README.md

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
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:

• Keras
• Tensorflow 1.15.0
• Numpy
• Skimage
• Matplotlib

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

Out of Cultivated area classified Crop Types

Quantitative Results

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

UNet

Results	Values
Accuracy	89.5
Dice co-efficient	89.2

SegNet

Results	Values
Accuracy	74.69
Dice co-efficient	73.43

DeepLabv3

Results	Values
Accuracy	89.13
Dice co-efficient	88.6

Dataset Crop type Classification

UNet

Results	Values
Accuracy	67.3
Dice co-efficient	57.1

SegNet

Results	Values
Accuracy	49.5
Dice co-efficient	37.0

DeepLabv3

Results	Values
Accuracy	69.7
Dice co-efficient	62.02

Performance Graphs

UNet

Part 1: Cultivated/Non-Cultivated LandCover Classification

Loss: UNet vs SegNet vs DeepLabv3

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

Part 2: Crop Type Classification

Loss: UNet vs SegNet vs DeepLabv3

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

Maintainer

Asim Hameed Khan ([email protected])