Small Scale crop Mapping and Yield Prediction using artificial intelligence and machine learning
April, 2024
Executive Summary
The Small scale crop mapping and yield prediction using artificial intelligence and Machine learning (SSCM/AI, ML) is a project by students from Taita Taveta University in partnership with the Kenya Space Agency that aims to bridge the gap between modern technology and the small scale farmer by providing a platform where the farmer can gain insights about their farms without the need for complicated methods and expensive equipment. Existing yield prediction methods are highly statistical relying on samples and often provide low accuracy predictions.
The project aims to also bridge the knowledge gap between the decision makers, stakeholders and the local farmers. Yield information which is of great importance to the government as well as other research institutions barely gets to them. Cooperatives have been put in place to try and solve this problem but it is still a long process that is not always true. African local small scale farms often have a heterogeneous landscape which makes it difficult to map out using traditional method.
This project therefore creates a great opportunity for both the farmer and the decision maker in improving yield and gaining insights about the production and state of farms around the country.
The project also helps in the accomplishment of the UN’s sustainable development goals SDG13 on climate action and adaptability, SDG 2 zero hunger.
Introduction
Food security is a concern of almost every nation around the world. Coming up with measures to reduce loses, improve production and secure food for the people of any nation is in every government’s priority list. Over the past years Kenya has been struck by Agricultural as well as economic drought making it difficult for the survival of people to a point where the government had to rely on aid from donors to stabilize the situation.
With climate change bringing unpredictable weather patterns and changing rainfall patterns. Farmers are disadvantaged while trying to use traditional methods to predict rainfall or crop yield. This is leading to a reduction in the quality and quantity of food for our nation.
Small scale farmers account for about 78% of the countries food production while maize is the staple food for the country according to data from World Bank. The project focuses on maize as our pilot crop and Kitale a place known for its maize production as our pilot area.
The project aims to bring a modern and reliable solution to these problems by utilizing machine learning technology and satellite data for mapping as well as yield prediction and crop health monitoring.
The aim of this proposal is to raise funds for field data collection as well as facilitating production and deployment cost for the project including ground testing and initial user enrolment.
Problem Statement
The existing methods of crop mapping often suffer from limitations such as high costs, data inconsistencies, and limited spatial coverage. These challenges make it difficult to obtain up-to-date and accurate information about crop distribution and health, particularly on a small-scale level. Therefore, there is a need for an innovative approach that leverages AI and ML to overcome these limitations and provide reliable small-scale crop mapping solutions.
Justification
Cost and Data Inconsistencies: Traditional methods of crop mapping and yield prediction often come with high costs, making them inaccessible to small-scale farmers who form a significant portion of food producers in Kenya. Additionally, these methods frequently suffer from data inconsistencies due to limited spatial coverage and reliance on manual data collection processes.
Limited Spatial Coverage: Small-scale farms, which constitute a large percentage of agricultural production in Kenya, often have a heterogeneous landscape, making it challenging to accurately map and predict crop yield using traditional methods. This limitation hampers efforts to provide timely and precise information to farmers and decision-makers.
Need for Real-time Insights: With climate change leading to unpredictable weather patterns and changing rainfall distribution, small-scale farmers face increasing challenges in predicting crop yield and managing their farms effectively. There is a pressing need for real-time insights into crop health, pest infestations, and environmental conditions to enable farmers to make informed decisions and optimize their production processes.
Alignment with Sustainable Development Goals (SDGs): The proposed project aligns with the United Nations' Sustainable Development Goals (SDGs), particularly SDG 13 on climate action and adaptability, and SDG 2 on zero hunger. By leveraging artificial intelligence and machine learning technologies, the project seeks to enhance agricultural sustainability, improve food security, and mitigate the impact of climate change on small-scale farming communities.
Government and Stakeholder Interest: Yield information is crucial for government agencies, research institutions, and other stakeholders to formulate policies, allocate resources, and support agricultural development initiatives effectively. However, existing mechanisms for disseminating such information are often inadequate, leading to a disconnect between decision-makers and local farmers. Addressing this gap is essential for promoting collaboration, knowledge sharing, and data-driven decision-making in the agricultural sector.
Objectives and Goals
Main Objectives
The primary objective of this project is to create a Machine learning model capable of classifying crop type and predicting crop yield from small scale farms and in heterogeneous landscapes.
Specific Objectives
- Parameter Identification: To Identify hyperparameters and factors unique to maize yield prediction.
- Crop health Monitoring: To provide farmers with up to date information about the state of their crops in terms of health and pests.
- Data Collection: To create a repository of agricultural data for researchers and decision makers.
Methodology
Data
Data will be sourced from various sources ensuring ground truthed data and a wide temporal dataset is acquired. Sources include:
- Sentinel-2 Satellite imagery
- Landsat Satellite imagery
- Yield Data for five years from ground selected farms.
- Rainfall and precipitation data
Method
Earth Engine Environment Design
In this phase, an Earth Engine environment will be created to facilitate data processing and analysis. The Google Earth Engine platform will be utilized to access, process, and analyze Sentinel-2 satellite imagery.
Data Acquisition and Preprocessing
Sentinel-2 satellite imagery will be obtained for the study area. The imagery will undergo preprocessing procedures, including radiometric and atmospheric corrections, to ensure data accuracy and consistency.
Customized Maize Algorithm Development
A customized maize algorithm will be developed for classifying maize fields using the SMILE CART classification method. This algorithm will enable the accurate identification of maize crops within the satellite imagery.
Maize Area Mapping and Class Identification
A classification algorithm using random forest neural networks will be developed and applied to the satellite imagery to create maize area maps and classify maize fields. The classification results will be validated through ground truth data and accuracy assessment.
Vegetation Index Calculation
Various vegetation indices, including NDVI, RVI, NPCRI, GCI, NDMI, and EVI, will be calculated using the processed satellite imagery. These indices will provide essential information about crop health and growth.
Rainfall Data Integration
Rainfall data will be incorporated to assess its impact on crop health. Correlations between vegetation indices and rainfall patterns will be analyzed to derive insights into crop performance.
Gaussian Process Learning Framework
A Gaussian Process learning framework using TensorFlow will be established to model the trends between vegetation indices, rainfall, and crop health. The framework will capture complex relationships within the data.
Regression Model Development
A regression model, utilizing TensorFlow Estimator, will be developed to predict crop yield. The model will be trained using vegetation indices, rainfall data, crop area and additional production datasets.
Web Application Development
The predicted yield for the specified user period will be displayed in a web application. The application will provide an interactive interface for users to access and visualize the yield predictions.
Expected Outcomes
- Crop Yield Prediction: A prediction on the expected yield at least for 2 years.
- A web map showing crop health across a field.
- Statistics on performance of the farm in production, trends and indicators such as moisture Quantity and soil fertility.
- A web Application allowing for user interaction.
- A database of agricultural data.