Reason: Patents pending
until file(s) become available
EVALUATING GENERALIZABILITY OF DEEP LEARNING NETWORKS FOR WEED IDENTIFICATION AND THEIR OPTIMIZATION FOR DEPLOYING ON EDGE DEVICE
Recent studies have indicated the need for agriculture automation to reduce the excessive use of herbicides, its detrimental effects on the environment, and the rise of herbicide-resistant weeds. In addition, the predicted increase in population to 9.7 billion by 2050 requires an increase in food production that cannot be met without proper weed control. At the current forefront of automation, deep learning has achieved high accuracy for weed classification and localization. Limited efforts have gone into developing and evaluating the generalizability of deep learning networks (DLN) over crop growth stages and weed heights. Further, even less work has been done for optimizing these computation-hungry DLN so that they could be deployed on lightweight edge devices for potential integration with an Unmanned Aerial System (UAS). Hence, in this research, the generalizability and deployment ability of four DLNs were evaluated for two computer vision tasks, i.e., object detection and image segmentation. For each task, the best performing network was optimized on two edge devices, namely, NVIDIA® Jetson NanoTM and NVIDIA® Xavier NXTM. Finally, studies were conducted to determine the edge devices’ frame rate for weed identification. For image segmentation, neither DeeplabV3+ nor UNet could generalize accurately for early season weed identification. For object detection, the YOLOv4 network trained on the V1 growth stage of soybean and 7.62 cm average weed height (AWH) of Palmer amaranth generalized the best with generalizability mean average precision score of 70.33 %. When optimized using tensorRT (floating-point precision of 16) on the edge devices, YOLOv4 resulted in 4.6 fps on Jetson Nano and 27.8 fps on Xavier NX, resulting in the highest fps achieved on an edge device for weed identification from UAS images. This research has resulted in developing foundational data, identifying promising deep learning-based algorithms, and evaluating edge devices that could lead to designing a real-time weed identification and UAS-based smart weed management system.