Abstract:
This work investigates the use of advanced image classification techniques for
improving the accuracy and efficiency in determining agricultural areas from
satellite images. The United Nations Office on Drugs and Crime (UNODC) need
to accurately delineate the potential area under opium cultivation as part of their
opium monitoring programme in Afghanistan. They currently use unsupervised
image classification, but this is unable to separate some areas of agriculture from
natural vegetation and requires time-consuming manual editing. This is a
significant task as each image must be classified and interpreted separately. The
aim of this research is to derive information about annual changes in land-use
related to opium cultivation using convolutional neural networks with Earth
observation data.
Supervised machine learning techniques were investigated for agricultural land
classification using training data from existing manual interpretations. Although
pixel-based machine learning techniques achieved high overall classification
accuracy (89%) they had difficulty separating between agriculture and natural
vegetation at some locations.
Convolutional Neural Networks (CNNs) have achieved ground-breaking
performance in computer vision applications. They use localised image features
and offer transfer learning to overcome the limitations of pixel-based methods.
There are challenges related to training CNNs for land cover classification
because of underlying radiometric and temporal variations in satellite image
datasets. Optimisation of CNNs with a targeted sampling strategy focused on
areas of known confusion (agricultural boundaries and natural vegetation). The
results showed an improved overall classification accuracy of +6%. Localised
differences in agricultural mapping were identified using a new tool called
‘localised intersection over union’. This provides greater insight than commonly
used assessment techniques (overall accuracy and kappa statistic), that are not
suitable for comparing smaller differences in mapping accuracy.
A generalised fully convolutional model (FCN) was developed and evaluated
using six years of data and transfer learning. Image datasets were standardised
across image dates and different sensors (DMC, Landsat, and Sentinel-2),
achieving high classification accuracy (up to 95%) with no additional training.
Further fine-tuning with minimal training data and a targeted training strategy
further increased model performance between years (up to +5%).
The annual changes in agricultural area from 2010 to 2019 were mapped using
the generalised FCN model in Helmand Province, Afghanistan. This provided
new insight into the expansion of agriculture into marginal areas in response to
counter-narcotic and alternative livelihoods policy. New areas of cultivation were
found to contribute to the expansion of opium cultivation in Helmand Province.
The approach demonstrates the use of FCNs for fully automated land cover
classification. They are fast and efficient, can be used to classify satellite imagery
from different sensors and can be continually refined using transfer learning.
The proposed method overcomes the manual effort associated with mapping
agricultural areas within the opium survey while improving accuracy. These
findings have wider implications for improving land cover classification using
legacy data on scalable cloud-based platforms.