters on high-dimensional inputs and also capture complex temporal correlations. To address these restrictions, this research presents a new framework for predicting groundwater-level changes up to five months in advance, built on the Vision Transformer (ViT) and optimised with Quokka Swarm Optimisation (QSO). ViT enables strong global feature extraction and long-range dependency modelling by processing time-series data as sequential image-like patches, in contrast to traditional neural networks. Drawing on quokka’s adaptive survival behaviour, the QSO procedure optimises the transformer's hyperparameters, such as patch size, attention heads, and depth, in real time to enhance prediction accuracy. The ViT+QSO outperformed baseline deep learning methods on groundwater datasets from Southern Africa in terms of RMSE, MAE, correlation coefficient (R), and Nash-Sutcliffe Efficiency (NSE). Quantile regression uncertainty quantification further improves the model's reliability for water resource planning. Hydrological variables, in addition to climate indices, influence groundwater fluctuations, as confirmed by ablation research. The proposed ViT-QSO achieved 93% R and 0.118 MAE, whereas the basic ViT achieved only 89.1% R and 0.152 MAE for groundwater-level prediction. Scalability, interpretability, and suitability for areas with limited monitoring infrastructure are hallmarks of the proposed methodology. This research provides valuable insights into how to better withstand the effects of climate change, in addition to human activities, on groundwater supplies. ABSTRAK: Ramalan paras air bawah tanah yang boleh dipercayai adalah penting bagi pengurusan jangka panjang sumber air dan kesejahteraan manusia; namun, pendekatan pembelajaran mesin konvensional berhadapan kekangan pengendalian input berdimensi tinggi serta model korelasi temporal kompleks. Kajian ini mencadangkan satu rangka kerja baharu berasaskan Pengubah Visi (ViT) yang dioptimum menggunakan Optimisasi Kawanan Quokka (QSO) dalam meramal perubahan paras air bawah tanah pada lima bulan lebih awal. ViT memproses data siri masa sebagai tampalan jujukan menyerupai imej bagi membolehkan pengekstrakan ciri global dan model kebergantungan jarak jauh, manakala QSO mengoptimumkan hiperparameter pengubah secara adaptif dalam meningkatkan ketepatan ramalan. Model ViT-QSO menunjukkan prestasi unggul berbanding kaedah pembelajaran mendalam asas pada set data air bawah tanah di Afrika Selatan, dengan peningkatan ketara dari segi RMSE, MAE, pekali korelasi (R), dan Kecekapan Nash–Sutcliffe (NSE), serta mencapai nilai R sebanyak 93% dan MAE 0.118 berbanding ViT asas masing-masing mencatatkan 89.1% dan 0.152. Pengkuantitian ketidakpastian melalui regresi kuantil meningkatkan kebolehpercayaan model bagi perancangan sumber air, manakala kajian ablasi mengesahkan peranan pembolehubah hidrologi dan indeks iklim terhadap turun naik paras air bawah tanah. Secara keseluruhan, metodologi yang dicadangkan adalah berskala, boleh ditafsir, dan sesuai pada kawasan infrastruktur pemantauan terhad, serta memberikan sumbangan penting dalam menangani kesan perubahan iklim dan aktiviti manusia terhadap sumber air bawah tanah.