Convolutional neural network hardware implementation for soil roughness estimation

đź“… July 2023

🧑 S. Popa, K. Marandskiy, G. Feldioreanu, M. Ivanovici

#VGG #convolutional neural network #soil roughness estimation

In the Agriculture 5.0 context, computer vision and artificial intelligence are two key technologies for the automatic agricultural field and crop monitoring and management. For enabling the usage of both technologies on portable or mobile devices, including unmanned aerial vehicles, hardware implementations are required, for real-time performance, high-precision running and low power consumption. We implemented a Very Deep Convolutional Neural Network (VGG-11) for Soil Surface Roughness’s (SSR) Random Roughness (RR) parameter estimation from digital images of a line laser beam projection on the analysed surface. We used the Python PyTorch framework for modelling the VGG-11 in software. We compare the results against the classical contact measurement pinboard method on both artificial and real soil surfaces. SSR is defined as the irregularities of the soil surface, as a consequence of various factors including soil texture, aggregate size, rock fragments, vegetation cover and land management. SSR is directly related to soil water storage, infiltration and overland flow. Our VGG-11 is trained for 120 epochs and provides a 99.42% accuracy on the test set. This software model for SSR’s RR estimation is intended as a golden model for the validation of a hardware implementation using the Verilog hardware description language. The hardware implementation is mainly targeted at Field Programmable Gate Array (FPGA) devices and can easily be interfaced with a microprocessor system. A similar hardware implementation of a VGG-11 is presented in [1].

https://ai4agri.unitbv.ro/wp-content/uploads/2024/04/M_Ivanovici_abstract_EARSeL2023.pdf