Integrated field and satellite based decision support system for climate-resilient and sustainable ranches and rangelands across California

This project focused on California’s rangelands,  where climatologically, ecologically,  and topographically diverse landscapes pose the biggest challenge for sustainable ranching (Fig. 1).  

Research Objective 1. Build an extensive field-based geospatial database on the aboveground biomass metrics over California’s rangeland, to be used to calibrate and evaluate the remote sensing-based monitoring algorithm.

In collaboration with UCCE and FSA collaborated (FSA funded field supplies and Loss Adjuster labor cost), we identified and established approximately 40 monitoring sites (700 m2) at the beginning of the project, providing a representative distribution of annual rangelands across all sub-ecoregions statewide (Fig. 1). In late spring of 2022, we reconvened our partners, revisited the exclosure site selection, and revised sampling and measurement protocol, based on the experiences during previous field season (September 2011 to March 2022). A total of 42  sites were finalized,  representing the diversity of the annual grasslands across all sub-ecoregions statewide (Fig. 1). The exact locations of each sites were recorded by high precision Trimble GPS. To match the footprint size of available moderate to high spatial resolution satellite remote sensing data (~10 to 30 m), each site was situated in a large open grassland (> 45 x 45 m, a.k.a., Macro plots) and includes one large exclosure (14.63 x 14.63 m) to accommodate remote sensing and field clip-plot quantification. A total of 17 county based UCCE farm advisors and 5 regional FSA “Loss Adjusters” collected field data in 2021, 2022, and 2023.

Forage clippings were done once or twice a month from January to March to capture the rapid growth period, once in April/May to capture peak standing crop, and once from late September or October to capture RDM. Plant functional group composition (percent grass vs forb cover) was recorded at each sampling event (2021/2022 and 2022/2023 seasons); additionally, dominant species were recorded via the dry-weight-rank method at each site at peak standing crop (2022/2023 season). For the 2021/2022 season, clippings were specifically collected within non-grazed exclosures and in adjacent grazed areas. A subset of samples were sorted to separate new growth from old-growth (i.e., RDM) to measure the current year’s forage biomass to match satellite based estimations. This approach allowed us to assess the ability of the remote sensing platform to quantify forage biomass in paired grazed vs non-grazed areas, which can differ in build-up of RDM. For the 2022-2023  season, we continued to clip forage samples within grazing exclosures, field sorting samples with greater than 15% RDM cover.

For each monitoring site, 10-15 samples were clipped during each clipping event at each site. The clipping samples were averaged by date for each site, resulting in a total of 150 site-date data points for further analysis in Year 1. During Year 2 of the project, we continued forage clipping for another season following the same protocol, resulting in a  total of 278 field clipping samples for model optimization (Obj. 2). RDM clipping was repeated during September - November, and a total of 85 data points were collected for RDM analysis and model building. Additionally, we also conducted quality assurance and quality control on all clipping data spanning three years. Overall, 32 sites had consistent field sampling between February and May in 2021, 2022, and 2023. By following the same protocols for sampling, field data collection, and preprocessing, we built a large geospatial database of forage production with high quality and consistency.

Research Objective 2. Develop a well-calibrated remote sensing algorithm scalable statewide for consistent forage production estimates in near real-time. 

With this unique field dataset, we improved upon the remote sensing approach that we developed for forage production estimation using MODIS daily 250-m and Landsat 16-day 30-m satellite observations (Liu et al., 2021). Our prior approach was calibrated with limited exclosure site data. Moreover, it was applied only to herbaceous-dominated rangelands, to avoid the mixed woody-herbaceous pixel issue at 30-m resolution. In order to provide daily estimates for the majority of rangelands without the masking effects from shrubs and trees, higher spatial resolution imagery is needed as a main data stream. We  relied on Sentinel 2A&B at 10 m every 5 days (plus Landsat), both publicly-available and free, for operational near real-time forage monitoring (Fig. 2). 

Our focus was on all 10 ×10-m pixels without trees or shrubs. A first version of the base mask for herbacious cover at 10m was generated by first combining herbaceous or grasslands pixels from CalFire’s vegetation map and the 2023 National Land Cover Database (NLCD). Only pixels with NDVI lower than 0.3 in mid-August were retained for further analysis since the presence of trees and other perennial vegetation within a 10m pixel would increase the NDVI value compared to senescent grasses.  

To estimate forage production, we adopted the similar biophysical light use approach, developed from our previous studies (Liu et al., 2019, 2021), and further improved the accuracy and robustness, especially with higher spatial and temporal resolution of Sentinel 2 satellite imagery than previously used Landsat 30m imagery, and leveraging the substantial field data from the newly established network of the exclosure sites for model training and testing.

We assembled weather data,  including incoming solar radiation, temperature, precipitation,  from CIMIS and PRISM, and the topographical indices such as elevation and soil data such as soil organic matter. We have downloaded Sentinel 2A & 2B imagery at 10m to 20m with repeated observations every 5 days from 2021 to May 2024.  Images with more than 30% cloud cover were removed from the analysis.  An automatic preprocessing workflow has been developed to generate a daily time series of remote sensing metrics, e.g., the Savitzky-Golay filter was applied to the original NDVI time series to fill the gap and remove the abnormal values; a linear interpolation was then performed to generate the daily time series of NDVI.

For each individual 10 m pure grass pixel, we have already identified the phenological metrics in California annual grasslands such as the start of the growing season (SOS), end of the growing season (EOS), and length of the growing season (LOS).  The Sentinel-2 NDVI time series of each pixel was fitted with two sigmoidal curves (for dates before and after NDVI reached peak value) and the rate of change was then used to identify the SOS and EOS and estimate the growing season length, which was fed into the daily accumulation of forage production.

We improved the light use efficiency optimization with the available field data and extracted Sentinel-2 data. The baseline model was based on a constant light use efficiency determined by the accumulated absorbed photosynthetically active radiation (APAR) and the field measured forage production before and at the peak production. The improved optimization of LUE factored the impact of environmental stressors such as temperature and moisture stressor. In particular, the moisture scalar was represented by the Land Surface Water Index (LSWI), a remote sensing indicator sensitive to vegetation water content and soil moisture based on thee contrast of the surface reflectances in Sentinel 2 band B8A (Red Edge 4) and in the shortwave infrared band (B11). 

The model parameters, including LUEmax and Topt, were optimized by the Stochastic Gradient Descent (SGD) algorithm, which updates model parameters via randomly selected subsets of the training data . We used 70% of the field data, randomly selected, for the optimization process, and the remaining 30% of the data for validation. Model performance was evaluated by the coefficient of determination (R2) and the root mean squared error (RMSE). Independent model validation was also conducted using the field data from the Pepperwood Preserve. Additionally, we also compared the herbaceous biomass data from field clipping and our estimation with those from the existing Rangeland Analysis Platform (RAP).

We applied the optimized light use efficiency model statewide and generated the forage production maps on Google Earth Engine (GEE). To characterize interannual variability, the temporal dynamics of forage production were analyzed for two recent years (2023 and 2024). In addition, San Luis Obispo (SLO) county and the Pepperwood Preserve were selected for case studies to further examine the spatial patterns and interannual variability of forage production from 2021 to 2024. Two years of statewide annual forage production was also generated to test the scalability of our approach. These products were shared with county advisors. 

Research Objective 3. Develop a locally-optimized RDM estimation tool with multi-sensor imagery. 

 Recent work on variability and loss in RDM over the summer dry period (Larsen et al., 2020) suggests important implications for using remote sensing estimates of both growing and non-growing season above-ground biomass. To map RDM operationally during the dry season, our goal was to develop an innovative and scalable approach using locally optimized machine learning algorithms. We designed a two step approach:  (1) scaling up the field measured RDM with unique hyperspectral measurements by PRISMA imager, operated by ASI (Agenzia Spaziale Italiana), over selected large areas, and (2)  wall to wall cost-effective RDM mapping with the calibrated machine earning models to integrate improved growing season production estimates, factors related to decomposition, and coincident Landsat/Sentinel imagery. 

We regularly submitted tasking orders  to acquire PRISMA hyperspectral imagery at 30-m resolution from August 2022 to cover our monitoring sites. However, due to the low priority of US researchers to acquire PRISMA data, only a total of 10 and 11 scenes were acquired at a 30 × 30 km swath in 2022 and 2023, respectively. We also acquired another hyperspectral data set from NASA EMIT satellite.  We are still exploring the relationship between cellulose absorption features at ~2100 nm and many other spectral features, such as Cellulose Absorption Index and the traditional broadband indices (NDRI) with field measured RDM. Work is ongoing to build a statistical model to identify the most important hyperspectral features and indices for RDM estimation across sites and estimate RDM with PRISMA/EMIT data. The calibrated model will then be applied to all herbaceous-dominated 30-m pixels to map RDM for each PRISMA scene. We will build a separate model to incorporate those broadband indices available from Landsat and Sentinel in fall, preceding weather, solar energy, and topography for a more robust RDM estimation, and then evaluate the improvement of PRISMA-based RDM method over existing approaches. 

Research Objective 4. Prototype an online interactive app to help stakeholders improve ranch management and risk mitigation decision-making.

We have tested the capability of our approach in generating daily, weekly, and annual forage production at 10m across all California’s annual grassland. The calibrated algorithm for statewide application has been implemented on Google Earth Engine (GEE). This workflow automatically extracts and preprocesses remote sensing data from  Sentinel 2 at 10m every 5 days. Time series of these remote sensing metrics were then integrated with weather data to automatically estimate forage production at daily time scale.

As an important aspect related to ranch management and risk mitigation decision-making, we analyzed the year-to-year fluctuation of historical forage production derived from remote sensing observations. The coefficient of variation (CV), defined as the ratio of the standard deviation divided by the mean,  was used to represent the interannual variability in forage production as well as other vegetation and climatic factors. The 8-year moving window from 2001 to 2018 was applied to reduce the uncertainties caused by the disturbances such as drought. We applied the Mann-Kendall and the Theil-Sen methods to test the significance of  and quantify the trends of the variability of forage production, for each 30m pixel, and then summarized over each sub-ecoregion. To further understand the trends of variability in forage production, the random forest model was built using a suite of remote sensing indices, climate indices as well as soil and topographical indices were used as predictors. The partial dependence plot was used to quantify how the trends of variations of forage production vary with each independent variable. 

We have obtained the inputs from ranchers and local UCCE advisors on the key functional needs of an interactive API. A user-friendly dashboard will be designed to allow users to query, visualize, and summarize forage production and RDM at various scales. Various tabs will also permit users to examine spatial heterogeneity within and across management units. By combining with our previous historical daily forage production estimates at 30m resolution (Liu et al., 2020), end users can visualize deviations from average production (anomalies) at ranch, region, or state-wide scales, and make adaptive decisions.