SpectraNode: An Embedded TinyML Platform for Autonomous Nonlinear Spectral Correction in Controlled-Environment Agriculture

SpectraNode is a low-cost embedded prototype designed for autonomous real-time spectral correction in sensor nodes for controlled-environment agriculture. The system acquires spectral measurements at fixed monitoring points within horticultural LED-lit growing facilities, delivering more reliable spectral data without dependence on network infrastructure or external processing.

The prototype integrates an AS7265x 18-channel multispectral sensor (410–940 nm) connected via I²C to an Arduino Nano 33 BLE Sense Rev2 with an ARM Cortex-M4F microcontroller (nRF52840, 64 MHz, 256 KB SRAM). The system pipeline includes: acquisition of 18 raw spectral band values, on-device normalization, inference through an 8-bit quantized multilayer perceptron (MLP) running on TensorFlow Lite Micro, and wireless transmission of the 18 corrected spectral values via BLE 5.0 to an Android application for real-time visualization and field data export. The MLP architecture consists of three hidden layers with 64, 32, and 16 neurons using ELU activation and an 18-neuron linear output layer covering the full spectral range (410–940 nm). The model occupies approximately 10–16 kB, under 7% of available SRAM, allowing fully standalone operation without network connectivity.

The model was validated against a commercial spectroradiometer (OHSP-350C) under horticultural LED lighting conditions representative of controlled-environment agriculture. A linear regression baseline trained on the same data reduced the normalized RMSE by 45%. The proposed embedded MLP achieved an 80% RMSE reduction, outperforming the linear baseline. Inference latency on the Cortex-M4F confirmed real-time operation, with a memory footprint below 7% of available SRAM.

To the best of the authors' knowledge, this is the first system to combine on-device nonlinear spectral correction with BLE 5.0 transmission to a mobile Android application on a standalone Cortex-M4F node for controlled-environment agriculture, in contrast to prior work that depends on Wi-Fi connectivity and server-side processing. Future work will address on-device spectral drift detection and adaptive retraining under varying lighting conditions and sensor aging.