Title: Overcoming Data Scarcity in Land Use Monitoring with Time-Weighted Dynamic Time Warping Slides 📎
Land use monitoring using machine learning and Earth observation data is usually challenging due to the lack of training samples, especially for large areas and long periods where gathering in-situ information is costly or sometimes impossible. This work proposes a machine learning approach called Time-Weighted Dynamic Time Warping (TWDTW) for data-scarce applications. TWDTW is a satellite image time series classification algorithm that uses a Dynamic Time Warping (DTW) distance. DTW is a widely used algorithm in various fields, including speech recognition, medicine, industry, and finance, and has shown promising results in land use mapping due to its ability to deal with gaps in time series, robustness to noise, matching time series of different lengths and intervals, and to keep its classification performance on small training sets.
However, DTW has limitations in matching events regardless of when they occur, which can result in out-of-season alignments and misclassifications—for example, aligning a summer crop to a winter one. TWDTW overcomes this limitation by introducing a time weight to matches deviating from an expected date in the training set. This temporal constraint improves classification performance by controlling for out-of-season alignments while keeping DTW’s flexibility to smaller phenological fluctuations of vegetation.
This presentation will demonstrate the effectiveness of the TWDTW method for land use classification using the open-source R package dtwSat. Overall, this machine learning method is suitable for data-scarce regions and can contribute to land use monitoring, supporting the environmental targets proposed by the European Green Deal and the United Nations’ Sustainable Development Goals.