Biodiversity plays a crucial role in maintaining the stability and functionality of agroecosystems. Effective biodiversity management and sustainable agricultural production rely heavily on accurate identification and classification of plant species within plant communities. Traditional methods of species identification are often labor-intensive and time-consuming, presenting challenges in dataset acquisition, preparation, and model selection for image classification. Recent advancements in remote sensing technologies and machine learning have provided promising solutions to these challenges. This study explores the potential of using unmanned aerial vehicles (UAVs) to collect RGB images and employing transfer learning techniques to accurately classify plant species in heterogeneous plant areas.
The research presented in this paper introduces a comprehensive method combining object-based supervised machine learning for dataset preparation and a pre-trained transfer learning model, EfficientNetV2, for plant species classification. The study area is an agricultural property within Parco delle Cave in Brescia, Lombardy, Italy, covering approximately one hectare. UAVs equipped with high-resolution cameras captured images of the study area, which were then processed to create an orthomosaic photo and a digital surface model (DSM).
- Image Acquisition and Pre-Processing: The UAV flight was conducted using a Mavic 2 Pro UAV equipped with a Hasselblad L1D-20-megapixel camera. The images were captured at a height of 30 meters, covering the entire study area. A total of 319 images were processed using Agisoft Metashape software to generate an orthomosaic photo and a DSM.
- Object-Based Segmentation and Supervised Data Preparation: The orthomosaic image was segmented into meaningful image-objects using multi-resolution segmentation techniques. Essential parameters such as scale, shape, and compactness were fine-tuned to achieve optimal segmentation results. Training samples representing different plant species were selected, and the K-nearest neighbor (KNN) algorithm was used for supervised classification.
- Training Dataset Creation: Seven classes of plant species were identified for classification based on sufficient image data availability. A total of 1374 images were generated, with 70% used for training and 30% reserved for testing.
- Transfer Learning with EfficientNetV2: The pre-trained EfficientNetV2 model was employed for classification. Transfer learning allowed leveraging pre-trained model features and fine-tuning them for the new task, significantly improving classification accuracy even with a limited dataset. Image augmentation techniques were applied to increase the dataset size from 782 to 1374 images.
The EfficientNetV2 model achieved a remarkable 99% classification accuracy for seven plant species. Performance metrics such as precision, recall, and F1-score were also evaluated, demonstrating the robustness of the implemented approach. Comparative studies were conducted to contrast EfficientNetV2 with other widely used transfer learning models, including ResNet50, Xception, DenseNet121, InceptionV3, and MobileNetV2. EfficientNetV2 outperformed these models, highlighting its superior accuracy and computational efficiency.
This research addresses significant gaps in plant species classification by utilizing UAV-collected RGB images and advanced transfer learning techniques. The study underscores the importance of creating localized and specialized image datasets for accurate plant classification. The use of UAVs provides high flexibility, low cost, and real-time data acquisition, making them ideal for assessing diverse plant areas. The study’s findings contribute to the development of efficient methods for plant species identification and biodiversity management in agroecosystems.
The proposed method for plant species identification and classification using UAV-collected RGB images and transfer learning demonstrates high accuracy and efficiency. The study’s methodology is particularly valuable for ecological and agricultural studies, offering a robust tool for biodiversity monitoring and management. Future research should explore the applicability of this method across diverse geographical regions and environmental conditions to further enhance its robustness and generalizability.
Researcher Profile: Girma Tariku Woldesemayat
Girma Tariku Woldesemayat is a prominent researcher currently pursuing his Ph.D. in Information Engineering at the Università degli Studi di Brescia, Italy. Girma holds a master’s degree in Communication Technology and Multimedia from the same university, demonstrating a strong academic foundation in advanced communication systems and multimedia technologies. His academic journey reflects a deep commitment to leveraging technological advancements to address complex challenges in various domains.
#AI #MachineLearning #Biodiversity #SustainableAgriculture #EnvironmentalMonitoring #UAVTechnology #PhDResearch #InformationEngineering #SoilMonitoring #AgriculturalInnovation #TransferLearning #EcologicalStudies #TechForGood #SmartAgriculture #ConservationTech
Leave a Reply