Case Study: Multi-Label Classification of Satellite Images with Fast.AI

Satellite image classification is the most significant technique used in remote sensing (GIS) for the computerized study and pattern recognition of satellite GIS, which is based on diversity structures of the image that involve rigorous validation of the training samples depending on the used ML/AI classification algorithm.

Satellite imagery is important for many applications including disaster response, law enforcement, and environmental monitoring. These GIS applications require the automated AI-powered identification of objects and facilities in the imagery.

In this post, we focus on the satellite image segmentation using Fast.Ai.

Satellite imagery is being used together with AI and deep learning in many areas to produce stunning insights and discoveries. Today we look at applying this approach to recognising buildings, woodlands & water areas from satellite images.

Conventionally, we use 4 classes for identifying objects in GIS images:

• Building
• Woodland
• Water
• Background (i.e. everything else).

For this multi-label image classification problem, we will use the Planet dataset, where it’s a collection of satellite images with multiple labels describing the scene. 

The entire workflow consists of the following steps:

• Grab our input data
• Train a model with fastai
• QC with fastai metrics.

Let’s set the working directory YOURPATH

import os
os.chdir(‘YOURPATH’)
os. getcwd()

and import the following libraries

from fastai.vision.all import *

import pandas as pd

import torch
from torch import nn

from fastcore.meta import use_kwargs_dict

from fastai.callback.fp16 import to_fp16
from fastai.callback.progress import ProgressCallback
from fastai.callback.schedule import lr_find, fit_one_cycle

from fastai.data.block import MultiCategoryBlock, DataBlock
from fastai.data.external import untar_data, URLs
from fastai.data.transforms import RandomSplitter, ColReader

from fastai.metrics import accuracy_multi, BaseLoss

from fastai.vision.augment import aug_transforms
from fastai.vision.data import ImageBlock
from fastai.vision.learner import cnn_learner

from torchvision.models import resnet34

Let’s import the input dataset

planet_source = untar_data(URLs.PLANET_SAMPLE)
df = pd.read_csv(planet_source/’labels.csv’)

Let’s check the content

df.head()

Let’s edit the data columns

df = df[df[‘tags’] != ‘blow_down clear primary road’]

batch_tfms = aug_transforms(flip_vert=True, max_lighting=0.1, max_zoom=1.05, max_warp=0.)

planet = DataBlock(blocks=(ImageBlock, MultiCategoryBlock),
get_x=ColReader(0, pref=f'{planet_source}/train/’, suff=’.jpg’),
splitter=RandomSplitter(),
get_y=ColReader(1, label_delim=’ ‘),
batch_tfms = batch_tfms)

dls = planet.dataloaders(df)

and plot the first 9 selected images

dls.show_batch(max_n=9, figsize=(12,9))

We can also invoke the lambda function

blocks = (ImageBlock, MultiCategoryBlock)

get_x = lambda x:planet_source/’train’/f'{x[0]}.jpg’

val = df.values[0]; val

array(['train_21983', 'partly_cloudy primary'], dtype=object)

get_x(df.values[0])

get_y = lambda x:x[1].split(‘ ‘)

planet = DataBlock(blocks=blocks,
get_x=get_x,
splitter=RandomSplitter(),
get_y=get_y,
batch_tfms=batch_tfms)

dls = planet.dataloaders(df)
dls.show_batch(max_n=9, figsize=(12,9))

Let’s invoke planet.dataloaders

def _planet_items(x): return (
f'{planet_source}/train/’+x.image_name+’.jpg’, x.tags.str.split())

planet = DataBlock.from_columns(blocks=(ImageBlock, MultiCategoryBlock),
get_items = _planet_items,
splitter=RandomSplitter(),
batch_tfms=batch_tfms)

dls = planet.dataloaders(df)
dls.show_batch(max_n=9, figsize=(12,9))

Let’s train the model

from torchvision.models import resnet34

from fastai.metrics import accuracy_multi

learn = cnn_learner(dls, resnet34, pretrained=True, metrics=[accuracy_multi])

class BCEWithLogitsLossFlat(BaseLoss):
“Same as nn.BCEWithLogitsLoss, but flattens input and target.”
@use_kwargs_dict(keep=True, weight=None, reduction=’mean’, pos_weight=None)
def init(self, *args, axis=-1, floatify=True, thresh=0.5, **kwargs):
if kwargs.get(‘pos_weight’, None) is not None and kwargs.get(‘flatten’, None) is True:
raise ValueError(“flatten must be False when using pos_weight to avoid a RuntimeError due to shape mismatch”)
if kwargs.get(‘pos_weight’, None) is not None: kwargs[‘flatten’] = False
super().init(nn.BCEWithLogitsLoss, *args, axis=axis, floatify=floatify, is_2d=False, **kwargs)
self.thresh = thresh

def decodes(self, x):    return x>self.thresh
def activation(self, x): return torch.sigmoid(x)

learn.loss_func = BCEWithLogitsLossFlat()

learn.lr_find()

SuggestedLRs(valley=0.0020892962347716093)

lr = 1e-2
learn = learn.to_fp16()

learn.fit_one_cycle(5, slice(lr))

learn.save(‘stage-1’)

Path('models/stage-1.pth')

learn.unfreeze()
learn.lr_find()

SuggestedLRs(valley=7.585775892948732e-05)

learn.fit_one_cycle(5, slice(1e-5, lr/5))

learn.show_results(figsize=(15,15))

Summary

• Instead of cats & dogs, the Planet Competition Dataset consists of satellite images from the Amazonian region.
• The task here consists of classifying which types of land covers are present on each image. We can have multiple landcovers types present on one image.
• Here the task is a multi-label classification problem, where each image can belong to multiple classes.
• Using pre-trained models is a good practice in general.

Explore More

fast.ai’s superresolution model on satellite imagery.

Multi-label classification using fastai

Multi-Label Keras CNN Image Classification of MNIST Fashion Clothing

ML/AI Wildfire Prediction using Remote Sensing Data

ML/AI Wildfire Prediction

Embed Socials