Slider
Slider
Slider
Slider

Adigar - A Drone Simulator for Agricultural Pest Control that Simulates Planning and Spraying Processes . . .

Features


Simulating robustly:


  1. a two dimensional grid world of an arable land from a collage of aerial/satellite photographs available,

  2. flight of a swarm of autonomous drones amidst different weather conditions,

  3. synthesised optimal planning policies over the arable land,

  4. generated predictive maintenance policies for drone battery and pesticide liquid,

  5. safe optimal pesticides spraying.

Problem


Adigar is a drone simulator specifically developed for agricultural pest control using swarms of drones. Adigar collectively assists simulating planning and spraying processes. Adigar is developed to address the following research question. "How to develop a swarm of drones solution for spraying pesticides in arable lands with optimal safe pesticide usage and minimum human intervention?" Adigar is the simulation wing and is part of the overall research approach we follow to solve the said research question.


Optimal pesticide usage corresponds to the correct dosage of a given pesticide as recommended by a national or international public health agency. Spraying pesticides manually takes up a considerable amount of time and requires a large human effort. One may utilise a single drone to spray pesticides appropriately with minimum time and human intervention. However, we suggest a swarm of drones here, since a single drone may prove to be highly inefficient in covering a large area of land. Before moving on with real drone hardware or gear or environment, Adigar can be used to completely plan and execute pesticide spraying on a virtual simulated environment, thereby assisting stakeholders to gauge the effectiveness of our research approach.

Significance


Tenants living near commercial farms face distinct health problems such as abdominal pain, nausea, vomiting, bloody diarrhea, headache, dizziness, drowsiness, weakness, lethargy, delirium, shock, kidney insufficiency, neuropathy, etc., due to overexposure to pesticides. Our research approach suggests using a multitude of drones that support autonomous refilling from a pesticide reservoir and recharging batteries using charging bay(s) located on site. We believe that our solution will minimise human intervention in the spraying process and effectively contribute to reducing the risk of contracting diseases that result from overexposure.


The suggested research approach is geared towards optimal pesticides usage, which can assist to reduce pesticide overuse, decrease pesticides runoff to surface water and groundwater. This will help to safeguard aquatic life and the purity of surface and groundwater. As said earlier, it will also help to lessen the damage to the micro-organisms in soil and reduce the rate of driving soil to a non-fertile state. Therefore, the approach can contribute to minimising the contamination of the environment. On the other hand, optimal pesticides usage will help increase the quality and the quantity of the harvest and save money otherwise spent on an excessive amount of pesticides.


When spraying pesticides manually, it is bound to take a considerable amount of time to cover the target area but a single spraying drone can fly over and cover a large arable area more speedily than a human. Furthermore, a swarm of spraying drones can cover a much larger area more speedily than a single spraying drone. Thus, we believe that the proposed solution concept will progressively help save time and the money of the user.


A typical commercial agricultural drone costs at least $4,000. If we assume a swarm of drones consisting of at least 3 such drones, the associated cost will be at least $12,000 ($4,000 x 3). A usual farmer may not be able to bear such an exorbitant cost. This motivates to build a cost-effective fully customised swarm of drones. It is expected to employ agent-based modelling techniques to solve the overall problem.


The benefits of the suggested approach are:


  1. Minimise human intervention in the spraying process.

  2. Contribute to reducing the risk of contracting diseases that result from overexposure.

  3. Reduce pesticide overuse.

  4. Minimise contamination of the environment.

  5. Increase the quality and the quantity of harvest.

  6. Save money spent on an excessive amount of pesticides.

  7. Save time and the money of the user.

  8. Enable farmers to use the resulting solution without the need to incur unreasonable costs.

  9. A better and safer platform for the pesticide industry to market different pesticides brands.

Challenges


Described in this section, several open challenges that reign in the paradigm of our research.


Battery Lifetime

A drone is powered by a rechargeable battery. Usually, a drone has a limited flying time due to limited battery lifetime. DJI is one of the top-rated drone producers in the world. Even the battery life of a DJI Phantom 4 drone for a normal flight is around 23 - 24 minutes and takes a considerable amount of time to recharge. When managing a customised swarm of drones, it is highly important to consider the status of their batteries and finding ways to recharge them efficiently.


Load Balancing

A single slave drone (slave drones will be discussed under Design) should carry a payload of liquid pesticides for crop spraying but the maximum weight that a drone can carry depends on many factors including any requirements on its speed and stability.


Covering Area

As per the scope of this research, we consider arable land areas of no more than one acre that require spraying a chosen common pesticide generally available in the Sri Lankan market. Furthermore, we consider avoiding unsafe areas which are traps that if reached, the drone may not be able to get out; and forbidden areas including dams, ponds, wastelands, streams, private properties, dead zones, etc. We are also interested in considering government regulations, guidelines from national and international public health agencies, and other international standardisation requirements for both operations of drones and safe pesticide usage.


Capturing Images of the Entire Area

Capturing images of an area up to one acre in one go is difficult for a single drone. When using a multi-UAV system for capturing images, another issue arises over the way images should be captured without having a single portion of land captured in two or more images.


Camera Quality

The quality of the drone camera is another crucial factor in aerial monitoring drones. In here, the quality of the camera corresponds to capturing clearer images of the farmland.


Additionally, there are several other obstacles such as involving onboard sensors, embedded processing, communicating through wireless links and limited sensing coverage. Robustness, flexibility, navigation and communication with other drones are some of the important features that drones in a swarm should entail.


Unemployment

Many people who live near the commercial farms depend on undertaking secondary support jobs related to farming such as pesticides spraying. Since the aim of this research is to introduce a swarm of drones solution to spray optimal amounts of pesticides with minimum human intervention, there exists a risk factor that may cause such personnel to lose their jobs.

Methodology


Herein outline our proposed methodology.


Development of a Solution Concept

An overview of the proposed system is shown in the following figure. The plan for a robust development of viable solution concept is outlined as follows:


  1. Develop a drone (discovery drone), to mark the boundary of the selected farmland for spraying a designated pesticide. The drone should consist of a global positioning system (GPS) sensor for geo-location identification, a general packet radio service (GPRS) sensor for communication via internet, a camera to capture pertinent images and a charging port for recharging its battery from a charging station.

    Rather than using a discovery drone, we also consider using a handheld GPS sensor for marking the boundary of the farmland as an alternate solution. Then the farmer can walk along the boundary of the farmland to capture boundary geo-locations.

  2. Develop a mobile application (app-A), to facilitate the farmer to connect with the above described discovery drone to mark the boundary of target farmland before spraying can commence. This application should also be able to monitor the flying path of the drone.

  3. Develop several comm-drones (similar in operation to the discovery drone) and associate with each, a set of slave drones to create clusters of workgroups each guided and controlled by a single comm-drone. Comm-drones have collision detection and avoidance mechanisms that support the whole cluster to detect and avoid such complications.

    Slave drones are primarily worker drones that actually perform the spraying. Each slave drone consists of a GPS sensor for geo-location identification, a spray tank and a battery charging port. More details are provided in Design.

  4. Develop a cloud-based (central) system to centrally manage the swarm operations and synthesise policies for optimising safe pesticide spraying. After the discovery drone completes its flight, boundary geo-locations are uploaded to the central system.

  5. The central system commands all comm-drones to fly over the farmland and capture ground images. The flying paths are determined by a planning algorithm. Develop a second mobile application (app-B), for connecting the farmer with the central system.

  6. Captured images are subsequently uploaded to the central system.

  7. With the images at hand, the central system generates the terrain of the farmland and subsequently produces the 2D grid world of the terrain via image processing techniques. This will identify and locate safe areas (e.g. areas with crops) and unsafe areas (e.g. traps, dams, ponds, wastelands, streams, etc.,) in the terrain.

  8. The central system plans by dividing safe areas to clusters according to the available number of comm-drones. Each cluster consists of a set of squares and each square consists of some circles. A capture of a sample farmland is shown in the following figure. The size of a circle is the maximum area that a slave drone can cover at a given time instance, when spraying pesticides.

  9. The central system assigns each comm-drone and several slave drones to a single cluster.

  10. The farmer provides user input to the central system via app-B. User input may include the total amount of spray liquid available, specific weather conditions, selecting a specific standardised procedure (e.g. WHO recommendation), etc. Using the given user input and already available data, the central system synthesises policies for the behaviour of the entire swarm, inside clusters and that for each slave.

  11. Apart from spraying pesticides, any slave drone can proceed to the charging or filling stations when needed by informing its associated comm-drone, the comm-drone subsequently informs the central system. The comm-drone then receives instructions to assign the former slave drone's remaining workload to another slave in the cluster. Inter-cluster slave transfers can also occur. If a comm-drone wants to perform similar maintenance tasks, it should also inform the central system. The central system then assigns the salves in that cluster to that of another comm-drone, if the former comm-drone could not finish its task before pursuing maintenance.

    After a slave drone has finished recharging or refilling, it uses the uplink in the station to inform the central system. After a comm-drone completes its tasks in the associated cluster, the central system assigns to it another safe area.

  12. After finishing all spray jobs in the selected farmland all drones are directed to a hangar.

Overall Workflow

The following figure illustrates the overall workflow of our proposed solution. Initially, we capture an aerial/satellite image of the farmland. Then, a 2D grid world of the farmland is generated using our 2D Grid Maker procedure which incorporates a Faster R-CNN for object detection. Purple, green, red, and blue nodes on the grid illustrate starting position, goal states (areas where we want to spray pesticides), unsafe and safe areas respectively. Afterwards, we perform offline path generation with the help of our drone simulator: Adigar to generate the drone flying path on the grid. For the purpose of succinctness, the following figure only presents the flight path of a single drone in yellow colour.


During the actual flight of the swarm, online planning is performed. It has three sub planning procedures: predictive maintenance, shielding and spraying pesticides. Shielding is specifically used here to learn a backup controller online to override (as needed) the controller learned during the offline path generation step, to tackle real-world adversaries such as wind, battery and pesticide liquid depletion, etc.


The below figure shows how a spray job is coordinated. The central system divides goal states to clusters according to the available number of comm-drones. Each cluster (which is usually administered under a single comm-drone) consists of a set of zones and each zone consists of some minimum spraying areas (MSAs). The size of a MSA is the maximum area that a slave drone can cover at a given time instance, when spraying pesticides. Therefore, several slave drones will usually be managed by a single comm-drone.


Moreover, all the drones we use consist of sensors which can detect battery level, rain, wind speed with direction. They also have collision detection sensors, global positioning system (GPS) sensors and general packet radio service (GPRS) connections.


Training Users

Target user group of this research is farmers. In the training phase, farmers will be trained to basically handle a smartphone, install and use aforementioned mobile applications. Also, mobile applications will inform farmers the optimal location to place both charging and filling stations.


Evaluation

On the whole evaluation will be based on empirical studies. We plan on collecting feedback from the target user group via interviews and questionnaires. Furthermore, a chemical analysis will be performed on crops and soil collected from test sites.


Possible Extensions

As mentioned earlier, the current scope of the research plans to cover a limited arable farmland of size no more than one acre. In future extensions, we hope to extend the coverage to at least ten acres. Further enhancements can include extending stationary charging stations with wireless charging capabilities. This enhancement will help minimise recharging time and increase spraying efficiency.


Design


A viable system design is proposed in this section.


Mapping the Farmland

The discovery drone flies over the boundary of the selected farmland under the supervision of the farmer through mobile application app-A while identifying the relevant boundary geo-locations. This is illustrated in the following figure. These geo-locations are uploaded to the central system using the uplink on the drone.


Capturing Images of the Farmland

To capture aerial photographs of the farmland, the central system directs comm-drones to scan the area surrounded by the boundary geo-locations. Furthermore, the farmer can monitor all related incidents via mobile application app-B. The associated workflow is shown in the following figure.


Generating the 2D Grid World

According to the following figure, the 2D grid world is generated from the set of aerial photographs taken by drones previously.


As shown in the below figure, the safe areas are divided into clusters, with each cluster consisting of squares, each designated as the workspace for single slave drone at a given instance and each square consists of circles. Each circle is a single spray area for a stationary slave and upon completion of all circles, a slave may be assigned a new square.


Spraying Pesticides

The workflow for spraying pesticides is as follows:


  1. The farmer informs the type of crop (e.g. paddy) to central system via mobile application app-B. Central system determines the appropriate dosage from prior knowledge.

  2. Central system commands comm-drones to spray pesticides using slaves, inside their allocated cluster.

  3. Each comm-drone issues commands to its assigned slave drones to spray pesticides in the allocated squares.

  4. Each slave drone sprays pesticides being stationary over the centre of its current circle.

  5. Meanwhile, any comm-drone may proceed to charging station(s) and any slave drone may also advance to charging and/or filling station(s) as discussed in Research Methodology

  6. After finishing the whole spray job, all drones proceed to a hangar.

Design of a Slave Drone

We hope to use two embedded computing platforms to design the slave drone: Raspberry Pi and Arduino. The proposed design is shown in the following figure.


Spray Tank

Slave drones consist of a spray tank. The spray tank is located under the lower part of the drone as illustrated in the above figure. It consists of a single solid cone (or full cone) nozzle attached to an irrigation solenoid valve with a flow meter. The tank also has two level switches (or level sensors) to guard against the liquid volume inside the tank decreasing or increasing beyond the allowed lower and upper limits respectively.


Charging Port

Each drone (including the comm-drones) has a charging port for recharging its battery when the battery has drained beyond 10% of its capacity.


Fill Valve

The filling station contains a one-way fill valve to control the passage of pesticide liquid from the station into the spray tank. Drones are provided the geo-locations of the closest charging and filling station(s).

Financial


Let service provider be the business institution that adopts our research for commercial purposes and transforms it as one (or more) of latter's products. The community-based business model we propose inline with this research is illustrated in the above figure. It is assumed that potential farmers are part of or belong to a locality known as Farmers' Community Centre. These may be run by farmers themselves. It is also assumed that these centres are spread across the country and has the financial capacity to purchase the product hereto from the service provider, which is termed as Package in the figure; containing the swarm of drones (of preferred size), mobile applications, a tablet computer, charging and filling stations. The business model entails a subscription fee per centre, for using the central system for data processing and control of the swarm. To ensure that farmers do not mix excessive amounts of raw pesticides to create the liquid pesticide solution for spraying, the model assumes that a commercial liquid pesticide solution is provided to the community centres via a pesticide provider, in sealed containers (say in cauldrons). These cauldrons are safely stored in the community centres. The centres can rent the package to potential farmers who are interested, along with a couple of cauldrons containing the pesticide solution. These cauldrons can easily be connected to a filling station. Having several packages in the centre will increase the efficiency of spraying among the community and periodic maintenance of them will be carried out by the service provider at the request of the centre. This model allows farmers to agree on easy financing schemes among themselves for using packages and pesticide solutions.

Publications


2018

  1. Proposal: A Swarm of Crop Spraying Drones Solution for Optimising Safe Pesticide Usage in Arable Lands

    A Amarasinghe, V B Wijesuriya, D M Ganepola, K L Jayaratne

    University of Colombo, University of Oxford, University of Colombo, University of Colombo

    PDF


  2. Presentation: A Swarm of Crop Spraying Drones Solution for Optimising Safe Pesticide Usage in Arable Lands

    A Amarasinghe, V B Wijesuriya, D M Ganepola, K L Jayaratne

    University of Colombo, University of Oxford, University of Colombo, University of Colombo

    PDF


  3. A Swarm of Crop Spraying Drones Solution for Optimising Safe Pesticide Usage in Arable Lands: Poster Abstract

    A Amarasinghe, V B Wijesuriya, D M Ganepola, K L Jayaratne

    University of Colombo, University of Oxford, University of Colombo, University of Colombo

    PDF


Investigators


Team

Akarshani Amarasinghe

  • caa[at]ucsc.cmb.ac.lk

Team

Viraj B. Wijesuriya

  • viraj.wijesuriya[at]cs.ox.ac.uk

Team

Dilshan Ganepola

  • dmg[at]adigar.com

Team

Lakshman Jayaratne

  • klj[at]ucsc.cmb.ac.lk