Adaptive Gradient-Boosting-Based Sensor Selection for Energy-Efficient Wireless Sensor Networks (WSNs)

Energy conservation remains a fundamental challenge in wireless sensor networks (WSNs), where battery-powered nodes are often deployed in locations where recharging is impractical. Because spatially correlated sensors frequently report redundant readings, deactivating less informative sensors can extend network lifetime while retaining nearly the full-network classification performance. Prior sensor-selection studies, however, have often evaluated fixed sensor counts on public classification benchmarks without ground-truth labels identifying which sensors are genuinely informative, while estimating energy savings primarily from sensor-count ratios rather than topology-dependent physical energy models. This paper presents an adaptive sensor-selection framework that uses gradient-boosted trees with permutation importance to rank sensors and identify the smallest subset that retains a predefined fraction of full-network classification accuracy. The framework is evaluated using a purpose-built synthetic benchmark with known ground-truth sensor informativeness and an explicit network topology governed by a first-order radio energy model, together with eight independently re-implemented baseline methods evaluated under identical conditions across eight data-partition seeds. At the selected operating point, GBM-Permutation retains 99.3% of the full-sensor classification accuracy (84.6% versus 85.2%) using only 10 of 54 sensors, corresponding to an 82.6% reduction in modeled per-round energy consumption; modeled first-node-dies network lifetime increases from 615 rounds (using all 54 sensors) to 883 rounds (a 1.4-fold increase), distinct from the sensor-count-based Lifetime Extension Factor (LEF) of 5.4. The proposed method significantly outperforms all baseline methods in the primary synthetic evaluation (p < 0.01). A redundancy–severity sensitivity analysis and validation on two generic real-world tabular datasets and one real sensor-derived dataset show that the advantage is strongest under moderate redundancy rather than being universal. These findings demonstrate the value of ground-truth-validated and topology-aware evaluation for identifying sensor-selection methods that are effective under energy-constrained WSN conditions.