Development of portable soil organic matter system based image and spectral fusion

Abstract Soil organic matter (SOM) content is an important indicator of agricultural soil fertility. A portable detection device was designed by combining near-infrared (NIR) spectroscopy with soil image information technology to rapidly and accurately determine the SOM content. The system extracts the RGB color histogram from pre-processing soil images, such as image cropping and overexposure removal, to improve the validity of image data. Subsequently, the color histogram information is fused with near-infrared spectral data. Meanwhile, a self-attention generative adversarial network (SA-GAN) is proposed to expand SOM fusion data, addressing the challenge of limited soil sample availability for deep learning. 120 soil samples and their corresponding NIR data, image data, and true values of organic matter were collected from the North China Plain, China. Three models, namely, Support Vector Machine (SVM), Partial Least Squares Regression (PLSR), and Convolutional Neural Network (CNN) were used for SOM content prediction. The experimental results show that after data fusion and expansion, the R² values of SVM, PLSR, and CNN models improved from 0.59, 0.55, and 0.60 to 0.73, 0.76, and 0.88, respectively. Concurrently, the RMSEs decreased from 7.84, 8.11, and 5.65 to 3.60, 3.21, and 2.08, indicating higher predictive accuracy across all models. In addition, the portable device integrated with the prediction model was validated in the field, achieving R² of 0.80. It is proven that the system can effectively detect the SOM content in real-time, which provides important technical support and a reference basis for guiding smart agricultural production.