Research on soil testing and planting in STM32 based on lightweight neural networks

To facilitate scientific planting based on soil conditions, this paper presents a soil detection and tillage assistance system powered by an STM32 embedded microprocessor. The centerpiece of the system is the STM32F103 microcontroller, which interfaces with sensors for temperature and humidity, light intensity, and soil pH to collect pertinent soil data and transmit it to the host computer. On the host side, an agricultural knowledge graph is constructed, enabling an agronomic Q and A system that melds soil data with agricultural expertise. TensorFlow Lite is employed on the host to implement a quantized neural network training regimen, achieving significant model compression to ensure effective operation within the resource-constrained embedded environment. The textual matching algorithm employs Bert+DSSM, enhancing accuracy from 79% to 86% over the DSSM alone. This breakthrough leads to an information-rich visual decision support platform, augmenting decision-making in agricultural endeavors.