A decision-level integration approach for detection and classification of leaf diseases in greenhouse tomato crops using deep learning and machine learning models

Accurate disease classification of tomato leaf diseases is crucial for reducing yield losses and supporting sustainable agricultural practices. This study proposes a decision-level integration approach for the comprehensive assessment of phytosanitary status and soil and environmental conditions in tomato crops, combining EfficientNetB0-based disease classification with Random Forest-based analysis of edaphic and environmental variables. The proposed framework incorporates a deep learning model based on EfficientNetB0, fine-tuned to classify eleven tomato leaf disease classes using 38,500 annotated images, alongside a Random Forest model trained on 2,880 records comprising soil pH, electrical conductivity, nitrogen (N), phosphorus (P), and potassium (K), soil moisture and temperature, as well as environmental and solar radiation variables collected under controlled greenhouse conditions. The outputs of both models are integrated at the decision level using a qualitative decision-level integration matrix, enabling the identification of relationships between disease incidence and soil properties, categorized into three qualitative levels: Deficient, Optimal, and Excess. The deep learning model achieved high performance, with precision and recall values above 97% and an F1-score exceeding 98%, exhibiting minimal confusion among visually similar disease classes. The Random Forest model reached an overall 93.7% of precision demonstrating stable and balanced classification across the three qualitative levels. Furthermore, qualitative integration matrix validated the integrated framework by revealing consistent associations between soil properties categories and specific leaf diseases. Overall, the results show that decision-level integration of visual and physiological data improves the system’s interpretability and enhances its ability to anticipate plant health risks, making the proposed approach an effective decision-support tool for precision agriculture.