Acquisition of oilseed rape seedling population based on visible light imagery from unmanned aerial vehicles

Abstract

Seedling density of oilseed rape has a significant effect on seed yield and quality, and accurate and timely estimation of seedling density in the field can guide later field management to ensure high yield and oil quality is of great significance. Currently, object detection is the most commonly used seedling counting method, which counts seedlings by accurately identifying them, but it will lead to unavoidable errors when the target overlap rate is large. This study aimed to reduce this error by obtaining seedling counts through regression prediction constructing a model of shape characteristics and seedling counts of oilseed rape seedlings in the field. Ultra-high resolution RGB images were captured using an unmanned aerial vehicle (UAV) during the growth stage when the oilseed rape seedlings had at least two leaves. The study used three regression modeling methods, namely, fully connected neural network (DNN), random forest algorithm (RF), and multiple linear regression (MLR), to construct a regression prediction model, and the threshold segmentation method (OTSU) was used to segment oilseed rape seedling targets from the processed images of the ExG color vegetation index, and the segmented oilseed rape seedling objects were extracted and selected as the total area of the connective domain, tal_are, the total outer connective rectangle The total area of the connected domain tal_are, the total perimeter of the outer rectangle S, and the total height of the outer rectangle h-all were extracted and selected as model inputs. The results show that: (1) the relative error of prediction of regression modeling is 8.68 %, which is much lower than the 17.75 % relative error achieved by the SSD object detection model; (2) there is a strong correlation between the shape features of oilseed rape seedlings and the actual number of seedlings measured, and the DNN fully-connected neural network performs the best among the different modeling methods, and in the five-fold cross validation, its average R can reach 0.81, and the average MAPE of 7.22 %, RMSE of 11.61, and prediction error of 8.68 % determined that the DNN prediction model can better estimate the number of rape seedlings. In summary, this study establishes a practical and fast remote sensing counting method for the early growth of oilseed rape seedlings, reduces the error of object detection due to high overlap, and uses regression prediction method to invert the model in oilseed rape land to verify the validity of the model. This method provides a scalable solution for large-scale field monitoring, enabling precision agriculture practices such as optimized planting density and early yield prediction.