Optical-SAR Decision Fusion with Markov Random Fields for High-Resolution Large-Scale Land Cover Mapping

Decision Fusion at Pixel Level of Multi-Band Data for Land Cover Classification—A Review

“…The fusion of MODIS and Landsat data in [48] proves promising in improving overall accuracy but faces challenges in the classification accuracy of MODIS data and preconditions for linear fusion models. The integration of optical and SAR images [77] demonstrates improved accuracy and efficient computation but is sensitive to speckle and demands higher memory occupation. Lastly, the decisionlevel combination of multi-band information in SAR image fusion [71] presents enhanced classification accuracy but introduces computational complexity and longer processing times.…”

“…Specifically, the methodologies and techniques discussed in this review paper offer significant contributions with wide-ranging applications across various domains. The fusion of decision-making algorithms with hyperspectral-multispectral data, as highlighted by the works [18,24] and with optical-SAR fusion [77], presents opportunities for enhanced precision in applications such as precision agriculture, environmental monitoring, urban planning, disaster management and infrastructure inspection. For instance, in precision agriculture, farmers can utilize hyperspectral imaging combined with classification techniques to monitor crop health, detect early signs of diseases and optimize irrigation strategies.…”

“…CNNs were used in both [75,76] to propose state-of-the-art classification approaches with fuzzy rules, Bayesian margin Dempster-Shafer evidence theory and multi-structure joint decision-making strategies. Finally, in [77], Markov random fields were employed for classification with the final land cover labels formed by a Bayesian decision fusion approach.…”

“…In the following sections, as depicted in Figure 1, certain topics of data fusion are covered by the chosen papers. Limited spatial resolution but rich semantic information [18], the robust fusion of more than two decision sources [24], the full exploitation of shape or texture features [48], reducing computational demands of large images using convolution operators [77] and weight-sharing CNN for reducing weights and connections across different layers [71]. Analytically, Ouerghemmi, Le Bris et al [18] focused on resolving the spatial and spectral resolution tradeoff in remote sensing by proposing a two-step fusion method that combines multispectral and hyperspectral imagery.…”

“…This approach is especially beneficial in regions with challenging weather conditions, such as cloud cover and rainfall, where traditional high-resolution data may be less effective. Additionally, Maggiolo, Solarna et al [77] addressed the crucial topic of climate change monitoring by combining optical and SAR sensors. Decision fusion is employed to combine probabilistic decisions from these sources, considering their unique characteristics and associated uncertainties.…”