Abstract. Soil moisture, an essential parameter for hydroclimatic studies, exhibits significant spatial and temporal variability, making it challenging to map at fine spatiotemporal resolutions. Although current remote sensing products provide global soil moisture estimate at a fine temporal resolution, they are mostly at a coarse spatial resolution. In recent years, deep learning (DL) has been applied to generate high-resolution maps of various soil properties, but DL requires a large amount of training data. This study aimed to map daily soil moisture across Tasmania, Australia at 80 meters resolution based on a limited set of training data. We assessed three modelling strategies: DL models calibrated using an Australian dataset (51,411 observation points), models calibrated using the Tasmanian dataset (9,825 observation points), and a transfer learning technique that transferred information from Australian models to Tasmania. We also evaluated two DL approaches, i.e. Multilayer perceptron (MLP) and Long Short-Term Memory (LSTM). Our models included data of Soil Moisture Active Passive (SMAP) dataset, weather data, elevation map, land cover and multilevel soil properties maps as inputs to generate soil moisture at the surface (0–30 cm) and subsurface (30–60 cm) layers. Results showed that (1) models calibrated from the Australia dataset performed worse than Tasmanian models regardless of the type of DL approaches; (2) Tasmanian models, calibrated solely using Tasmanian data, resulted in shortcomings in predicting soil moisture; and (3) Transfer learning exhibited remarkable performance improvements (error reductions of up to 45 % and a 50 % increase in correlation) and resolved the drawbacks of the Tasmanian models. The LSTM models with transfer learning had the highest overall performance with an average mean absolute error (MAE) of 0.07 m³m^-3 and a correlation coefficient (r) of 0.77 across stations for surface layer and MAE = 0.07 m³m^-3, and r = 0.69 for subsurface layer. The fine-resolution soil moisture maps captured the detailed landscape variation as well as temporal variation according to four distinct seasons in Tasmania. The best performance of soil moisture models were made available live to predict near-real-time daily soil moisture of Tasmania, assisting agricultural decision making.

中文翻译：

---

通过迁移学习绘制塔斯马尼亚近实时土壤湿度动态图


摘要。土壤湿度是水文气候研究的重要参数，表现出显着的空间和时间变化，使得以精细的时空分辨率绘制地图具有挑战性。尽管当前的遥感产品以精细的时间分辨率提供全球土壤湿度估计，但它们大多以粗略的空间分辨率提供。近年来，深度学习（DL）已被应用于生成各种土壤属性的高分辨率图，但深度学习需要大量的训练数据。本研究旨在基于一组有限的训练数据，以 80 米的分辨率绘制澳大利亚塔斯马尼亚岛的每日土壤湿度图。我们评估了三种建模策略：使用澳大利亚数据集（51,411 个观察点）校准的深度学习模型、使用塔斯马尼亚数据集（9,825 个观察点）校准的模型以及将信息从澳大利亚模型转移到塔斯马尼亚的转移学习技术。我们还评估了两种深度学习方法，即多层感知器（MLP）和长短期记忆（LSTM）。我们的模型包括土壤湿度主动被动（SMAP）数据集、天气数据、海拔图、土地覆盖和多级土壤属性图作为输入，生成地表（0-30厘米）和地下（30-60厘米）的土壤湿度层。结果表明：(1) 无论深度学习方法的类型如何，根据澳大利亚数据集校准的模型的表现都比塔斯马尼亚模型差； (2) 塔斯马尼亚模型仅使用塔斯马尼亚数据进行校准，在预测土壤湿度方面存在缺陷； (3) 迁移学习表现出显着的性能改进（错误减少高达 45%，相关性增加 50%），并解决了塔斯马尼亚模型的缺点。 采用迁移学习的 LSTM 模型具有最高的整体性能，表层的平均绝对误差 (MAE) 为 0.07 m ³ m ^-3 ，相关系数 (r) 为 0.77，MAE = 0.07 m ³ m ^-3 ，对于地下层，r = 0.69。高分辨率土壤湿度图捕捉了塔斯马尼亚四个不同季节的详细景观变化以及时间变化。土壤湿度模型的最佳性能可实时预测塔斯马尼亚的每日土壤湿度，协助农业决策。