What if regenerative farming could be monitored by a team of different drones that learns to fly safely, coordinate its tasks, and send field data through a satellite-linked network? This paper builds a proof of concept for that idea. Regenerative farming is regarded as the next evolution in the agricultural sustainability cycle, covering soil health, plant diversity and biodiversity, climate-change mitigation, and sustainable food production.

That goal requires both a change in farmer attitudes and perceptions and the development and integration of new technologies into farming. In smart farming, UAVs can monitor crop health and yield, assess soil quality, image hard-to-reach zones, and plant seeds fired into the ground from the air.

UAVs are also increasingly used for irrigation management, especially where water pooling or capping makes irrigation levels difficult to assess from the ground. The challenges include airspace regulations, privacy and safety, farmer training, deployment costs, data management, battery life, and the lack of GPS in some farmland areas.

The paper also identifies a structural limitation: UAV fleets are usually homogeneous, meaning they comprise the same type of UAV. With many cooperative UAVs in the sky, collision avoidance becomes a central issue. Route planning alone is therefore not enough; operational success also requires managing synchronization, speed, altitude, and mission time alongside the route.

Clustering heterogeneous UAVs raises issues of autonomy and synchronization, so an intelligent operational brain is presented as necessary for collective and cohesive behavior. The proposed framework is described as two machine-learning brains, together with mathematical formulations and link-budget predictions.

The workflow is a four-stage process: first, the framework makes the UAVs fly autonomously and in synchronization while scheduling task priorities. Second, MAVLink coordinates the leader-follower topology and supports data exchange, flight control, safety, and routing to the UAV leader’s action plan.

A LEO satellite coordinates multiple fleets equipped with MIMO-LoRa antennas to improve data-rate transmission and link reliability. In stage three, a gateway node receives underground WSN data such as soil moisture, temperature, humidity, and water levels, then sends it through the UAV fleet to the leader and LEO satellite.

In stage four, the LEO satellite transmits the data to the cloud through a GCC for storage and data analytics. The first brain enables UAV fleet autonomy using DRL and is called the DRL brain. The second brain enables UAV synchronization and task scheduling using ACO and is called the ACO brain.

The DRL brain’s primary mission is safe flight without human intervention, enabling the autonomous UAV fleet to manage unexpected and unpredictable emergencies. Its three main parameters are deep agents for learning the best course for rewards, states, and actions; flight-path planning; and proximity information for avoiding obstacles with proximity sensors and front-facing cameras.

The ACO brain’s mission is UAV fleet synchronization and task scheduling, enabling the fleet to carry out scheduled tasks while flying in formation. Its three main parameters are multi-agents, social learning, and dynamic leader selection.

Within a self-organizing swarm, every UAV learns and adapts to its surroundings, motion, and speed, while a leader and followers synchronize as a colony before scheduled tasks begin. Figure four shows a Mission Planner interface guiding a UAV across a regenerative farm in a yellow strip-pattern search.

The numbered markers indicate the search begins at one and ends at two, with the flight path covering the visible field in repeated passes. This July twenty twenty-three trial in Taif, Saudi Arabia, was flown at an altitude of one hundred twenty meters with a footprint reach of one point five square kilometers, supporting monitoring of soil-restoring agriculture alongside food production.

Each colony’s leader establishes direct communication links with leaders in other colonies to support autonomy, synchronization, and task scheduling. Each UAV colony targets one crop, aiming to improve monitoring time and power consumption, while a control signal transfers data between colony leaders and the LEO satellite.

The time needed for a complete sweep is relatively similar across all fleets, at fourteen point two minutes. Figure six shows the DRL brain network’s average discounted reward across training episodes. The curve rises from near zero, then levels off around roughly zero point two-five to zero point two-seven, while the shaded region and lighter fluctuations indicate variation across repetitions.

The authors describe this reward as being adjusted through trial and error, with training devoted to obstacle avoidance and safe flying; rewards reset to zero at the start of each cycle. The ACO brain network’s convergence is shown through average path cost as a function of iterations.

The figure reveals a consistent average path cost and highlights the efficacy of dynamic leader selection, social learning, and multi-agents. Figure nine plots mean squared error across training, testing, and validation over forty-nine epochs, with the curves falling sharply at first and then leveling off.

The validation curve reaches its best recorded performance at iteration forty-three, marked on the chart, while training stops when the error begins to increase. The authors use this pattern to argue that the model’s errors remain small and that no major overfitting occurs before the selected stopping point.

The proof of concept combines overground and underground sensors in a smart farm and presents the setup as space and ground segments. The space segment contains UAV colonies with cameras, proximity sensors, and five-G MIMO-LoRa antennas that receive data from the wireless gateway and transmit it to the LEO satellite, with the data destined for the cloud.

The ground segment includes the GCC, sensors measuring temperature, humidity, rain, and solar radiation, and a wireless LoRa gateway server transmitting sensor data to the UAVs. Further validation evaluates link quality between the UAVs and wireless gateway using the power spectral density ratio, written as Eb over No, and bit error rate.

The evaluation considers a Gaussian distribution channel in relation to altitude, frequency, transmission power, antenna gains, bit rate, and distance. The framework responds to battery depletion and lack of GPS in homogeneous UAV fleets by setting up a fleet of heterogeneous UAVs.

Heterogeneity is intended to prevent all UAVs from experiencing battery depletion at the same time, while including a LEO satellite is intended to prevent loss of connectivity. A MIMO-LoRa antenna is used to ensure high data rate and link reliability.

The paper combines heterogeneous UAVs, DRL, ACO, IoT, and MIMO-LoRa into an autonomous regenerative-farming framework, then validates parts of it through simulation and a practical proof of concept.