Revolutionizing Sugarcane Growth Monitoring: High-Throughput Phenotyping Through UAVs and Machine Learning for Enhanced Decision-Making

Sri Lanka's sugarcane industry plays a crucial role in the country's economy. However, comprehensive decision-making, including breeding, is hindered by a lack of growth data, primarily due to the extensive cultivation that limits resources available for collecting data through traditional ground-level methods. To overcome these limitations, we adopted a novel approach using UAV images for the assessment of plant growth over large cultivation areas which enables efficient monitoring and analysis by developing a Crop Surface Model (CSM) generated via the Digital Elevation Model (DEM). The experiment was conducted at Lanka Sugar Company (Pvt.) Ltd. in Pelwatte, Sri Lanka. Twelve experimental plots were established following the basic ridge and furrow land preparation, cultivating the SL 96 128 Sugarcane variety. A UAV with a multispectral sensor was used to capture images the sixth month after planting. On the same date, sugarcane plant height was measured in 144 sugarcane plants manually. Image processing was carried out using QGIS software For Generate DEM, and Agisoft Software was used for orthomosaic map generation. CSM, derived from DEM, was established to assess plant heights across 144 data sets. Their accuracy was verified by comparing them to plant heights measured manually. Results demonstrated a significant correlation between CSM height and manually collected height data, as evidenced by Pearson correlation coefficients in the 6th month (r = 0.873) after planting. The study employed three machine learning techniques, namely Simple Linear Regression, Support Vector Regression, and Random Forest Regression. Results indicated that Random Forest Regression exhibited the highest accuracy, boasting an R2 value of 81.22% and an RMSE of 3.354. Results revealed that multispectral UAV images can be used to predict sugarcane heights with good accuracy and further analysis will be carried out to test the accuracy of yield predictions.