Vertical green living walls (VGWs)—growing plants on vertical walls inside or outside buildings—have been suggested as a nature-based solution to improve air quality and comfort in modern cities. However, as with other greenery systems (e.g., agriculture), managing VGW systems requires adequate temporal and spatial monitoring of the plants as well as the surrounding environment. Remote sensing cameras and small, low-cost sensors have become increasingly valuable for conventional vegetation monitoring; nevertheless, they have rarely been used in VGWs. In this descriptive paper, we present a first-of-its-kind remote sensing high-throughput monitoring system in a VGW workplace. The system includes low- and high-cost sensors, thermal and hyperspectral remote sensing cameras, and in situ gas-exchange measurements. In addition, air temperature, relative humidity, and carbon dioxide concentrations are constantly monitored in the operating workplace room (scientific computer lab) where the VGW is established, while data are continuously streamed online to an analytical and visualization web application. Artificial Intelligence is used to automatically monitor changes across the living wall. Preliminary results of our unique monitoring system are presented under actual working room conditions while discussing future directions and potential applications of such a high-throughput remote sensing VGW system.
Leaf water potential (leaf) is a key indicator of plant water status, but its measurement is labor-intensive and limited in spatial coverage. While remote sensing has emerged as a useful tool for estimating vegetation water status, leaf remains unexplored, particularly in mixed forests. Here, we use spectral indices derived from unmanned aerial vehicle-based hyperspectral imaging and machine learning algorithms to assess leaf in a mixed, multi-species Mediterranean forest comprised of five key woody species: Pinus halepensis, Quercus calliprinos, Cupressus sempervirens, Ceratonia siliqua, and Pistacia lentiscus. Hyperspectral images (400–1000 nm) were acquired monthly over one year, concurrent with leaf measurements in each species. Twelve spectral indices and thousands of normalized difference spectral index (NDSI) combinations were evaluated. Three machine learning algorithms—random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM)—were used to model leaf. We compared the machine learning model results with linear models based on spectral indices and the NDSI. SVM, using species information as a feature, performed the best with a relatively good leaf assessment (R2 = 0.53; RMSE = 0.67 MPa; rRMSE = 28%), especially considering the small seasonal variance in leaf ( = 0.8 MPa). Predictions were best for Cupressus sempervirens (R2 = 0.80) and Pistacia lentiscus (R2 = 0.49), which had the largest leaf variances ( > 1 MPa). Aggregating data at the plot scale in a ‘general’ model markedly improved the leaf model (R2 = 0.79, RMSE = 0.31 MPa; rRMSE = 13%), providing a promising tool for monitoring mixed forest leaf. The fact that a non-species-specific, ‘general’ model could predict leaf implies that such a model can also be used with coarser resolution satellite data. Our study demonstrates the potential of combining hyperspectral imagery with machine learning for non-invasive leaf estimation in mixed forests while highlighting challenges in capturing interspecies variability.