LSTM Neural Networks Anomaly Detection for Biotic and Abiotic Early Stress Detection on Tomato Plants

This study explores time series anomaly detection using a long short-term memory (LSTM) neural network to identify abnormal plant conditions—both biotic and abiotic—by analyzing a stem frequency parameter. This parameter, closely linked to plant stem conductivity, serves as a novel indicator of plant hydration and overall physiological status. Four independent experiments were conducted between 28 May 2024, and 16 December 2024, and 20 tomato plants were maintained under different conditions (healthy, biotic, and abiotic stress) to assess physiological differences and develop models for early symptom detection.We used samples from healthy plants to train an LSTM neural network to predict stem frequency values 24 h ahead. The trained model was then used to forecast the behavior of stressed plants, and an anomaly was flagged whenever the predicted values significantly deviated from the measured ones. Our LSTM-based anomaly detection model successfully detected water stress conditions several days before visible symptoms appeared. However, the algorithm struggled to distinguish early signs of Fusarium infection. Despite this limita- tion, in most cases, the model provided early warnings of biotic stress before any visual symptoms were evident. Future research will focus on expanding the dataset to enhance the model’s ability to differentiate between various types of plant stress. This article builds upon the work presented at CAFE 2024 Cum et al., 2024, utilizing the same experimental setup but exploring a completely different approach for the early identification of stress symptoms in tomato plants.