Predicting Ground Cover with Deep Learning Models—An Application of Spatio-Temporal Prediction Methods to Satellite-Derived Ground Cover Maps in the Great Barrier Reef Catchments

• Published in: Remote sensing (Basel, Switzerland) Vol. 16; no. 17; p. 3193
• Main Authors: Mao, Yongjing, Turner, Ryan D. R., McMahon, Joseph M., Correa, Diego F., Chamberlain, Debbie A., Warne, Michael St. J.
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.09.2024
• Subjects: Accelerated erosion; Catchment areas; Catchments; Datasets; Deep learning; Error analysis; Grazing; Great Barrier Reef Catchments; Ground cover; Hydrology; Lagoons; Land management; Land use; Land use management; Land use planning; Livestock; Livestock grazing; Machine learning; Neural networks; Prediction models; Predictions; Recurrent neural networks; Regression analysis; Seasonal variations; Seasons; Sediments; Soil erosion; Soil water; spatio-temporal prediction; Time series; time series analysis; Trends; Vegetation; Water quality; Australia; Great Barrier Reef
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs16173193

Livestock grazing is a major land use in the Great Barrier Reef Catchment Area (GBRCA). Heightened grazing density coupled with inadequate land management leads to accelerated soil erosion and increased sediment loads being transported downstream. Ultimately, these increased sediment loads impact the water quality of the Great Barrier Reef (GBR) lagoon. Ground cover mapping has been adopted to monitor and assess the land condition in the GBRCA. However, accurate prediction of ground cover remains a vital knowledge gap to inform proactive approaches for improving land conditions. Herein, we explored two deep learning-based spatio-temporal prediction models, including convolutional LSTM (ConvLSTM) and Predictive Recurrent Neural Network (PredRNN), to predict future ground cover. The two models were evaluated on different spatial scales, ranging from a small site (i.e., <5 km2) to the entire GBRCA, with different quantities of training data. Following comparisons against 25% withheld testing data, we found the following: (1) both ConvLSTM and PredRNN accurately predicted the next-season ground cover for not only a single site but also the entire GBRCA. They achieved this with a Mean Absolute Error (MAE) under 5% and a Structural Similarity Index Measure (SSIM) exceeding 0.65; (2) PredRNN superseded ConvLSTM by providing more accurate next-season predictions with better training efficiency; (3) The accuracy of PredRNN varies seasonally and spatially, with lower accuracy observed for low ground cover, which is underestimated. The models assessed in this study can serve as an early-alert tool to produce high-accuracy and high-resolution ground cover prediction one season earlier than observation for the entire GBRCA, which enables local authorities and grazing property owners to take preventive measures to improve land conditions. This study also offers a new perspective on the future utilization of predictive spatio-temporal models, particularly over large spatial scales and across varying environmental sites.