IEEE Geoscience and Remote Sensing Magazine - June 2023 - 71

Beidou Navigation Satellite System, StarFL provides more
security guarantees for each participant in urban computing,
which includes autonomous driving and resource exploration.
They specified that the independence of the satellite
cluster makes it easy for StarFL to support vertical FL.
Jiang et al. [101] also pointed out that interactive learning
between vehicles and their system environments through
vertical FL can help assist with other city sensing applications,
such as city traffic lights, cameras, and roadside units.
FEDERATED TRANSFER LEARNING
This category suits cases where neither the sample space nor
the feature space overlap, as shown in Figure 8(c). In view of
the problems caused by the small amount of data and sparse
labeled samples, federated transfer learning is introduced to
learn knowledge from the source database, and to transfer
it to the target database while maintaining the privacy and
security of the individual data. In real applications, Chen
et al. [102] constructed a FedHealth model that gathers the
data owned by different organizations via FL and offers personalized
services for health care through transfer learning.
Limited by the available data and annotations, federated
transfer learning remains challenging to popularize in
practical applications today. However, it is still the most effective
way to protect data security and user privacy while
breaking down data barriers for large-scale ML.
FUTURE PERSPECTIVES
With the exponential growth of AI applications, data security
and user privacy are attracting increasing attention in geoscience
and RS. For this purpose, FL can aggregate a desired
global model from local models without exposing data, and
has been applied in various topics such as real-time image
classification, forest fire detection, and autonomous vehicles,
among others. Based on the needs of currently available
local data and individual users, existing systems are
best served by focusing more on horizontal FL. There are
other possible research directions in the future for vertical
FL and federated transfer learning. Here we list some examples
of potential applications that are helpful for a comprehensive
understanding of geoscience and RS through FL.
GENERATING GLOBAL-SCALE GEOGRAPHIC
SYSTEMATIC MODELS
The geographic parameters of different countries are similar,
but geospatial data often cannot be shared due to national
security restrictions and data confidentiality. A horizontal
FL system could train local models separately, and
then integrate the global-scale geographic parameters on
the server according to the contribution of different owners,
which could effectively avoid data leaks.
INTERDISCIPLINARY URBAN COMPUTING
As is known, much spatial information about a specific city
can be recorded conveniently by RS images. Still, other information,
such as the locations of people and vehicles, and
JUNE 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
the elevation information of land covers, are usually kept
privately by different industries. Therefore, designing appropriate
vertical FL systems will be helpful for increasing urban
understanding, such as estimating population distributions
and traffic conditions, and computing 3D maps of cities.
OBJECT DETECTION AND RECOGNITION,
CROSS-SPATIAL DOMAIN AND SENSOR DOMAIN
The RS data owned by different industries is usually captured
by different sensors, and geospatial overlap is rare.
Considering that the objects of interest are usually confidential,
local data cannot be shared. In this case, the federated
transfer learning system can detect objects of interest effectively
by integrating local models for cross-domain tasks.
UNCERTAINTY
In the big data era, AI techniques, especially ML algorithms,
have been applied widely in geoscience and RS missions. Unfortunately,
regardless of their promising results, heterogeneities
within the enormous volume of EO data, including noise
and unaccounted-for variation, and the stochastic nature of
the model's parameters, can lead to uncertainty in the algorithms'
predictions, which may not only severely threaten the
performance of the AI algorithms with uncertain test samples
but also reduce the reliability of predictions in high-risk RS
applications [103]. Therefore, identifying the occurrence of
uncertainty, modeling its propagation and accumulation,
and performing uncertainty quantification in the algorithms
are all critical to controlling the quality of the outcomes.
PRELIMINARIES
AI techniques for geoscience and RS data analysis aim to
map the relationship between properties on the earth's surface
and EO data. In practice, the algorithms in these techniques
can be defined as a mathematical mapping that transforms
the data into information representations. For example,
neural networks have become the most popular mapping
function that transforms a measurable input set X into a
measurable set Y of predictions, as follows:
f :XY " i
(8)
where f denotes the mapping function, and i represents
the parameters of the neural network.
Typically, as shown in Figure 9, developing an AI algorithm
involves data collection, model construction, model
training, and model deployment. In the context of supervised
learning, a training dataset D is constructed in the
data collection step, containing N pairs of input data sample
x and labeled target y, as follows:
DX Y (, ){ ,} .xy
N
==
ii i =1
(9)
Then, the model architecture is designed according to the requirement
of EO missions, and the mapping function as well
as its parameters i are initialized (i.e., fi
is determined). Next,
71
IEEE Geoscience and Remote Sensing Magazine - June 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2023
