A Recommendation Engine for Subsistence Farmers

Daniel P. Stormont and Samantha L. Snabes

The Pineapple Project, Root2Market, Mountain View, U.S.A.

Keywords:

Expert Systems, Planning and Scheduling.

Abstract:

Many of the world’s agricultural workers, especially in the equatorial regions, are subsistence farmers. This

paper describes the design of a Recommendation Engine that is the core of a crowd-sourced tool to assist sub-

sistence farmers in better utilizing their land. This tool is being developed by a worldwide team of developers

through hackathons and volunteer work. The paper provides an introduction to the project, an overview of the

system design, current and future designs planned for the Recommendation Engine, and the next steps for the

project.

1 INTRODUCTION

According to the World Bank, 75% of the world’s

poor live in rural areas where the majority of them

are subsistence farmers (WorldBank, 2007). Subsis-

tence farming is deﬁned as “a system of farming that

provides all or almost all the goods required by the

farm family usually without any signiﬁcant surplus

for sale” (Dictionary, 2012). As also noted in (World-

Bank, 2007), one of the “pathways out of poverty” in

the World Bank agriculture-for-development agenda

is to “improve livelihoods in subsistence agriculture

and low-skill rural occupations.” An essential ele-

ment of improving livelihoods is providing farmers

with better information: about crops, growing condi-

tions, pests, growing tips, produce handling sugges-

tions, and market information. Providing that infor-

mation is the objective of the Pineapple Project.

This project was incubated in a number of global

hackathons focused on using technology to assist

those in need, especially in developing nations. A

multi-national team was formed and key members of

the team have continued on with the project to provide

continuity between hackathons. Presently, a number

of partnerships have been formed with other ongo-

ing research projects to be able to make continuous

progress, although there is still an intent to utilize the

energy and new ideas resulting from participating in

hackathons in the future. The project was launched at

the National Aeronautics and Space Administration

(NASA)-sponsored International Space Apps Chal-

lenge (ISAC) in April 2012 (ISAC, 2012). Since its

launch, the project has been advanced in two Random

Hacks of Kindness (RHoK) global events (RHoK,

2012) and at the United States Agency for Interna-

tional Development (USAID) Hacking4Hunger event

(USAID, 2012). The Pineapple Project was an award

winner in all four hackathons.

The remainder of this paper will present the de-

sign of the system, as it is currently envisioned, and

then focus on the design of the core of the system: the

Recommendation Engine. This engine combines ge-

ographic location, growing conditions data, and crop

data to assist farmers with planting decisions. The pa-

per will conclude with a description of the next steps

for the project.

2 SYSTEM DESIGN

The crop recommendation system is being built in

phases by team members around the world, so the

design is distributed, extensible, and highly modular.

Figure 1 shows a block diagram of the system. There

are a number of services and data sources that inter-

act with the Recommendation Engine – the common

component in the system.

2.1 Crop Database

The crop database is presently limited to a small sub-

set of tropical fruits common to the equatorial region

of the world. The ideal growing condition parameters

for these fruits are stored in the database, with the type

of fruit as the key value. This data is sufﬁcient for the

planned pilot testing of the system, but it will need to

578

P. Stormont D. and L. Snabes S..

A Recommendation Engine for Subsistence Farmers.

DOI: 10.5220/0004331805780581

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 578-581

ISBN: 978-989-8565-39-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: The crop recommendation system block diagram.

be expanded for future versions of the system, either

manually or by incorporating additional data sources.

2.2 Growing Conditions Database

The growing conditions database contains historical

climatological data, indexed to a grid associated with

latitudes and longitudes on Earth. For the pilot test-

ing, the system is making use of the Prediction Of

Worldwide Energy Resource (POWER) data set pro-

vided by NASA (Hoell et al., 2007). This data set

provides daily integrated surface radiation, daily av-

eraged dew point temperature, daily maximum and

minimum temperatures, and daily precipitation on

a one-degree latitude/longitude grid. The database

also includes soil conditions data from the Food and

Agriculture Organization of the United Nations (FAO,

2012). All of the data is stored in a NoSQL database

for fast retrieval and because of the sparse nature of

the data sets (Strauch, 2012). (Much of the data is

omitted since the only data of interest to the project is

the data that covers land masses.)

In future iterations of the system, the growing

conditions data will be expanded with other sources–

including crowd-sourced data–and additional param-

eters will be added, such as polyculture, diseases, and

pests.

2.3 Price Database

This database is not part of the pilot project, which is

focused on growing conditions only. There are a num-

ber of projects that are looking at the issue of crowd-

sourced pricing information and producer-driven mar-

kets at the national and regional levels, such as the

OpenMarkets Foundation (OpenMarketsFoundation,

2011), but none of them are to a point that their sys-

tems can be integrated with the crop recommendation

system. Future iterations of the system will consider

methods for providing price and demand data as a

source to the Recommendation Engine.

2.4 Geolocation

Geolocation is critical for ﬁltering the data provided

to the Recommendation Engine. The geolocation in

the current system will take advantage of the Global

Positioning Satellite (GPS) data provided by a device

that is so equipped or by doing a lookup on a loca-

tion provided through the interface. If the user knows

the latitude and longitude of the location they want a

recommendation for, that can be entered into the sys-

tem. Otherwise a lookup of a provided location, such

as a municipality, will be used to query a geocoding

service. At present, the system is using the Google

Geocoding API (Google, 2012), but this will need to

be supplemented with other services for areas that are

not well represented in the Google data.

2.5 User Interface

Three user interfaces have been prototyped during the

hackathons: a web interface (actually, several web in-

terfaces), a smart phone app, and a Short Message

Service (SMS) interface for standard cellphones.

2.5.1 Web Interface

The web interface is the one used most by the col-

laborating team members in developing the crop rec-

ommendation system. It provides facilities for tex-

tual, graphical, and pointing device interaction and

displays; thus it is more visually appealing and it is

easier to assess the quality of the data and the results.

However, for the pilot program, it is not likely that a

subsistence farmer in a rural area of a developing na-

tion will have access to a computer and a high-speed

data connection to utilize this interface.

2.5.2 Smart Phone

Some of the international participants in the ISAC

prototyped an iPhone app for accessing basic crop

data. It is functional and could be modiﬁed to ac-

cess a mobile version of the website, but smart phone

adoption has not reached the level of penetration in

developing nations that would allow the project to fo-

cus on this as a primary interface for the system...yet.

Smart phone prices are coming down and adoption is

accelerating, even in less afﬂuent areas of the world,

so this will likely be a secondary option for accessing

the system.

ARecommendationEngineforSubsistenceFarmers

579

2.5.3 Cellphone Text

Simple cellphones (or “ﬂip phones”) are ubiquitous

throughout the world and even the poorest nations

typically have at least one cellphone provider for

voice and SMS services. Thus, the project has fo-

cused on the SMS interface as the primary interface

for the pilot test of the system. The SMS interface

is simplistic, but the system can still gather enough

information in one or two text messages to make a

recommendation. All of the processing is done on the

server side by the Recommendation Engine through

its interfaces to other services and data sources.

3 RECOMMENDATION ENGINE

DESIGN

Central to everything that the crop recommendation

system does is the Recommendation Engine. This en-

gine is responsible for taking the inputs from the user

interface and the geolocation/geocoding service and

ﬁltering the data in the database to come up with a rec-

ommendation or recommendations to the farmer that

will be appropriate to the location of the planting and

the time of the year.

3.1 Interface Design

The distributed nature of the teams working on this

project and the desire to keep the system modular and

extensible to allow for replacing elements of the sys-

tem when better data sources or sources or services

that are more appropriate to a given location are iden-

tiﬁed drives the need for well deﬁned interfaces be-

tween all of the system components. The interfaces

are deﬁned with APIs for services (at present these

are usually APIs deﬁned by external organizations,

such as the Google Geocoding API (Google, 2012))

and with a JavaScript Object Notation (JSON) format

(JSON, 2012) for data interfaces. An example of a

JSON data interface deﬁnition for the elevation asso-

ciated with a geolocation is as follows:

{ "Elevation" :

{

"MinEl" : null,

"AvgEl" : 456,

"MaxEl" : null

}

JSON has the advantage of being readily under-

stood by humans reviewing the data and also being

fairly easy to parse and to convert from other formats.

The recommendation system uses Python scripts on

the server extensively for data conversion and parsing.

Converting JSON to Python is a simple direct con-

version from JavaScript objects and arrays to Python

dictionaries and lists. Conversion from extensible

Modeling Language (XML) to JSON is also straight-

forward with Python libraries.

All of the interfaces used by the Recommendation

Engine at present are deﬁned in a JavaScript Interface

Format speciﬁcation that has been made available to

the developers working on the project.

3.2 Current RE Design

The current design of the Recommendation Engine

makes use of logic programming to determine the

crop or crops that meet the growing criteria. This

is a method expert systems have used since the early

years of expert systems programmed in Prolog. In this

case, the expert system is programmed in server-side

Python, but the logic is the same.

To begin the process, the location of the area to be

planted is determined either through the geolocation

service on the device, a geocoding service, or by di-

rect entry of a latitude and longitude. This location is

used as a hash into the growing conditions database.

The retrieved growing conditions contain the temper-

ature, sunlight, and precipitation parameters for the

full year. The current date or a user-entered planting

date is used to limit the growing conditions for plant-

ing to the data for a single month.

With the growing conditions data retrieved, the ex-

pert system in the Recommendation Engine selects

crops from the crop database that satisfy the grow-

ing conditions. For example, the Recommendation

Engine will determine if either of the temperature ex-

tremes for the month (minimum or maximum tem-

perature) would violate the temperature range for a

crop in the database. If so, that crop is removed from

consideration. The same is done for the sunlight and

precipitation data. The crops that meet the selection

criteria form the initial recommendation. If the initial

recommendation list is empty, the user is informed

that no crops meet the criteria for that area and time

of year and the process terminates. Otherwise, the

growth rate of the crops in the recommendation list is

used to retrieve additional months of data from the

growing conditions data to make sure none of the

constraints on the selected crops are violated at some

point during the growth period. The crop recommen-

dations that survive this second check are presented

to the user.

The advantage of the logical programming ap-

proach is that the decision process can be logged and

used for troubleshooting program and/or data errors.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

580

It can also serve an educational purpose by providing

an answer to the question of why a particular crop in

the crop database was not selected by the Recommen-

dation Engine. However, while the logic program-

ming approach is feasible for the small size of the crop

database and the limited parameters being considered

for the pilot testing, it will not scale well to some of

the desired future capabilities for the system. That

will require a different approach.

3.3 Proposed Future Design

To truly be a useful tool for subsistence farmers, the

crop recommendation system needs to have more than

just a small subset of tropical fruits in it and it needs

to take many more factors into consideration. In the

crop database, any crop that could be grown in a re-

gion should be considered by the Recommendation

Engine. The price database also needs to be imple-

mented and incorporated into the recommendation.

The farmer should be able to get an estimate of the

market value of the crop when it is harvested, as well

as the best place to take the crop to market. The

growing conditions database needs to be expanded to

include parameters like polyculture (simultaneously

planting multiple crops) or crop rotation when mak-

ing a recommendation. And tips on dealing with pests

and diseases, as well as methods for harvesting, stor-

ing, and transporting the produce to market should be

made available.

These enhancements to the functioning of the

Recommendation Engine will require a redesign,

since logical programming doesn’t scale well when

multiple factors have to be examined. The planned

approach to the redesign will be to replace the expert

system in the Recommendation Engine with a neural

network that takes all of the growing conditions, pest,

crop, and price data as input and outputs crop recom-

mendations and tips to the grower. The disadvantage

is that the decision process is no longer observable;

but the neural network should be faster and more ﬂex-

ible.

4 NEXT STEPS

The immediate objective for the project team is to

get an SMS-based version of the system into the

ﬁeld for testing with subsistence farmers and Non-

Governmental Organization (NGO) personnel on the

ground in one of the equatorial regions.

In the longer term, it will be necessary to continue

expanding the capabilities and scope of the crop rec-

ommendation system, in order to increase its value

to the farmers that use it and their families and com-

munities. The ultimate goal is to reduce hunger and

poverty in the developing world. This will require ad-

vancing the design and implementation of the system

and it will require close monitoring of trends, like in-

creasing use of smart phones, that could make the sys-

tem more useful to its users.

5 CONCLUSIONS

This project is an ambitious undertaking with the po-

tential to really make a difference in people’s lives.

It has inspired enthusiasm from sponsors, hackathon

judges, and especially from the participants in the

project. Within the year we hope to take a pilot ver-

sion of the system into the ﬁeld and begin to translate

this passion for helping others into tangible results.

REFERENCES

Dictionary (2012). subsistence farming. Merriam-

Webster.com, http://www.merriam-webster.com.

FAO (2012). pH - TOPSOIL. Food and Agriculture Or-

ganization of the United Nations, http://www.fao.org/

geonetwork.

Google (2012). The Google Geocoding API. Google,

https://developers.google.com.

Hoell, J. M., Stackhouse, P. W., Westberg, D. J., Chandler,

W. S., Whitlock, C. H., and Zhang, T. (2007). Clima-

tological to near real time global meteorological data

for agriculture, range, and forestry applications. In

Fall American Geophysical Union Meeting, San Fran-

cisco, CA.

ISAC (2012). The Pineapple Project. International Space

Apps Challenge, http://2012.spaceappschallenge.org.

JSON (2012). Introducing JSON. JSON.org, http://json.org.

OpenMarketsFoundation (2011). Evaluating the impact of

an SMS-based Market Information System on small-

holder farmers in the Volta Region. GNAFF.

RHoK (2012). The Pineapple Project. Random Hacks of

Kindness, http://www.rhok.org.

Strauch, C. (2012). Nosql databases. Technical report,

Hochschule der Medien, Stuttgart.

USAID (2012). The PineApple Project. United

States Agency for International Development,

http://idea.usaid.gov.

WorldBank (2007). World Development Report 2008: Agri-

culture for Development. The World Bank, Washing-

ton, DC.

ARecommendationEngineforSubsistenceFarmers

581