Plantation Monitoring Using Drone Images: A Dataset and Performance Review

TL;DR

This paper tackles the problem of autonomous plantation health monitoring using RGB imagery captured from lightweight drones, addressing the lack of accessible UAV-based datasets. It introduces a CVAT-annotated dataset of 9,534 tree instances from 255 drone images across three health classes and benchmarks multiple CNN architectures, along with YOLO-based individual-tree detection. Findings indicate depth-wise convolution, as exemplified by XceptionNet, yields the best accuracy on this small UAV dataset, while YOLO-based detection achieves notable performance with pretraining. The publicly available dataset and benchmark offer a practical, low-cost tool for farmers in developing regions and establish a reference for future plantation-monitoring research.

Abstract

Automatic monitoring of tree plantations plays a crucial role in agriculture. Flawless monitoring of tree health helps farmers make informed decisions regarding their management by taking appropriate action. Use of drone images for automatic plantation monitoring can enhance the accuracy of the monitoring process, while still being affordable to small farmers in developing countries such as India. Small, low cost drones equipped with an RGB camera can capture high-resolution images of agricultural fields, allowing for detailed analysis of the well-being of the plantations. Existing methods of automated plantation monitoring are mostly based on satellite images, which are difficult to get for the farmers. We propose an automated system for plantation health monitoring using drone images, which are becoming easier to get for the farmers. We propose a dataset of images of trees with three categories: ``Good health", ``Stunted", and ``Dead". We annotate the dataset using CVAT annotation tool, for use in research purposes. We experiment with different well-known CNN models to observe their performance on the proposed dataset. The initial low accuracy levels show the complexity of the proposed dataset. Further, our study revealed that, depth-wise convolution operation embedded in a deep CNN model, can enhance the performance of the model on drone dataset. Further, we apply state-of-the-art object detection models to identify individual trees to better monitor them automatically.

Figures (6)

• Figure 1: Sample drone images in the proposed dataset. These images are annotated using CVAT annotation tool for individual tree identification.
• Figure 2: Sample annotations performed on the proposed dataset, using the CVAT annotation tool cvat. Two samples from each class are shown here: Column (a) Good, Column (b) Stunted, and Column (c) Dead.
• Figure 3: Train versus test accuracy and loss curves for AlexNet and VGG19 models: (a) AlexNet accuracy curve, (b) AlexNet loss curve, (c) VGG19 accuracy curve, (d) VGG19 loss curve. The huge differences between the train and test accuracy and loss indicate overfitting by both the models.
• Figure 4: Train versus test accuracy and loss curves for GoogleNet and EfficientNet models: (a) GoogleNet accuracy curve, (b) GoogleNet loss curve, (c) EfficientNet accuracy curve, (d) EfficientNet loss curve. The huge differences between the train and test accuracy and loss indicate overfitting by both the models.
• Figure 5: Train versus test accuracy and loss curves for VGG16 and ResNet50 models: (a) VGG16 accuracy curve, (b) VGG16 loss curve, (c) ResNet50 accuracy curve, (d) ResNet50 loss curve. The differences between the train and test accuracy and loss is less for VGG16, and even lesser for ResNet50 model.