A neural network approach for real-time monitoring of cannabis sativa L. germination

Abstract

One of the primary challenges in detecting crop germination in seedbeds using computer vision is the difficulty of accurately analysing cases of overlapping seedlings. This study investigates the effectiveness of Convolutional Neural Network (CNN) models and Long Short-Term Memory (LSTM) models in addressing this issue. Utilizing a substantial dataset comprising over 80,000 labelled images of Cannabis Sativa plants, our research aims to compare the performance of standalone CNN models with hybrid architectures (CSS+LSTM model). A noteworthy aspect of our experimental methodology is the deliberate omission of data augmentation techniques during dataset preparation. This decision enables us to evaluate the inherent quality and utility of our curated dataset without introducing artificial modifications. Furthermore, recognizing the significance of incorporating a temporal component in germination detection, we conduct a specific assessment of the hybrid model (CNN+LSTM) against the standalone CNN model. Through comprehensive experimentation and analysis, we evaluate the relative effectiveness of each model in accurately classifying germination status across different levels of seedling overlap. Preliminary results reflect a better performance of the hybrid model compared to the standalone CNN model. However, it is crucial to consider the computational resources, time, and effort required for training and testing the model. Thus, our study provides valuable insights into the intricate interplay among model architecture, dataset characteristics, and the complexity of the germination detection task. Moreover, this research contributes to the advancement of practical applications of deep learning methodologies in agricultural monitoring, emphasizing the necessity of tailored model designs to overcome the unique challenges encountered in greenhouse environments.