Resource-constrained environmental sensing systems must simultaneously satisfy conflicting requirements for predictive intelligence, communication reliability, memory efficiency, and energy conservation. Although Tiny Machine Learning (TinyML) enables local inference on microcontroller-class devices, existing environmental monitoring systems frequently treat environmental inference, transmission scheduling, and radio-frequency resource allocation as separate optimization problems. This study proposes a value-aware RF-enabled TinyML architecture in which environmental severity, prediction uncertainty, temporal novelty, information freshness, and RF-link state jointly determine whether information is transmitted and the communication resources assigned to its delivery. A computational experimental evaluation was performed using a real environmental time series containing 2,231 usable temporal sequences. Four candidate models Decision Tree, Random Forest, multilayer perceptron, and compact one-dimensional convolutional neural network were evaluated using chronological training, validation, and independent test partitions. The Random Forest achieved the highest test performance, with 89.85% accuracy, 88.08% macro-F1, and 91.18% recall for the highest-severity class. A Value-of-Information controller subsequently combined model severity, predictive entropy, temporal novelty, and information age. Four communication strategies were compared: periodic fixed-RF monitoring, periodic TinyML telemetry, event-triggered TinyML telemetry, and the proposed value-aware adaptive RF architecture. The proposed architecture reduced transmission count by 72.54% and transmitted bytes by 74.22% relative to periodic monitoring. Across 300 modeled RF-link realizations, mean packet-delivery ratio increased from 70.78% for conventional periodic telemetry to 94.44%, while mean critical-event delivery increased from 76.41% to 87.04%. Mean modeled communication energy decreased by 32.13% relative to periodic transmission, although a simpler event-triggered architecture consumed less energy at the cost of substantially poorer reliability. The results demonstrate that communication efficiency in intelligent environmental sensing should be defined not only by energy minimization but by the reliable delivery of high-value environmental information. Physical microcontroller, LoRa, and power measurements remain necessary for final hardware validation.

Keywords: TinyML; environmental monitoring; edge intelligence; LoRa; adaptive RF telemetry; Value of Information; embedded machine learning; resource-constrained IoT; low-power sensing; Internet of Things.