Predicting plant stress using SAM-L: novel self-adaptive-meta learner with XAI based on soil moisture and chlorophyll...

Recent advancements in precision agriculture have introduced innovative approaches to addressing plant stress, a critical factor influencing crop productivity and agricultural sustainability. Accurate, real-time prediction of plant stress has become essential for optimizing water utilization and promoting healthy crop development. While existing machine learning methods have demonstrated efficacy, they often lack the adaptability required to accommodate the dynamic conditions of agricultural environments. Prior research has identified soil moisture and chlorophyll content as key indicators of plant health and stress, with conventional models relying on simplistic algorithms for stress prediction. However, these models exhibit limitations in scalability, adaptability and interpretability. To overcome these challenges, this study employed sparse additive models with learning (SAM-L) algorithms, integrated with explainable artificial intelligence (XAI), to provide a flexible and transparent solution. In this paper, we proposed a novel framework that integrates SAM-L and XAI to predict plant stress using soil moisture and chlorophyll content. The SAM-L algorithm is a machine learning method that focuses on sparsely selecting relevant features through additive models. It aims to enhance model interpretability while maintaining high prediction accuracy by learning sparse representations of input data. The SAM-L algorithm enhances interpretability while preserving high predictive accuracy by learning sparse feature representations from input data. Additionally, XAI was incorporated to ensure interpretable decision-making, enabling farmers and stakeholders to comprehend the rationale behind irrigation recommendations. The model's architecture incorporates a three-layer Long Short-Term Memory (LSTM) network to process sequential data effectively. The proposed framework achieved a high performance on publicly available dataset, yielding an overall accuracy of 89.2% on the multi-class classification task. Further analysis of the results across the three predefined stress categories (healthy, moderate stress, and high stress) revealed strong performance, with the model obtaining a macro F1-score of 0.88 and a macro recall of 0.88. The proposed framework not only can enhance prediction accuracy but also can promote sustainable farming practices by reducing water wastage and improving crop resilience.