Growth prediction of tomato seedlings based on causal LSTM and GAN

The stable production of seedlings is very important for seedling growers. Predicting the growth of seedlings helps growers promptly adjust management strategies and production expectations. Traditional methods rely on historical growth data or assess current plant physiological parameters to estimate future growth. This study aims to predict future images directly from historical growth images of tomato seedlings. Specifically, a dataset of 10-d image sequences of tomato seedlings was collected. Then, an algorithm based on several neural networks was applied to predict the images of the next 5 d based on the images of the first 5 d. The algorithm was composed of a causal long short-term memory (LSTM) unit, a gradient highway unit (GHU), and a pix2pix unit. The experimental results showed that the introduction of a Generative Adversarial Network (GAN) further enhanced the clarity and realism of the predicted images, ensuring higher quality and more accurate visual results. From the perspective of image similarity, the average mean squared error (MSE) reached 394.97 and the average structural similarity (SSIM) reached 0.90 over 5 d. From the perspective of biological information, the average prediction errors of the plant area were 1.7, 1.4, 1.5, 0.9, and 3.2 cm² over the 5 d, and the average prediction errors of plant height were 1.7, 1.9, 4.6, 6.9, and 4.5 mm, respectively. The extracted biological information such as plant area and height showed good following performance compared with the real growth information. The research results show that predicting future plant images from historical images has the potential to become a useful tool for nursery growers to adjust management strategies and production expectations.