Search

• A Machine Learning method is developed by integrating remotely sensed feedback of vegetation to antecedent rainfall to forecast seasonal soil moisture. • The method showed (>50%) and (>80%) variability in root-zone(0–1 m) and deep (1–6 m) soil can be explained 3 months ahead of time. • Integrating remotely sensed vegetation feedback to antecedent rainfall has improved soil moisture forecast by up to 33.3% • This model could be a steppingstone to developing and deploying a computationally light operational seasonal soil moisture forecast model. Seasonal forecast of soil moisture at large spatial scale over forested landscape has numerous implications in forest hydrology and bushfire risk planning. Remotely sensed plant response to rainfall, input meteorological forcing, and site-specific landscape attributes were integrated into a data-driven Gradient Boosting Machine Learning (ML) model to forecast summer season (December to February) soil moisture equivalent to those from Australian Water Resource Assessment-Landscape (AWRA-L) at root-zone (0–1 m) and deep (1–6 m) layers of the soil. Multispectral and thermal infrared bands from MODIS (Band 1–12, LST (day, night)) and meteorological forcing for 2000 – 2018 during and prior to winter (before August) and site-specific landscape attributes were used to generate the explanatory input variables to the model. Spatial and temporal forecasting skills of the model were evaluated using two types of cross-validation: two-fold cross-validation by space over the entire period (2000–2018) and two-fold cross-validation by time for all grid cells (2160 cells). In the first method, the model showed a high skill to forecast deep soil moisture (R2 = 0.81, NSE = 0.79, RMSE = 58.01 mm) and root-zone soil moisture (R2 = 0.65, NSE = 0.63, RMSE = 24.34 mm). In comparison, the second evaluation resulted in more accurate forecast deep-soil moisture (R2 = 0.88, NSE = 0.86, RMSE = 47.39 mm) but in lower performance for root-zone soil (R2 = 0.55, NSE = 0.47, RMSE = 29.2 mm). These outcomes have highlighted that the (>80%) variability in summers deep (1–6 m) soil moisture and (>50%) variability in root-zone (0–1 m) soil moisture over forested landscape can be explained at the end of winters. Furthermore, integrating remotely sensed observations has improved the results by showing an enhancement in the RMSE by 22.8–33.3% in the deep (1–6 m) soil moisture and 8.8–14.3% in root-zone (0–1 m) soil moisture. Overall, the integrated system using remotely sensed plant response, climate forcing and machine learning exhibit great potential to forecast summer soil moisture three months ahead in the Mediterranean climates. [ABSTRACT FROM AUTHOR]

• The effect of training samples on within-season crop classification was examined. • The sample size should be sufficiently large in within-season crop classification. • Proportions of each crop type should be similar for training and application. • Training using samples from previous years led to a minor decline in accuracy. • Overall classification accuracy exceeded 86 % for the third and later months. Within-season crop classification using multispectral imagery is an effective way to generate timely crop maps that can support water and crop management; however, developing such models is challenging due to limited satellite imagery and ground truth data available during the season. This study investigated ways to optimize the use of multi-year samples in a within-season crop classification model, aiming to enable accurate within-season crop mapping across years. Our study focused on classifying field-scale corn/maize, cotton, and rice in south-eastern Australia from 2013 to 2019. The crop classification model was based on the random forest and support vector machine algorithms applied to Landsat 8 multispectral bands. We designed four experiments to understand the influences of training sample selection on model accuracy. Specifically, we analyzed how the within-season classification accuracies are affected by 1) training sample size; 2) proportions of classification classes; 3) the inclusion of a non-crop class (e.g., fallow land) in the training sample, and 4) training samples collected from different years. We found that 1) the training sample size should be sufficiently large to ensure within-season classification accuracy; 2) using training samples for each crop type in proportion to their occurrence within the landscape results in more accurate multi-year classification; 3) the inclusion of the non-crop class can reduce the accuracy with which crop types are distinguished, so the proportion of the non-crop class should be maintained at a relatively low level, and 4) predicting the current year with training samples from previous years can lead to a minor decline in accuracy compared to using samples only from the current year. These training sample settings were adopted to develop a final model. We found that the model accuracy continues to improve as more input imagery is added as the cropping season progresses, with a rapid rate of initial improvement which then slows. December, the third month of the summer growing season, is the earliest time that reliable maps were generated, with an overall accuracy of 86 % and user's accuracies for all crops exceeding 80 %. Our proposed experiments are robust and transferable to other regions and seasons to assist the development of within-season crop maps, and can thus be valuable tools to support agricultural management. [ABSTRACT FROM AUTHOR]

Real-time crop canopy nitrogen concentration (CNC) and aboveground biomass (AGB) sensing capability can enable precision agriculture with significant economic and ecological benefits. Canopy spectral response in visible near-infrared (VNIR, 400–979 nm) has been widely used to estimate CNC and AGB of canopies but is often confounded by the soil background, crop type, phenological changes and environmental conditions. Machine (ML) and deep learning (DL) have emerged as promising techniques for mapping CNC and AGB using canopy spectral reflectance in VNIR. However, ML and DL models' performance relies on extensive labelled training datasets, which are often limited due to being expensive and labour intensive. We estimated CNC and AGB using state-of-the-art hyperspectral (HS) modeling approaches using partial least squares regression (PLSR), random forest (RF), and one- and two-dimensional convolutional neural network (1D and 2D CNN). These models, together with multispectral (MS) approaches, were evaluated in the context of a small (422) dataset for model development collected from diverse fertilization and contrasting growth conditions. A handheld field spectroradiometer was used to collect spectral reflectance samples across a wide range (more than ten-fold difference in AGB) of canopy development stages in two seasons (winter and summer) in ryegrass, and across phenological stages (from GS-25 to GS-60) in barley. Results can be summarized in four components: (1) the median values of performance metrics in 1D CNN (R2 test =0.83, NRMSE test =0.15 for CNC; R2 test =0.79 (ryegrass), 0.61 (barley), NRMSE test = 0.29 (ryegrass), 0.30 (barley) for AGB) and PLSR (R2 test =0.71, NRMSE test = 0.20 for CNC; R2 test =0.62 (ryegrass), 0.70 (barley), NRMSE test = 0.38 (ryegrass), 0.27 (barley) for AGB) showed the best performance for CNC and AGB prediction in the test data; (2) when evaluated with the independent measurements, 1D CNN demonstrated the best (up to 56% decrease in NRMSE) performance for AGB and was comparable to PLSR for CNC (up to 25% decrease in NRMSE than their HS and MS counterparts); (3) predictive models underperformed at low canopy cover for CNC estimation and were adversely affected by the high biomass saturation effects for AGB prediction; and (4) visible spectrum is relatively more informative to CNC models whereas the red edge to NIR spectrum is more informative to AGB models. This study demonstrates the potential and limitations of commonly available field VNIR spectral reflectance data for CNC and AGB mapping across contrasting conditions in a small dataset (<500). • 1D convolutional neural network (CNN) exhibited the best and robust performance. • 1D CNN showed a relatively higher tolerance for biomass saturation effects. • Canopy nitrogen concentration (CNC) models underperformed at low canopy cover. • Visible spectrum was more informative for CNC prediction. • Red edge to NIR spectrum showed more relevance for aboveground biomass prediction. [ABSTRACT FROM AUTHOR]