Interpretable Deep Learning Framework for Land Use & Land Cover Classification in Remote Sensing using SHAP

Abstract

Land Use and Land Cover (LULC) classification
plays a pivotal role in a range of applications,
including urban planning, environmental
monitoring, agricultural management, and climate
change analysis. With the rapid advancement in
satellite and aerial remote sensing technologies,
vast volumes of high-resolution multispectral and
hyperspectral imagery are now accessible. While
deep learning has emerged as a powerful tool for
automating LULC classification with high
accuracy, its black-box nature poses significant
challenges to transparency, trustworthiness, and
adoption in critical domains. This research
proposes an interpretable deep learning
framework that not only delivers accurate LULC
classification but also ensures model
explainability through the use of SHAP (SHapley
Additive exPlanations), a game-theoretic
approach to interpreting machine learning
models.
The proposed framework integrates a
convolutional neural network (CNN)-based
architecture trained on satellite imagery to
classify different land cover types such as
vegetation, water bodies, built-up areas, barren
land, and agricultural fields. The CNN is trained
using a labeled remote sensing dataset, such as
Sentinel-2 or Landsat 8 imagery, with
preprocessing steps including radiometric
calibration, normalization, and augmentation to
handle class imbalances and improve
generalization.