Canopy height and biomass prediction in Mombaca guinea grass pastures using satellite imagery and machine learning

Abstract

Remote sensing can serve as a promising solution for monitoring spatio-temporal variability in grasslands, providing timely information about different biophysical parameters. We aimed to develop models for canopy height classification and aboveground biomass estimation in pastures of Megathyrsus maximus cv. Mombaça using machine learning techniques and images obtained from the Sentinel-2 satellite. We used different spectral bands from the Sentinel-2, which were obtained and processed entirely in the cloud computing space. Three canopy height classes were defined according to grazing management recommendations: Class 0 (< 0.45 m), Class 1 (0.45–0.80 m) and Class 2 (> 0.80 m). For modeling, the original database was divided into training data (85%) and test data (15%). To avoid dependency between our training and test datasets and ensure greater generalization capacity, we used a spatial grouping evaluation structure. The random forest algorithm was used to predict canopy height and aboveground biomass by using height and biomass field reference data obtained from 54 paddocks in Brazil between 2016 and 2018. Our results demonstrated precision, recall, and accuracy values of up to 73%, 73%, and 72%, respectively, for canopy height classification. In addition, the models showed reasonable predictive performance for aboveground fresh biomass (AFB) and dry matter concentration (DMC; R2 = 0.61 and 0.69, respectively). We conclude that the combined use of satellite imagery and machine learning techniques has potential to predict canopy height and aboveground biomass of Megathyrsus maximus cv. Mombaça. However, further studies should be conducted to improve the proposed models and develop software to implement the tool under field conditions.