Improved Detection of Mixed Pesticide Solution Concentrations Using Deep Learning with Segmented Stride 1D-CNN and Color Rendering Index Features

Highlights CRI features were integrated with spectral data into a 1D-CNN, enhancing prediction accuracy. A segmented stride design was implemented to optimize feature extraction in critical spectral regions. The CRI-integrated 1D-CNN outperformed traditional methods, achieving superior predictive accuracy. The model's flexibility and adaptability were validated, proving its robustness for real-time applications. ABSTRACT. Accurate detection of pesticide mixture concentrations is crucial for optimizing pesticide utilization efficiency. This study introduces an innovative approach using a One-Dimensional Convolutional Neural Network (1D-CNN) integrated with Color Rendering Index (CRI) features for real-time liquid concentration measurements. A portable spectrometer was employed to capture the spectral curves of transmitted light from mixtures processed by a dispensing mixer. The 1D-CNN model, enhanced by the integration of CRI features (R1-R15), demonstrated superior predictive accuracy and stability compared to traditional methods such as Partial Least Squares Regression (PLSR). The proposed segmented stride design in the 1D-CNN model optimized feature extraction by focusing on critical spectral regions, significantly improving model performance. Key findings include the CRI-integrated 1D-CNN model achieving an R 2 of 0.9912 and an RMSE of 0.3207, outperforming traditional PLSR models (R 2 =0.9792, RMSE=0.5063), and the segmented stride 1D-CNN model exhibiting a mean R 2 of 0.9738 and an RMSE of 0.5549, marking a 4.4% improvement in R 2 and a 38.1% reduction in RMSE compared to a uniform stride 1D-CNN with a stride length of 2. Furthermore, the maximum relative errors for solution concentrations of 6, 7, 8, and 15 mg/L were reduced from 11.1%, 14.1%, 16.8%, and 3.9% to 6.6%, 6.8%, 8.25%, and 3.6%, respectively. This study presents a novel deep learning method combining a segmented stride 1D-CNN with CRI features to predict the concentration of mixed solutions. The segmented stride 1D-CNN utilizes different convolutional kernel strides for various spectral bands, allowing the model to capture local features at different scales across the spectrum, significantly enhancing predictive performance. The findings validate the potential of integrating advanced machine learning techniques with spectral analysis for the precise and efficient online measurement of pesticide concentrations, offering significant advantages for agricultural monitoring. This approach ensures optimal pesticide use, enhanced pest control efficacy, minimized pesticide wastage, and reduced environmental impact by preventing over-application or under-application of pesticides. The model's flexibility and adaptability were validated, proving its robustness for real-time applications. Keywords: Color rendering index (CRI), Convolutional neural network (CNN), Deep learning, Feature extraction, Mixed solution concentration, Partial least squares regression (PLSR), Segmented stride, Spectral analysis.