Edge Intelligence in Agriculture: Quantifying the Payload-Latency Trade-offs of Dynamic LSTM Quantization on ARM Architectures

The integration of Digital Twins in Agriculture 4.0 necessitates continuous, high-fidelity microclimate prediction. While Long Short-Term Memory (LSTM) networks provide exceptional accuracy for greenhouse state estimation, their deployment on resource-constrained Edge Gateways (e.g., Raspberry Pi) is severely hindered by high memory payload and energy consumption required for Over-The-Air (OTA) updates. This paper investigates the empirical trade-offs of applying Affine Post-Training Dynamic Quantization (PTDQ) to agricultural LSTM models, converting 32-bit floating-point (FP32) weights to 8-bit integers (INT8). Using a localized Cyber-Physical System (CPS) dataset from a semi-arid greenhouse in Biskra, Algeria, a highly compressed LSTM architecture was trained and dynamically quantized using PyTorch Native mechanisms (QNNPACK). Empirical hardware profiling directly on an ARM Cortex-A72 processor revealed a structural payload compression of 65.13% (73.23 KB down to 25.54 KB). While INT8 inference exhibited an architectural latency penalty (62.93 ms ± 6.49 ms vs. 1.76 ms ± 0.48 ms for FP32) due to runtime dynamic casting on the aarch64 CPU, this delay is operationally negligible given the high thermal inertia and 15-minute sampling rates of agricultural greenhouses. The massive 65% reduction in memory bandwidth and radio transmission payload yields significant theoretical energy savings for remote agricultural Edge deployments over LoRaWAN networks, heavily outweighing the extended processor active time. Furthermore, the quantization noise was restricted to a Root Mean Square Error (RMSE) of 0.0001°C, rendering it entirely imperceptible given the 1.52°C baseline ground-truth error of the Digital Twin. While bounded to a semi-arid dataset and Cortex-A72 architecture, these results provide an empirical baseline for future multi-climate and RISC-V extensions.