Design and Field Validation of a Modular Vision-Guided UAV System for Real-Time Adaptive Vegetative Restoration

Vegetative restoration in degraded landscapes requires deployment strategies that can scale while adapting to heterogeneous terrain conditions. Conventional aerial seeding is typically performed in an open-loop manner, where seeds are distributed uniformly without accounting for local suitability for plant establishment. This paper describes a modular, unmanned aerial vehicle (UAV)-independent system for vision-guided aerial seeding, integrating onboard sensing, embedded processing, and real-time actuation within a closed-loop framework. The system combines a downward-facing visible-spectrum camera, a lightweight embedded computing unit, and a custom seed-dispensing mechanism organized in a perception–decision–actuation pipeline. Terrain suitability is evaluated in real time using three convolutional neural network (CNN) models and a conventional color-based greenness ratio method, enabling classification of sowable and non-sowable areas based on soil exposure, vegetation density, and obstacle presence. A confidence-based decision strategy, combined with temporal filtering, reduces noisy measurements, while an altitude-adaptive pulse-width modulation (PWM) controller regulates seed release to maintain a target seed density across varying flight heights. Field experiments conducted under semi-arid conditions show that terrain classification accuracy exceeds 85%, with inference latency below 100 ms per frame on an embedded Jetson Nano platform. In addition, the proposed control strategy maintains consistent seed density across different altitudes. These results indicate that onboard perception can be effectively coupled with adaptive aerial actuation, enabling more selective and efficient UAV-based vegetative restoration.