Degree Type

Creative Component

Semester of Graduation

Spring 2020

Department

Electrical and Computer Engineering

First Major Professor

Brian K Gelder

Second Major Professor

Chinmay Hegde

Degree(s)

Master of Science (MS)

Major(s)

Computer Engineering

Abstract

The Agricultural Conservation Planning Framework (ACPF) is a framework for watershed analysis that is supported by a unique land management database. Implementing the ACPF Framework comprises several steps. One of the most important steps in this framework is manually editing the United States Department of Agriculture (USDA) Farm Service Agency (FSA) Common Land Unit (CLU) boundaries to match cropping patterns per USDA National Agricultural Statistics Service (NASS) Cropland Data Layer (CDL) and National Agricultural Imagery Program (NAIP) aerial imagery. This step uses lot of man-hours and is highly susceptible to human errors. The use of latest deep-learning techniques will help alleviate some of these issues. In this project, various Machine learning techniques have been implemented to assist in this particular step of ACPF and the correctness of those techniques have been analyzed in detail. The ACPF Database also facilitates data for the Daily Erosion Project (DEP), a daily estimator of sheet and rill erosion across the western US Corn Belt [1]. In this report, we will detail field boundary digitization for the ACPF which can then be applied to DEP as well.

We calculate the accuracy of the machine learning models by comparing their output produced with the manually edited boundaries. The accuracy is quantized using Kappa Coefficient. The machine learning techniques used for this process include, Maximum likelihood Classification, Random Trees Classification, and Support Vector Machine Classification. Our last experiment is to create a Convolutional Neural Network (CNN) model to classify the crop type present in each field area. We use a small set of image chips of fields for training the CNN, and then apply it to the remaining images, and display the correctness of the Neural network model. Sample results from counties in Kansas and Nebraska are presented in this report.

Copyright Owner

Suresh Kumar, Rishikumar

File Format

Word

Share

COinS