Forest Attributes Segmentation

Update on 2022/10/02. Provide reclassification step to enhance the resolution of the result image (only for forest species)

Introduction

This is the Pytorch (1.9.1) implementation of Deep Learning model in "FOREST-RELATED SDG ISSUES MONITORING FOR DATA-SCARE REGIONS EMPLOYING MACHINE LEARNING AND REMOTE SENSING - A CASE STUDY FOR ENA CITY, JAPAN".

• Support different 2D/3D UNET based architecture with Atrous Convolution Blocks for forest attribute (species, age) segmentation
• Support training/validation dataset from Sentinel 1/2 in GIFU prefecture - Japan.

Method

Deep learning architecture

Reclassification

The Reclassification method will be updated soon

Study area and data collection

Training data was collected from 国土数値情報ダウンロードサービス

Results

Installation

The source code is test with Anaconda and Python 3.9.7.

1. Clone the repo:

    git clone https://github.com/anhp95/forest_attr_segment.git
    cd forest_attr_segment

2. Create a conda environment from as follows:

    conda env create -f environment.yml

Training

Follow these steps to train the model with our dataset

1. Download the dataset via Google Drive

2. Configure the dataset path in mypath.py

3. Activate your Anaconda environment

4. Input arguments: (see the full set of input arguments via python train.py --help)

   usage: train_nn.py [-h] [--forest_attr {spec,age}]
                  [--backbone {2d_p2,2d_p1p2,2d_p1p2p3,3d_org,3d_adj,3d_adj_dec_acb,3d_adj_emd_acb,3d_org_emd_acb}]
                  [--num_epochs NUM_EPOCHS] [--batch_size BATCH_SIZE] [--lr LR] [--load_model LOAD_MODEL]
                  [--logs_file LOGS_FILE] [--pin_memory] [--no_workers NO_WORKERS]

Inference

1. Input arguments: (see the full set of input arguments via python infer_nn.py --help)

   usage: infer_nn.py [-h] [--forest_attr {spec,age}]
                  [--backbone {2d_p2,2d_p1p2,2d_p1p2p3,3d_org,3d_adj,3d_adj_dec_acb,3d_adj_emd_acb,3d_org_emd_acb}]
                  [--batch_size BATCH_SIZE] [--region {ena,nakat,mizunami,toki,tajimi,tono}] [--recls {0,1}]
                  [--n_clusters N_CLUSTERS]

Note. High-resolution inference (Only available for forest species)