BioTransNet: Detection of Plant Stress Through the Conversion of Biosensor Data into RGB Channels and Combination with Transformer Networks

Abstract

The primary goal of this research is to develop an AI-driven approach for assessing plant stress levels (healthy, moderately stressed, highly stressed) using biosensor data. The study aims to address the challenge of plant diseases that threaten agricultural productivity and food security, by efficiently identifying plant health status through advanced technologies. This study utilized a dataset consisting of environmental variables such as temperature, soil moisture, light intensity, humidity, nutrients, and pH levels, along with identifiers and timestamps for each plant. The data were transformed into RGB channels, enabling the use of 2D-transformer models (CaiT, DeiT3, ViT) for feature extraction. Different model combinations (CaiT & DeiT3, DeiT3 & ViT, CaiT & ViT, CaiT & DeiT3 & ViT) were tested, with the CaiT & ViT combination selected as the best. To optimize the model, the SelectKBest method was applied to identify the most significant features, with the Support Vector Machine (SVM) employed to refine the performance. The model identified five key features from a dataset with 960 features. These features were classified with a remarkable accuracy of 99.83%, demonstrating the effectiveness of the BioTransNet model in identifying plant stress levels. The results confirm that the BioTransNet approach is highly successful in assessing plant stress levels, offering an efficient and accurate solution for plant health monitoring. This AI-driven model could significantly improve agricultural productivity and contribute to food security by providing an effective method for detecting plant stress.