A Deep Hierarchical Feature Learning Framework for Multimodal Tomato Growth Stage Classification and Yield Prediction

Abstract

Precise monitoring of tomato plant development is vital for precision agriculture and sustainable crop management. In this study, we present a Deep Hierarchical Feature Learning Network (DHFL-Net) for automatic tomato growth-stage categorization and yield estimation using paired RGB and temperature pictures. The research employs 700 paired samples consisting of 700 RGB photos and 700 related thermal images of the same tomato plants. Preprocessing using Adaptive Noise Suppression Module (ANSM) and Plant-Aware Region Extraction Module (PAREM) to extract informative plant areas. DHFL-Net fuses and manipulates RGB morphology and Multi-Scale Thermal Texture Encoder (MTTE) descriptors. All the comparing algorithms were reimplemented and tested on the same dataset split for a fair comparison. The DHFL-Net attained an accuracy of 98.76%, precision of 98.42%, recall of 98.18% and F1-score of 98.30% respectively. It achieved R2=0.982, RMSE=0.084 and MAE=0.061 for yield prediction. Results indicate the value of combined RGB–thermal information for robust tomato phenotyping and yield calculation.