Answers are grounded in this paper and cite their sources.
This study evaluates selective dynamic 8-bit integer (INT8) quantization of long short-term memory (LSTM) recurrent weights while retaining a 32-bit floating-point (FP32) linear output layer, together with authenticated sensor processing, using 10,257 observations from the Secured and Optimized IoT-Based Irrigation System (SOIBIS) experimental farm at Federal University Lokoja, Nigeria. Five seeded runs compare eight models under chronological and feature-group evaluations. The chronological test contains only label pattern 101, for which a majority baseline also achieves 100% exact-match accuracy. On the class-diverse feature-group test, selective quantization achieves 98.40% mean exact-match accuracy, compared with 100% for a decision tree and 84.68% for a majority baseline. Selective quantization reduces serialized model size by 64.9% and single-record host model-only latency by 50.7%, but increases 12-record model-only latency by approximately 41.1%. Mean authenticated single-record inference-path time is 0.441 ms on the host. All 1,200 specified functional packet checks met their expected outcomes. These are host-software findings, not evidence of ESP32 deployment, independently verified pump actuation, water savings, or agronomic superiority.
No takes yet. Share an insight, caveat, or question.
Isiaka et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: