Use of synthetic images for training a deep learning model for weed detection and biomass estimation in cotton.

Site-specific treatment of weeds in agricultural landscapes has been gaining importance in recent years due to economic savings and minimal impact on the environment. Different detection methods have been developed and tested for precision weed management systems, but recent developments in neural networks have offered great prospects. However, a major limitation with the neural network models is the requirement of high volumes of data for training. The current study aims at exploring an alternative approach to the use of real images to address this issue. In this study, synthetic images were generated with various strategies using plant instances clipped from UAV-borne real images. In addition, the Generative Adversarial Networks (GAN) technique was used to generate fake plant instances which were used in generating synthetic images. These images were used to train a powerful convolutional neural network (CNN) known as "Mask R-CNN" for weed detection and segmentation in a transfer learning mode. The study was conducted on morningglories (MG) and grass weeds (Grass) infested in cotton. The biomass for individual weeds was also collected in the field for biomass modeling using detection and segmentation results derived from model inference. Results showed a comparable performance between the real plant-based synthetic image (mean average precision for mask-mAPm: 0.60; mean average precision for bounding box-mAPb: 0.64) and real image datasets (mAPm: 0.80; mAPb: 0.81). However, the mixed dataset (real image  + real plant instance-based synthetic image dataset) resulted in no performance gain for segmentation mask whereas a very small performance gain for bounding box (mAPm: 0.80; mAPb: 0.83). Around 40-50 plant instances were sufficient for generating synthetic images that resulted in optimal performance. Row orientation of cotton in the synthetic images was beneficial compared to random-orientation. Synthetic images generated with automatically-clipped plant instances performed similarly to the ones generated with manually-clipped instances. Generative Adversarial Networks-derived fake plant instances-based synthetic images did not perform as effectively as real plant instance-based synthetic images. The canopy mask area predicted weed biomass better than bounding box area with R2 values of 0.66 and 0.46 for MG and Grass, respectively. The findings of this study offer valuable insights for guiding future endeavors oriented towards using synthetic images for weed detection and segmentation, and biomass estimation in row crops.

Evaluating Cross-Applicability of Weed Detection Models Across Different Crops in Similar Production Environments.

Convolutional neural networks (CNNs) have revolutionized the weed detection process with tremendous improvements in precision and accuracy. However, training these models is time-consuming and computationally demanding; thus, training weed detection models for every crop-weed environment may not be feasible. It is imperative to evaluate how a CNN-based weed detection model trained for a specific crop may perform in other crops. In this study, a CNN model was trained to detect morningglories and grasses in cotton. Assessments were made to gauge the potential of the very model in detecting the same weed species in soybean and corn under two levels of detection complexity (levels 1 and 2). Two popular object detection frameworks, YOLOv4 and Faster R-CNN, were trained to detect weeds under two schemes: Detect_Weed (detecting at weed/crop level) and Detect_Species (detecting at weed species level). In addition, the main cotton dataset was supplemented with different amounts of non-cotton crop images to see if cross-crop applicability can be improved. Both frameworks achieved reasonably high accuracy levels for the cotton test datasets under both schemes (Average Precision-AP: 0.83-0.88 and Mean Average Precision-mAP: 0.65-0.79). The same models performed differently over other crops under both frameworks (AP: 0.33-0.83 and mAP: 0.40-0.85). In particular, relatively higher accuracies were observed for soybean than for corn, and also for complexity level 1 than for level 2. Significant improvements in cross-crop applicability were further observed when additional corn and soybean images were added to the model training. These findings provide valuable insights into improving global applicability of weed detection models.