Home

Awesome

CDL-Segmentation

This Repository contains code (will be published after paper acceptance) and benchmark results for the paper:

Deep Learning Based Land Cover and Crop Type Classification: A Comparative Study

Paper Link: https://ieeexplore.ieee.org/document/9441483 <br> Accepted at: 2021 International Conference on Digital Futures and Transformative Technologies (ICoDT2)

Segmentation Models

  1. UNet
  2. SegNet
  3. DeepLabv3+

Pre-requistes

Main packages required are:

Datasets

We used Google Earth Engine to generate the Dataset from Crop Land Data Layer provided by United States Department of Agriculture (USDA) and National Agricultural Statistics Service (NASS). Our Study area for analyzing performance of segmentation models is comprised of 9 different counties of state of Nebraska, United States of America. We emphasize on Classifying two categories, thus two different datasets are:- (1) Cultivated and Non-Cultivated Land Cover (2) Out of Cultivated area we classify the crop type i.e. Corn, Soya Bean, Winter Wheat, Alfalfa Hay and Others.

Visual Results

Top Images are Landsat8 image, Ground-Truth and Predicted Results for cultivated area respectively (Black pixels represent non-cultivated and white pixels are cultivated area, whereas Bottom Images are of Landsat8, Ground-Truth and classified Crops (Yellow as Corn, Green as Soyabean, orange as Winter Wheat, Pink as Alfalfa and Black as Other crops).

Cultivated/Non-Cultivated Land Cover

Visual_Results Visual_Results Visual_Results Visual_Results Visual_Results Visual_Results

Out of Cultivated area classified Crop Types

Visual_Results Visual_Results Visual_Results Visual_Results Visual_Results Visual_Results

Quantitative Results

Dataset: LandCover Classification (Cultivated/Non-Cultivated Land area)

UNet

ResultsValues
Accuracy89.5
Dice co-efficient89.2

SegNet

ResultsValues
Accuracy74.69
Dice co-efficient73.43

DeepLabv3

ResultsValues
Accuracy89.13
Dice co-efficient88.6

Dataset Crop type Classification

UNet

ResultsValues
Accuracy67.3
Dice co-efficient57.1

SegNet

ResultsValues
Accuracy49.5
Dice co-efficient37.0

DeepLabv3

ResultsValues
Accuracy69.7
Dice co-efficient62.02

Performance Graphs

UNet

Part 1: Cultivated/Non-Cultivated LandCover Classification

Loss: UNet vs SegNet vs DeepLabv3

Graphs Accuracy: UNet vs SegNet vs DeepLabv3 Graphs Dice Similarity Co-efficient: UNet vs SegNet vs DeepLabv3 Graphs

Part 2: Crop Type Classification

Loss: UNet vs SegNet vs DeepLabv3

Graphs Accuracy: UNet vs SegNet vs DeepLabv3 Graphs Dice Similarity Co-efficient: UNet vs SegNet vs DeepLabv3 Graphs

Maintainer

Asim Hameed Khan (asimhameed.cs@gmail.com)