Simple, Low-Cost Estimation of Potato Above-Ground Biomass Using Improved Canopy Leaf Detection Method

Above-ground biomass (AGB) is one of the most important indicators for evaluating potato growth and yield. Rapid and accurate biomass estimation is of great significance to potato breeding and agricultural production. However, high cost, large data volume, and poor model scalability are the main problems of hyperspectral remote sensing and LiDAR in existing AGB measurement methods, especially in small-scale farmland. One of the important methods for solving the above problems is extracting canopy structure features through RGB images. In this study, a new AGB estimation method for potatoes at the field scale was proposed by using canopy leaf detection and digital images. First, using the improved feature fusion network and the soft intersection over union (soft-IoU) layer, an improved detection network of dense leaves, DenseNet-potato, was developed to detect canopy leaves. Second, the detection network was used to extract the canopy structural features, and the corrected number and total area of canopy leaves were obtained. Finally, multilayer perceptron (MLP) regression was introduced to build prediction models for AGB using canopy features. It was found that the DenseNet-potato network had excellent detection effects on dense canopy leaves. The mAP(50) and mAP(75) of the two detection pipelines reached 76.63% and 64.35%, respectively, which were 9.17% and 6.05% higher than the state-of-the-art RetinaNet method. In addition, the results indicated a strong correlation between the estimated and field-observed AGB using the MLP method from the digital camera dataset (R-2 = 0.83, RMSE = 0.039 kg/plot, NRMSE = 12.16%), while the unmanned aerial vehicle (UAV) dataset was unsatisfactory (R-2 = 0.62, RMSE = 0.051 kg/plot, NRMSE = 15.32%). This study can provide a reference for efficiently estimating potato AGB using RGB images.