Preliminary Classification of Soil, Plant, and Residue Cover Using Convolutional Neural Networks

D Nahitiya,M N Bisheh,A Patrignani,R P Lollato

doi:10.4148/2378-5977.8081

Abstract

In agricultural fields, knowledge about the proportion of the soil surface covered with crop residue and vegetation canopy is key for improving soil and water conservation practices. In this study we trained a deep convolutional neural network to automate the classification of bare soil, crop stubble, and live vegetation from downward-facing images of agricultural fields. A comprehensive generic dataset, consisting of 3300 training and 645 test images, was collected from agricultural fields across Kansas State University Agricultural Experiment Stations and the Natural Resources Conservation Service Plant Material Center located near Manhattan, KS. Despite the intricate patterns and color textures resulting from different combinations of soil, canopy, and stubble the trained network showed good performance for automating the classification of land cover from images. The network achieved 87% accuracy over the training dataset and 84% accuracy over the test set.

Highlights

Soil cover by crop residue and actively growing vegetation is an important factor controlling soil erosion by wind and water
Knowledge about the proportion of the soil surface covered with crop residue and vegetation canopy is key for improving soil and water conservation practices
In this study we trained a deep convolutional neural network to automate the classification of bare soil, crop stubble, and live vegetation from downward-facing images of agricultural fields

Summary

Introduction

Soil cover by crop residue and actively growing vegetation is an important factor controlling soil erosion by wind and water. Several practical methods have been developed to quantify the soil cover in field conditions based on simple principles. Line transects are selected at random and are often repeated several times to obtain an accurate average of soil cover values per field. Another common method often used to quantify soil residue and canopy cover is the use of reference photographs. The classification of all three components—green canopy cover, crop stubble, and bare soil—still remains challenging because of the wide range of scenarios caused by the combination of soil types, crops, and soil moisture conditions in agricultural fields. The goal of this study was to quantify the fraction of green canopy cover, crop residue, and bare soil by using a deep neural network and a dataset of pixel-wise labeled images

Objectives

Results

Conclusion