Smart greenhouses have emerged as a promising solution to improve the efficiency and sustainability of agricultural practices. By incorporating advanced technologies such as sensors, automation systems, and machine learning algorithms, smart greenhouses can provide optimal growing conditions for crops while minimizing the use of resources such as water, energy, and fertilizers. In this work, the monitoring of relative Chlorophyll content (RCC) and leaf water content (LWC) using leaf-level spectral signature of cherry-tomato leaves through index and empirical PLSR estimation approaches inside three different greenhouses, namely a polycarbonate greenhouse (PCGH-Eva), a greenhouse covered in radiative cooling sheet (RCGH-Eva), and the shaded greenhouse (SGH) was investigated. 96 leaf samples were collected from the PCGH-Eva and the RCGH-Eva alike and 48 ones from the SGH totaling 240 leaf samples with their in-situ spectral signatures, LWC, and RCC. Four modelling cases were discussed, one dedicated to each greenhouse and a Universal one made up of the combined samples of all the greenhouses. For the index-based approach, the robustness of published 23 leaf Chlorophyll-sensitive and 30 leaf water-sensitive indices were investigated. A ranking system was made where the best index is the one that has the least numerical sum of ranks. The most robust Chlorophyll indices that showed the best estimation performances across all modelling cases were the indices that relied on the off-Chlorophyll absorption center wavebands near the red-edge range (680-750nm). The most robust RCC index was the mND705nm. On the other hand, the performance of the leaf water-sensitive indices was deemed unrobust against the change in the types of greenhouses. This was mainly due to their fixed reliance on certain water-sensitive bands in NIR(701-1300nm). This region also reflects the health of the cell structure of the leaf. For the case of the RCGH-Eva, this region was spectrally poor due to the used covering and no good performance was made. For the PCGH-Eva, it was spectrally clear, however, since its leaves were in a bad shape and their cell structure was bad, their NIR region was poor resulting in no good performance. The SGH was the only modelling case that responded positively to the LWC indices since its leaves have clear signatures and are good structure. For the use of the partial least squares regression (PLSR) in the modelling of RCC and LWC, this modelling method was done in two stages, the first one was to find the best modelling scenario for each one of the four modelling cases and the second was to reduce the size of the models to 10% of their original size by using only the feature wavebands which are the most 5% positively correlated and the most 5% negatively correlated ones. For the first stage, each modelling case was done in 9 different modelling scenarios. The input regions were the VIS (350-700nm), VIS+NIR(350-1300nm), and the full range for the RCC estimation across all modelling cases whereas the input regions for the LWC modelling were the NIR+SWIR(>700nm), the SWIR alone (>1300nm), and the full range. The three spectral transformation states used per used region were the original/ no-transformation state, the area normalization transformation, and the Stavisky-Golay 1st derivative (SG1) over the area normalized spectral signatures. For the second stage, many feature RCC wavebands were observed with the sense of correlation in between parentheses given as follows; 370nm(-), 390470nm(+), 660nm(+), 690nm(-), 720nm(+), 1450nm(-), and 1960-2060nm(–&+) and the LWC ones were at: 400nm(+), 550nm(+), 700nm(-), 760nm(+), 1470nm(-), and 1650nm(+), >2000nm(– &+).Based on the resulting R2 values, the optimized PLSR models performed the same or slightly better than the index-based approach regarding RCC modelling and were superior to them in the LWC modelling across all modelling cases. Consequently, for monitoring purposes, the index-based RCC models were used to monitor chlorophyll content of leaves instead of the RCC-PLSR ones due to their mathematical simplicity and their comparable performance to that of the RCC-PLSR ones. However, for LWC% monitoring purposes, the LWC-PLSR models were chosen. The only modelling case that gave inadequate performance was the RCGH-Eva one due to its radiative cooling covering which blocked 67% of the total incidental solar irradiance. The monitoring capabilities of the chosen index-based RCC and the optimized LWC-PLSR models were used to monitor 8 leaves for 15 days– 4 leaves from the RCGH-Eva, 2 from the PCGH-Eva and 2 from the SGH. The models did predict the RCC% and the LWC% of the with high accuracy when the leaves were still somewhat green (i.e., RCC% ≥20% and LWC ≥ 60%) and with lower accuracy with leaves that were more yellowish to brownish since such leaves at that state were not used while constructing the database. Finally, the use of spectral means to auto-estimate plants’ vitals planted at the RCGH-Eva is not advisable. Hence, the appropriate method for auto water content status monitoring of its plants should not depend on any non/remote electromagnetic means. In such a case, the use of soil moisture content sensors (SMCS) is advised as a means to directly estimate the water status of the plants. The auto-estimation of Chlorophyll for this type of greenhouse remains challenging. Therefore, the use of a handheld Chlorophyll device like the one in this work with a human operator is recommended. Conclusively, for polycarbonate-based greenhouses and shade houses in agriculture, there is no conceptual nor practical reason why the work presented here cannot be expanded further or reproduced elsewhere to make use of leaf-level spectral signatures in estimating Chlorophyll and water content of plants.
| Date of Award | 25 Nov 2024 |
|---|
| Original language | American English |
|---|
| Supervisor | Maryam Alshehhi (Supervisor) |
|---|