卫星—无人机两层组网葡萄园巡检与NDVI自适应路径规划(原标题:Satellite-Enabled Two-Tier UAV Vineyard Inspection with NDVI-Adaptive Path Planning)

1. Introduction

Viticulture is a high-value agricultural sector with both economic and cultural significance in many regions. Vineyards are exposed to a range of pressures, including disease, water stress, and climatic variability. Efficient monitoring is essential for informing timely interventions and management decisions, while keeping labour costs and effort low. Automated and data-driven monitoring approaches offer an increasingly practical path to vineyard monitoring, combining sensing, communication, and analytics for decision support. Relevant technologies include satellite imagery and communications, drone mapping, multispectral vegetation indices, and computer vision.

Satellite technologies can support crop monitoring both directly, through Earth observation, and indirectly, through communication. Satellite imagery can provide repeated, wide-area observations of vineyard condition and has been used to support precision viticulture by mapping vegetation vigour, water status, spatial variability, and other relevant indicators for improved sustainability and operational efficiency [1]. As data-driven approaches gain traction in agriculture, satellite communication is also becoming increasingly relevant, especially for remote regions where terrestrial connectivity may be unreliable or unavailable [2].

Drones provide a flexible and high-resolution sensing platform for precision agriculture, complementing the wider coverage of satellite imagery with detailed observations for crop mapping, plant-health assessment, irrigation support, yield estimation, spraying, and other monitoring and management tasks [3]. Drone-based multispectral mapping can provide vegetation indices, such as NDVI, which are widely used to quantify vegetation condition, growth, and stress [4]. Previous vineyard-focused research has examined such data for specific applications including water-status variability mapping and disease progression monitoring [5,6]. Nevertheless, UAV monitoring involves an inherent trade-off between coverage and detail: high-altitude flights can map larger areas efficiently, while low-altitude inspection can provide more detailed visual information but requires additional flight time and produces larger data volumes. At a broader monitoring scale, drone imagery is therefore often considered alongside satellite imagery, combining wide-area observation with finer spatial detail [7].

Computer vision and machine learning further expand the role of image-based crop monitoring by enabling automated interpretation of RGB, multispectral, and other data. In agriculture, these methods have been applied to a broad range of tasks, including plant and canopy segmentation, stress detection, disease identification, and yield estimation. For disease monitoring specifically, recent reviews highlight the growing use of UAV imagery, deep learning, and remote or proximal sensing tools for early detection and field-scale assessment [8,9,10]. In grapevines, recent work includes both leaf-level disease classification using convolutional neural networks and field-level disease monitoring using UAV multispectral data [6,11].

Building on these developments, this work presents a satellite-enabled, two-tier UAV workflow for semi-automated vineyard inspection. The proposed approach combines high-altitude multispectral scanning, identification of points of interest from vegetation index analysis, live flight path adaptation, low-altitude RGB inspection, and vision-based disease detection, classification, and severity estimation. Processing tasks are offloaded to a remote server reached through an emulated satellite communication environment, allowing the UAV-side system to remain lightweight while receiving multispectral analysis results during the mission. These results are used to adapt the flight path and guide selective low-altitude inspection of areas that may indicate plant stress or disease.

Unlike existing approaches that typically address multispectral crop monitoring, close-range disease inspection, communication constraints, or UAV path planning separately, this work integrates these elements into a single adaptive inspection workflow. The main contributions of this paper are as follows:

• A two-tier UAV inspection architecture combining high-altitude multispectral sensing, NDVI-based smart sampling, low-altitude RGB inspection, and remote image processing.

• On-the-fly path adaptation for inserting newly identified waypoints into an ongoing UAV mission, together with a comparative evaluation of alternative insertion and route-optimisation strategies.

• An experimental analysis of how satellite-link bandwidth affects image-transfer latency, waypoint availability, and total mission duration during adaptive inspection.

The experimental evaluation in Section 3 focuses on the effect of different path-adaptation strategies and communication bottlenecks on mission efficiency. A detailed evaluation of point-of-interest identification as a smart sampling mechanism for low-altitude image acquisition is beyond the scope of this work and will be addressed in future research. The disease-detection component is included as a downstream analysis module based primarily on established computer vision approaches, and is therefore not benchmarked as a separate contribution in this paper.

2. Materials and Methods

2.1. System Architecture

The overall system architecture and information flow are shown in Figure 1. The drone captures the images from the vineyard and sends them to the remote server for processing. The system incorporates high-altitude multispectral scanning, dynamic mission adaptation, and computer vision-based disease assessment. It is deployed as two interacting parts:

Figure 1. System Architecture and data flow.

A locally deployed application that interfaces with—and controls—the drone: For missions with a real drone, this is an Android app leveraging an SDK to communicate with the drone and deployed on the drone remote controller; for the tests presented in this work, a simulator was used instead, to facilitate repeat testing, as described in Section 2.6.

A remote server that handles image processing and computer vision tasks: In this work, a server hosted in a VM within the satellite emulation testbed presented in Section 2.2 was used.

Furthermore, the system comprises three main software modules, each with a distinct input and output, as shown in Table 1.

Table 1. System software modules.

The flow of information between the modules, during the course of a mission, is as follows:

Initially, a mission is created according to user input, comprising only high-altitude waypoints, placed evenly on a grid so as to cover the whole vineyard. At each high-altitude waypoint, the drone captures a multispectral image. The local application sends the pair of Red and NIR bands to the remote server. There, the POI identification module (Section 2.3) receives and processes them, and sends back the coordinates of any POIs identified. The Path adaptation module (Section 2.4) inserts these POIs into the existing path, adapting and optimizing it to minimise its total length. When a POI is visited, the drone captures a close-up RGB photo of the grapevine. All RGB close-ups are sent to the remote server after the mission for Disease detection (Section 2.5).

The satellite constellation is transparent and serves the purpose of providing connectivity between the drone and the ground node hosting the remote server.

2.2. Satellite Emulation Testbed

The satellite emulation testbed is capable of emulating a configurable number of satellites in an orbital plane. In this work, it is used to reproduce the behaviour of a Low Earth Orbit (LEO) satellite constellation with Inter-Satellite Links (ISLs), providing the communication environment between the drone-side system and the remote processing server.

The testbed was developed by DLR for research purposes using Mininet [12], a network emulator that can create networks of virtual hosts, switches, controllers, and links. In contrast to purely analytical or discrete-event simulators, Mininet networks run real code, including standard Unix/Linux network applications and the Linux kernel network stack. As a result, software developed and tested in Mininet can be transferred to real systems with minimal changes, making it suitable for network virtualisation, performance evaluation, and deployment-oriented testing.

In the emulated satellite network, ground nodes connect to the nearest visible satellite. Depending on the constellation configuration, communication between two ground nodes may traverse more than one satellite, making ISLs important for end-to-end connectivity. The performance of the communication path depends on the links between ground nodes and satellites, as well as on the ISLs between satellites. For this reason, the testbed supports multiple configurable parameters, including the following:

• The height of the orbital plane, which affects both the delay between a ground node and the nearest satellite and the duration of satellite visibility with respect to the ground node;

• The number of satellites between the source and destination ground nodes;

• Ground-to-satellite link parameters, such as bandwidth and packet loss rate;

• ISL parameters, including delay, bandwidth, and packet loss rate.

Although the testbed can support two orbital planes, only one orbital plane was considered in this work for simplicity. The testbed can also emulate orbits other than LEO, provided that the relevant parameters are selected carefully to avoid unrealistic scenarios. Communication between nodes is based on Internet Protocol (IP) routing, and both Transmission Control Protocol (TCP) and User Datagram Protocol (UDP) are supported.

The experimental setup is shown in Figure 2. The emulated satellite network runs inside a Virtual Machine (VM) on Ubuntu 22.04, which is accessed through a Virtual Private Network (VPN) connection. The drone-side system connects to this VPN and sends data to the VM. Upon receiving the data packets, the VM forwards them to the satellite emulation framework. Packets are then forwarded through the emulated satellite path, from the nearest satellite toward the gateway or destination ground node. In this work, end-to-end communication was carried out using TCP, and four satellites were modelled along the path between the drone and the ground node. The ground node hosts the Docker container running the remote processing service. After the images are processed, the results are returned through the same emulated communication path in the reverse direction.

Figure 2. Experimental testbed configuration.

Satellite emulation is considering a Starlink-like constellation. Hence, the ground nodes have continuous connectivity since at least one satellite is visible. This ensures that the ground nodes can always communicate without the need of a storage mechanism on the ground nodes. Considering that TCP is employed in the transport layer it is essential that the end-to-end connectivity is always present. Otherwise, the docker present on the other end of the constellation will not be readily reachable, which can adversely affect the overall performance.

The links between satellites and ground nodes, namely the drone and gateway, are modelled based on the orbital height of the satellite constellation and the relative movement between satellites and ground nodes. Link delay is highest when the satellite appears at the horizon of the ground node. It then decreases as the satellite moves toward the point above the ground node, before increasing again until the satellite is no longer visible. Satellite visibility duration also depends on orbital height, with higher orbital planes resulting in longer visibility duration. In the scenario considered in this work, the orbital height was set to 1221 km.

2.3. POI Identification

2.3.1. POI Identification Rationale and Objective

Scanning an area, or a vineyard in the present case, is a trade-off between flight altitude and effective ground resolution. The ground area captured by a single photo (also known as the footprint) depends on the altitude of flight, as well as the camera’s field of view (FOV). For simplicity and uniformity, in this paper we are considering only photos taken at nadir orientation (i.e., the camera is pointing straight down).

For a camera with

$F O V_{H} \times F O V_{V}$
field of view and the drone flying at altitude A, the footprint is a rectangle of dimensions

$X \times Y$
:

$$
X & = 2 A tan \left(\right. F O V_{H} / 2 \left.\right) \ Y & = 2 A tan \left(\right. F O V_{V} / 2 \left.\right)
$$

(1)

The ground sampling distance (GSD) describes the distance between adjacent pixels and is equal to the footprint size divided by the camera resolution. For a camera with a

$W \times H$
resolution:

$$
G S D_{x} & = 2 A \frac{tan \left(\right. F O V_{H} / 2 \left.\right)}{W} \ G S D_{y} & = 2 A \frac{tan \left(\right. F O V_{V} / 2 \left.\right)}{H}
$$

(2)

For a given drone model, the camera specifications are constant and GSD is reduced to a function of altitude, which can be derived from either the horizontal or the vertical GSD and resolution, equivalently. For the camera of a modern commercial drone (in this work, experiments were conducted using the DJI Mavic 3M (DJI, Shenzhen, China)), with a vertical resolution of 3956 pixels and a corresponding vertical FOV of 48°, the altitude required to achieve a desired GSD is:

$$
A = \frac{1}{2} \frac{tan \left(\right. 48 / 2 \left.\right)}{3956} G S D \approx 4443 \cdot G S D
$$

(3)

Grapevine leaves average about 20 cm in size, while discolouration spots that can point to disease are often about 1 cm. In disease detection datasets such as [13,14,15], discolouration spots that signify disease are usually 7–20 pixels wide. Hence, targeting a similar resolution and assuming that spots would be clearly identifiable if they are at least 10 pixels wide, at a GSD of no more than 0.001 m per pixel would be required, pointing at a maximum altitude of about 4.4 m for disease detection. This GSD is also in line with relevant research on unhealthy leaf detection in the wild [16]. This corresponds to a footprint roughly 5.2 m × 3.9 m, or an area of 20 m 2, meaning that even a relatively small vineyard of 3–4 hectares would require thousands of photos to scan exhaustively.

Non-exhaustive scans, or sampling, can provide insights regarding disease presence while keeping the number of images, the drone flight time, and the energy expenditure manageable. POI identification’s goal is to maximise the efficiency of low-altitude image sampling by identifying the most interesting, from a plant-health perspective, points in the vineyard.

2.3.2. Approach

Previous research has shown that NDVI bears a strong correlation with plant health and can be used to identify relevant risks [6,17]. Crucially, NDVI-based insights do not rely on specific details but on the average value of a plant or vineyard area, and hence can be acquired with a greater GSD (e.g., 0.02–0.05 m per pixel), at higher altitudes. Accounting for the comparatively lower resolution of multispectral cameras, this corresponds to an altitude of 45–110 m.

Hence, NDVI images constructed from high-altitude multispectral photos can provide insights regarding plant health. In this work, such photos are used to drive a smart sampling scheme, in which spots are identified as points of interest (POIs) based on their NDVI value, and are marked for low-altitude image acquisition.

2.3.3. NDVI Calculation and Correction

The NDVI is computed from the Red and Near Infrared (NIR) bands:

$$
N D V I = \frac{N I R - R e d}{N I R + R e d}
$$

(4)

While the NIR and Red bands are acquired as part of the multispectral photo, they each have slightly different fields of view, due to the placement of the sensors on the drone. In addition, both images have slight vignetting distortion. Before computing the NDVI, band images must be undistorted and aligned:

To compensate for vignetting distortion, the vignetting polynomial coefficients are extracted from each image’s metadata and used. For alignment, Oriented FAST and Rotated BRIEF (ORB) [18] is used to detect keypoints and compute compact binary descriptors in the Red and NIR images. ORB was selected because it combines low computational cost with robustness to the small rotational differences between the two bands, making it suitable for repeated processing of high-resolution multispectral images, while its binary descriptors can be matched efficiently using Hamming distance. A brute-force matcher compares the descriptors between the two bands, with cross-checking applied to retain only mutual best matches. The matches are sorted by distance, and the best 50 correspondences are used to estimate an affine transformation that aligns the NIR band to the Red band.

2.3.4. Grapevine Row Identification and Segmentation

From each NDVI image depicting a patch of the vineyard, individual segments of plants, corresponding to potential POIs, must be identified and the mean NDVI of each must be calculated:

Firstly, the NDVI images show rows of grapevines as well as soil. To separate the grapevines from the background, a binary mask is applied to NDVI with a threshold of 0.35. This threshold is in line with previous research, which suggests thresholds between 0.35 and 0.4, depending on plant development [19]. Grapevine rows are identified as continuous areas in the binary mask. Finally, each row is divided into smaller segments. Segmentation is performed along each row’s long axis, by lines perpendicular