Synchro-Transient-Extracting Transform Based Derived Feature Engineering for Environmental Stress Indicator Estimation in Crops

Crop classification and environmental stress prediction in smart agriculture depend on engineered features that capture agronomic meaning beyond raw sensor readings. This paper proposes a feature-engineering stage that (i) derives five domain-grounded environmental stress indicators — Temperature-Humidity Index (THI), Nutrient Balance Ratio (NBR), Water Availability Index (WAI), Photosynthesis Potential (PP), and Soil Fertility Index (SFI) — from raw soil and atmospheric measurements, and (ii) applies a Synchro-Transient-Extracting Transform (STET)-style time-frequency analysis to each derived index stream, extracting instantaneous-frequency, instantaneous-amplitude, and Hilbert-based transient descriptors as additional engineered features. Building on the Robust Maximum Correntropy Kalman Filter (RMCKF)-reconstructed and Efficient Multiplayer Battle Game Optimizer (EMBGO)-selected feature set from the preceding pre-processing stage, the proposed 29-dimensional engineered feature set is evaluated on the 2200-sample, 22-crop Crop Recommendation dataset using a Random Forest classifier under 5-fold cross-validation. The proposed feature set achieves 98.59% accuracy, 98.64% precision, 98.59% recall, 98.59% F1-score, 99.93% specificity, and 99.99% AUC, outperforming the 7-feature RMCKF baseline (98.23% accuracy), the 4-feature EMBGO-selected raw set (95.73%), and the derived indices used alone without STET descriptors (97.14%). An ablation study isolating the STET harmonic component (instantaneous frequency/amplitude) and transient component (Hilbert amplitude/frequency) shows both contribute complementary information, with the harmonic component contributing more (98.41% accuracy alone) than the transient component (97.82%) but their combination yielding the best result. These findings demonstrate that combining domain-grounded derived indicators with synchro-extracted time-frequency descriptors provides a compact, interpretable, and effective feature representation for downstream crop classification and stress prediction.