Advanced ML models with optimized sampling predict land subsidence with 98% accuracy

Predicting where the ground will sink is harder than it sounds. You need data on places where subsidence has happened, yes, but you also need reliable data on places where it hasn't—and that's where most prediction models stumble. A team of researchers has now developed a way to gather that absence data more intelligently, then trained machine learning models on it with such precision that they can identify high-risk subsidence zones with 98.5% accuracy.

The problem sits at the foundation of how these models learn. When you're trying to predict land subsidence susceptibility—the likelihood that ground in a given area will sink—you train your algorithm on two kinds of information: locations where subsidence actually occurred, and locations where it didn't. The second category is vast and shapeless. How do you choose which non-subsiding areas to include in your training set? Choose poorly, and your model learns the wrong patterns. The researchers proposed a new hybrid sampling approach grounded in two fundamental principles of geography: that nearby places tend to be similar to each other, and that processes operating at one scale influence processes at other scales. Using these principles, they generated absence data that better represented the actual geography of the study area.

With better training data in hand, they then tackled a second challenge: optimizing the machine learning models themselves. They employed two different algorithms—Support Vector Regression (SVR), a classical machine learning workhorse, and Convolutional Neural Networks (CNN), a deep learning approach that excels at finding spatial patterns. But both models have many adjustable settings, called hyperparameters, that dramatically affect performance. Finding the best combination is computationally expensive. The team used an advanced optimization algorithm inspired by the hunting behavior of grey wolves, enhanced with strategies borrowed from other nature-inspired algorithms and guided by Q-learning, a reinforcement learning technique. This multi-layered approach allowed them to simultaneously optimize both the features fed into the models and the hyperparameters controlling how they learn.

The results were striking. The SVR model, when optimized and trained on the new sampling dataset, achieved an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.9853—a measure of how well it distinguishes high-risk from low-risk areas. Its root mean squared error was 0.2449, and its R² value was 0.7600. The CNN model performed even slightly better: an AUC-ROC of 0.9833, a lower error of 0.2328, and an R² of 0.7831. When the researchers mapped these predictions back onto the actual study area, they found that 97.54% of known subsidence zones fell into the high or very high susceptibility classes according to the SVR model, and 96.57% according to the CNN model. The spatial agreement was remarkable—the models weren't just producing good numbers on a test set; they were correctly identifying where the ground actually sinks.

The researchers also used a technique called SHAP (SHapley Additive exPlanations) to understand which features—which input variables—mattered most to the models' predictions. This transparency matters because it lets geologists and engineers understand not just what the model predicts, but why. The combination of better data, smarter optimization, and interpretable results suggests a path forward for other regions vulnerable to subsidence. As cities expand, as aquifers deplete, as the ground shifts beneath us, the ability to predict where and how much it will sink becomes increasingly valuable. This work shows that with careful attention to how you prepare your data and how you tune your algorithms, you can achieve the precision needed to guide real decisions about where to build, where to monitor, and where to prepare for the ground to move.