Agricultural Land cover mapping by active learning from multispectral spot-7 satellite image
Résumé
Agricultural practices are major drivers of water flows in cultivated landscapes. Land Cover
mapping have a strong impact onto run off and soil erosion at the landscape and watershed
scales. In practice, there is a need for agricultural land cover mapping at both scales that
could be integrated into hydrological models to better understand the considered hydrological
behavior and thus improve water resources management
Field methods are inadequate for characterizing the spatial variability of crops at such scales.
Remote sensing appears therefore as a promising alternative. The classification of agricultural
land cover using a single multispectral satellite images has proven to be challenging due to the
high parcel intra variability over a wide study area.
Our objective is land cover mapping of agricultural fields, at a large scale, using active learning
techniques. The methods are applied to a spot-7 multispectral satellite image over a 35 km2 area
of Lebna catchment in the North Eastern Tunisia. The image high spatial resolution allows a
large scale land cover mapping on a wide extent.
The proposed method is based on a supervised classification, using field observations as
learning samples. The first difficulty consists is having a very small number of learning
samples. The classifier model is then constructed locally, making it suboptimal for the entire
area. Indeed, the learning samples are assumed to be representative of the whole data set which
is rarely confirmed in practice on a wide area.
To solve this issue, the proposed method is based on active learning techniques that build an
efficient learning set by improving iteratively the model performance by adding the most
informative samples. Samples are then labelled by the user and allow to construct a new
classifier model that should be more efficient. The selection of new samples needs a strategy to
rank pixels. Two criteria are used and often coupled: uncertainty and diversity. The samples
should be the most informative (I.e uncertain for the current classifier model) and diverse (I.e
non redundant).
The used uncertainty measure is based on a random forest classifier probability. This measure
can be applied to any other classifier that provides a probability of belonging to different
classes. For sample diversity, two metrics are used; a similarity-based and clustering-based
techniques.
Besides, this paper focuses on an operational strategy that allows mapping agriculture cover on
a large wide area. Two schemes are possible; either a pixel-based classification or a parcel-
based classification. The latter assumes having a digitized parcels; using GIS techniques, of the
study area but has the advantage of highly reducing computing times and leading to a land cover
map at the parcel scale that can be directly used in hydrological models.
First results show the effectiveness of active learning techniques to map a wide area while
maintaining a small number of training samples.