Potential of machine learning in leaf-based multi-source data driven tomato growth monitoring

Tomatoes (Solanum lycopersicum L.) represent a crucial fruit and vegetable crop whose leaf is a significant phenotypic parameter regulating photosynthesis and growth in a greenhouse environment. Chlorophyll content and leaf area index (LAI) are essential leaf indicators for directly monitoring alterations in tomato growth status. Therefore, this study conducted two years of experiments for collecting tomato growth parameters [tomato yield, vitamin C (VC), above-ground biomass (AGB)], sensors indicators [Chlorophyll content (SPAD), LAI], and image-based color indexes (CIs). Four machine learning methods and multiple-step regression (MSR) were applied to explore the monitoring model and methodology of the tomato growth parameters. The results demonstrated that leaf-based indicators performed well in nitrogen related indicators [leaf nitrogen content and N nutrition index (NNI)] and AGB estimation (|r|>0.7). CIs were important indicators in tomato yield and VC prediction (0.2<|r|<0.4). Machine learning methods could improve the prediction accuracy while decreasing the error based on the data of one type of sensor value (R2>0.31) or CIs (R2>0.31). The mixture of sensor values and CIs could predict the tomato indicators accurately (0.87>R2>0.42). MSR method measured a higher R2 and low error in tomato indicators (Linear: 0.81>R2>0.14, 19.56>RMSE>0.02; MSR: 0.85>R2>0.42, 10.73>RMSE>0.01). Machine learning obtained higher accuracy and lower error by combining sensors’ indexes and CIs for monitoring tomato growth indicators than the single indicator (sensor index or CI, 0.87>R2>0.45). Significantly, the least absolute shrinkage and selection operator measured the relatively higher R2 (>0.73) and lower error (RMSE<4.71, RE<5.46) on most tomato growth indicators’ prediction. Altogether, this study suggested that machine learning is promising in tomato indicator estimations of greenhouse crops, which contributed to developing a series of models (sensor value-based, CIs-based, and sensor value &amp; CIs-based) for quickly monitoring tomato crop growth and predicting yield.