CNViS-Net: A Self-Attention Framework for Predicting Soil Nutrients from Hyperspectral Data

Automating soil parameter analysis simplifies fertilization, resulting in time and resource savings. This efficiency reduces agricultural production costs and fosters better resource utilization, contributing to sustainable farming practices and improving food security. This research presents CNViS-Net, a novel self-attention framework for soil nutrient prediction using hyperspectral data, combining Deep Learning (DL) with Machine Learning (ML) techniques. The focus is on predicting four soil nutrients: Phosphorus (P), Potassium (K), Magnesium (Mg), and soil pH, using the HYPERVIEW hyperspectral dataset. The proposed CNViS-Net framework performs feature extraction through a combination of Convolutional Neural Network (CNN) and Vision Transformer (ViT), capturing complex spectral patterns across 150 channels of hyperspectral data. It passes the extracted features to four ML models, which are K-Nearest Neighbors (KNN), Random Forest (RF), Partial Least Squares Regression (PLSR), and Support Vector Regression (SVR) for final prediction. The combination of SVR with CNN and ViT outperforms other prediction models, achieving significantly lower score metrics of 0.164, 0.074, 0.307, and 0.66 for P, K, Mg, and pH, respectively. These remarkably lower scores demonstrate CNViS-Net's superior performance over previous state-of-the-art methods in accurately predicting soil nutrient levels, advancing precision agriculture through efficient hyperspectral data analysis.