Predicting plant traits in ecosystem modeling

V. V. Brykin

Surgut State University

Abstract: In this paper, we present a novel hybrid machine learning model designed to significantly enhance the accuracy and robustness of plant trait prediction. The model integrates a pre-trained convolutional neural network (CNN), adept at extracting visual features from plant images, with a comprehensive set of auxiliary variables meticulously represented as tabular data (CSV). Plant traits, such as leaf area, stem diameter, and flowering time, are crucial indicators of plant health, productivity, and adaptation to environmental conditions. Accurate prediction of these traits is essential for a wide range of applications, including precision agriculture, crop breeding, biodiversity monitoring, and ecological modeling. However, traditional methods for plant trait measurement are often time-consuming, labor-intensive, and limited in scale. The aim of this study is to overcome these limitations by developing a robust and scalable machine learning approach that leverages both visual information and rich contextual data to improve the accuracy and efficiency of plant trait prediction. The increasing availability of large-scale plant image datasets, coupled with the proliferation of auxiliary data sources, presents an unprecedented opportunity to develop sophisticated machine learning models for plant phenotyping. However, effectively integrating these diverse data modalities remains a significant challenge. Existing approaches often rely solely on visual information or neglect the valuable contextual information encoded in auxiliary variables. Therefore, we propose a hybrid approach that combines the strengths of both visual and tabular data analysis techniques. The model was trained on a carefully curated dataset consisting of 9147 plant images, representing a diverse range of plant species and growth stages. These images were paired with 167 meticulously selected auxiliary variables encompassing a wide spectrum of relevant information, including environmental conditions (temperature, humidity, light intensity), plant species, geographical location, and cultivation practices. To effectively capture the rich visual information contained in the plant images, we employed transfer learning, a powerful technique that leverages the knowledge acquired by pre-trained CNNs on large image datasets. The images were pre-processed using three distinct and widely recognized pre-trained convolutional neural network architectures: InceptionV3, ResNet, and VGG19. These CNNs have demonstrated remarkable performance in image recognition tasks and are well-suited for extracting relevant visual features from plant images. To effectively integrate the visual features with the auxiliary data, the outputs of the CNNs were first straightened and then combined with the auxiliary data using a "concatenate" layer, enabling the model to learn complex relationships between the different data modalities. The model was meticulously optimized using the Adam algorithm, a computationally efficient and adaptive optimization technique, and rigorously evaluated using a suite of relevant metrics, including root mean squared error (RMSE), mean absolute error (MAE), and R-squared, to comprehensively assess its performance. To ensure the robustness and generalizability of the model, we divided the dataset into training, validation, and testing sets. The training set was used to train the model, the validation set was used to optimize hyperparameters, and the testing set was used to evaluate the final performance of the model on unseen data. Furthermore, we explored the individual contributions of the different CNN architectures to the overall performance of the hybrid model. The results convincingly demonstrated that the hybrid model significantly outperforms the baseline CNN in plant trait prediction, achieving an improvement in prediction accuracy across a range of plant traits. This compelling finding confirms the effectiveness of leveraging additional contextual data to enhance the predictive power of machine learning models for plant phenotyping. This research also demonstrates that the auxiliary variables provide valuable contextual information that complements the visual features extracted from the plant images. The study proves the significant potential of hybrid models for analyzing plant data collected from diverse sources, including citizen science initiatives and remote sensing platforms, thereby enabling large-scale plant phenotyping and contributing to a deeper understanding of plant biology and ecology. These findings have significant implications for future research in plant phenotyping and precision agriculture, paving the way for the development of more accurate and efficient methods for monitoring plant health, optimizing crop yields, and conserving biodiversity. The developed model can be extended and applied to other application domains, by considering suitable CNN-architectures and auxiliary data, and fine-tuning it accordingly.

Keywords: ecosystems, feature extraction, hybrid model, neural networks, plant images, regression.

UDC: 004.81