Yield Estimates by a Two-Step Approach Using Hyperspectral Methods in Grasslands at High Latitudes

Ruminant fodder production in agricultural lands in latitudes above the Arctic Circle is constrained by short and hectic growing seasons with a 24-hour photoperiod and low growth temperatures. The use of remote sensing to measure crop production at high latitudes is hindered by intrinsic challenges, such as a low sun elevation angle and a coastal climate with high humidity, which influences the spectral signatures of the sampled vegetation. We used a portable spectrometer (ASD FieldSpec 3) to assess spectra of grass crops and found that when applying multivariate models to the hyperspectral datasets, results show significant predictability of yields (R2 > 0.55, root mean squared error (RMSE) < 180), even when captured under sub-optimal conditions. These results are consistent both in the full spectral range of the spectrometer (350–2500 nm) and in the 350–900 nm spectral range, which is a region more robust against air moisture. Sentinel-2A simulations resulted in moderately robust models that could be used in qualitative assessments of field productivity. In addition, simulation of the upcoming hyperspectral EnMap satellite bands showed its potential applicability to measure yields in northern latitudes both in the full spectral range of the satellite (420–2450 nm) with similar performance as the Sentinel-2A satellite and in the 420–900 nm range with a comparable reliability to the portable spectrometer. The combination of EnMap and Sentinel-2A to detect fields with low productivity and portable spectrometers to identify the fields or specific regions of fields with the lowest production can help optimize the management of fodder production in high latitudes.


Introduction
Ruminant milk and meat production above the Arctic Circle (~66.34 • N) takes place under unique climatic conditions, with long winters and short intense growing seasons. Forage is produced on agricultural grasslands at these latitudes with an average of a 132-day growing season, 689 growing degree days (using 5 • C as a base in the May-September period, as an average between 1989-2017, Agrimeteorology Norway, Holt station, Tromsø), resulting in a maximum of two harvests per year. Economic profitability for milk and meat production in this region is highly dependent on efficient forage production. Information on yields can help improve the management of the individual grass fields by e.g., assisting in decisions on stocking rates, fertilising, timing of harvest according to yield, identifying and intervening on low yielding areas and purchase of appropriate supplemental feed. attributes of grassland, help detect regions, fields or parts of fields in need of different management strategies, where the use of resources at a local scale can be optimized afterwards [32]. While the applicability of satellites in crop management is very promising [33], the environmental conditions present in high latitudes can limit the robustness of predictions derived from satellite data, thus introducing latitude-specific biases. Combining satellite data with portable devices such as unmanned aerial vehicles (UAVs) or portable spectrometers gives the advantage that data can be collected at higher temporal and spatial resolution, with fewer restrictions on cloud conditions. Satellite derived data can help find areas or fields in need of management actions, such as fertilization; afterwards, portable spectrometers can be used to achieve an optimized fertilization process [34,35] and ensure a more sustainable crop management [36].
Our study aims to provide stakeholders with better information systems on their agricultural production and test the applicability of a portable spectrometer in high latitudes (northern Norway) to estimate grass yield under varying weather conditions. The study investigates the use of multivariate approaches using data collected over four growing seasons to find a solution for the complex environmental variables present in the fields. In addition, we suggest a two-stage workflow to use both Sentinel-2 and the upcoming EnMap satellite data as a screening tool to detect fields with reduced productivity, which can subsequently be assessed in further detail with very high-resolution hyperspectral tools (i.e., portable spectrometers or UAV-mounted sensors) to create detailed maps of the variability in productivity within fields.

Site Description
We measured 254 reference squares (area = 0.25 m 2 ) during four years of data acquisition at different times of the growing season (Table 1). In 2014, 2015, and 2017, fields close to Tromsø (69 • 39 N 18 • 57 E) were used, while in 2016 we used fields located further south, close to Harstad (68 • 48 N 16 • 32 E), in northern Norway ( Figure 1). The three fields used in 2014 and 2015 were measured at three different dates from early in the growing season until days before the first harvest to ensure a large variation in yields and plant phenological status. The fields used in 2016 and 2017 were measured within a week before the first harvest. Grasses dominated the fields (Table 1), with a high presence of the weed Ranunculus repens in some fields. The weather conditions during data acquisition differed between years. Measurements in 2014, 2015 and 2016 were made with sunny weather or light overcast (hereafter, the "good weather" dataset). However, in 2017, weather conditions were challenging since it had been raining for several days, and the weather forecast indicated more rain in the coming days. Weather was cold and cloudy with light rain during sampling, which was performed to estimate a worst-case model predictive ability scenario (hereafter the "bad weather" dataset).

Spectra Acquisition
Reflectance spectra were recorded using a FieldSpec 3 portable spectrometer (ASD Inc., Boulder, Colorado) in a spectral range of 350-2500 nm, with a 1 nm interpolated resolution. The recorded spectra were internally interpolated by the spectrometer, with a resolution of 1.4 nm in the 350-1050 nm range (named SWIR-1 sensor in the instrument specification) and 2 nm in the 1050-2500 nm range (named SWIR-2 sensor in the instrument specification). Each spectrum utilized in the data analyses was the average of five independent samples taken from a nadir orientation from a 1.20 m height resulting in a 0.25 m 2 area. We used the fiber optic probe attached to the pistol grip, which has a field of view of 25 • . Sampling points were distributed to maximize the amount of points per plot in the available fieldwork time. Cross-transects were laid out in each field in the years 2014-2016 to cover the full extent of the field while constraining the time needed to perform the scans, thus ensuring a maximum spatial spread of sampled points to maximize the spectral and yield variability between sampled points. We performed grid transects in the year 2017 to ensure a high number of points during the sampling time, which was characterized by challenging weather conditions and rain showers shortly after the spectra acquisition was planned to be finalized. A white reference was taken before capturing the spectra for each point for instrument optimization. In addition to the full spectrum dataset (i.e., 350-2500 nm spectral range), we also developed models limiting the spectral range to the 350-900 nm region, as this region of the electromagnetic spectrum is more robust against spectral interferences produced by water vapor.

Reference Yield Measurements
Each 0.25 m 2 area that was scanned with the spectrometer was immediately cut at a height of 5 cm from the ground to simulate yield, then transported in mesh bags to dry at 60 • C for 48 h, which Remote Sens. 2019, 11, 400 6 of 15 is the standard practice for determining dry weight of fresh forage samples [37]. Dry weight of the yield was measured as reference weight for the models. We use the term yield in this study instead of above ground biomass as our measurements represent the yield a farmer would get from the field.

Hyperspectral Analyses and Modelling
Partial least squares regression (PLS) models were performed in R 3.4.2 [38] for all model development. A total of six models were developed (three models with the full spectrum, and three with the 350-900 nm spectrum) to study how removal of moisture-affected spectra affects the predictive ability of yield models. First, we ran a model including only the "good weather" plots (i.e., direct sunlight and very low air moisture, plots from 2014 to 2016), second, we ran a model based on the "bad weather" plots (samples from 2017), and third, we ran a model containing all the points (i.e., the "full model").
In both the FieldSpec models and the satellite simulations, spectra were presented as raw spectra or pre-treated with the standard normal variate (SNV) treatment [39], moving average filter or Savitzky-Golay derivatives [40] and modeled with PLS [41], contained in the package pls in the R Software [42]. The optimal model and number of latent variables (k) were selected based on achieving a combination of a high R 2 and a low root mean squared error of the prediction (RMSEP).

Satellite Simulation
We simulated the 13 Sentinel-2A bands from the FieldSpec dataset and created a predictive model for yield to estimate how the present satellites can predict grassland productivity. The bands were simulated by binning the original spectra following the specifications available on the Sentinel-2A mission website [43] using the prospectr package [44] in R [38]. All the bands were used for modeling: we did not create a limited range (i.e., 400-900 nm range) model due to the low number of bands available for model development.
In addition, we simulated the 240 EnMap bands and modeled the yield using the full spectral resolution of the simulated satellite sensor (i.e., 420-2450 nm) to estimate how accurate remote-sensed yield models might be using EnMap data compared to the Sentinel-2A modeling results. For that purpose, binning was performed on the original spectra following the specifications given on the official EnMap website (https://eoportal.org), i.e., one data point every 6.5 nm in the very near infrared VNIR region (420-1000 nm) and 10 nm in the SWIR short wave infrared regions (1000-2450 nm). In addition, in the case of EnMap, we developed a model using only the region between 420 and 900 nm to avoid the regions most affected by water and compare with the results of the model with the full spectral range.

Model Validation
The models assessing the effect of the environmental conditions (i.e., cloud cover and moisture) were validated against each other, i.e., the "good weather" model was used to predict the "bad weather" dataset and vice versa. The dataset for the "full model" and the satellite simulations were split into calibration and validation datasets (75% of the samples for calibration and 25% for validation) by first ordering the samples according to their yield and then sequentially assigning three samples for calibration and one for validation. Y-outliers (extreme predicted values) were identified when the fitted or predicted values showed abnormal values (i.e., outside the calibration range).

Two-Stage Workflow
We used a two-stage workflow ( Figure 2) to simulate our suggested application of this method, where farmers can first identify fields with reduced production at a large scale based on satellite data. The fields with reduced production can afterwards be studied with a portable spectrometer to find the areas in need of management strategies (e.g., fertilization or reestablish the grass sward in damaged parts, or other actions), in order to allocate time and economic resources more efficiently.

Two-Stage Workflow
We used a two-stage workflow ( Figure 2) to simulate our suggested application of this method, where farmers can first identify fields with reduced production at a large scale based on satellite data. The fields with reduced production can afterwards be studied with a portable spectrometer to find the areas in need of management strategies (e.g., fertilization or reestablish the grass sward in damaged parts, or other actions), in order to allocate time and economic resources more efficiently.

Results
A total of 254 samples were registered with grass yields that varied between 0.55 g/m 2 and 1212 g/m 2 : the quadrats scanned under optimal weather conditions (i.e., "good weather" datasets, from 2014 to 2016) had a wider range (0.55 g/m 2 to 1212 g/m 2 ) than the quadrats scanned under challenging weather conditions (i.e., "bad weather" dataset, of 2017), which had a narrow range of crop productivity (44 g/m 2 to 680 g/m 2 ) (Figure 3).

Results
A total of 254 samples were registered with grass yields that varied between 0.55 g/m 2 and 1212 g/m 2 : the quadrats scanned under optimal weather conditions (i.e., "good weather" datasets, from 2014 to 2016) had a wider range (0.55 g/m 2 to 1212 g/m 2 ) than the quadrats scanned under challenging weather conditions (i.e., "bad weather" dataset, of 2017), which had a narrow range of crop productivity (44 g/m 2 to 680 g/m 2 ) (Figure 3).  The year 2016 dataset was gathered in a location with high productivity, and that year was characterized by an extraordinary productivity that is unusual in the studied fields.

Hyperspectral Analyses
In the full spectrum models (i.e., the 350-2500 nm spectral range), the "good weather" model showed a good fit, with an R 2 value of 0.82 and RMSE of the cross validation (RMSECV) of 135 g/m 2 in the calibration dataset. When predicting the "bad weather" dataset using the "good weather" model, the R 2 value was 0.4 and a RMSEP of 286 g/m 2 (Table 2, Supplementary Figure S1). The "bad weather" model showed a lower fit, with an R 2 value of 0.37 and RMSECV of 86 g/m 2 in the calibration dataset. When predicting the "good weather" dataset, the R 2 value was of 0.48 and RMSEP of 105 g/m 2 (Table 2, Supplementary Figure S1). The "full model" (i.e., the model including The year 2016 dataset was gathered in a location with high productivity, and that year was characterized by an extraordinary productivity that is unusual in the studied fields.

Hyperspectral Analyses
In the full spectrum models (i.e., the 350-2500 nm spectral range), the "good weather" model showed a good fit, with an R 2 value of 0.82 and RMSE of the cross validation (RMSECV) of 135 g/m 2 in the calibration dataset. When predicting the "bad weather" dataset using the "good weather"  Figure S1). The "bad weather" model showed a lower fit, with an R 2 value of 0.37 and RMSECV of 86 g/m 2 in the calibration dataset. When predicting the "good weather" dataset, the R 2 value was of 0.48 and RMSEP of 105 g/m 2 (Table 2, Supplementary Figure S1). The "full model" (i.e., the model including all fields and years), on the other hand, showed a consistent predictive ability on both the calibration dataset (R 2 = 0.56, RMSECV = 166 g/m 2 ) and validation dataset (R 2 = 0.58, RMSEP = 178 g/m 2 ) (Table 2, Figure 4). Note that the "bad weather" points were evenly distributed along the calibration line.
In the "limited spectrum" models (i.e., limited to the 350-900 nm range), the "good weather" model showed a good fit, with an R 2 value of 0.82 and RMSECV of 135 g/m 2 in the calibration dataset. When predicting the "bad weather" dataset, the R 2 value was 0.36 and RMSEP 338 g/m 2 in the validation dataset, which is substantially worse than with the full spectrum model ( Table 2, Supplementary Figure S2). The "bad weather" model showed a similar performance to its full spectrum counterpart, with an R 2 value of 0.42 and RMSECV of 82 g/m 2 in the calibration dataset. When predicting the "good weather" dataset, the R 2 value was 0.05 and RMSEP 192 g/m 2 , which is also worse than the model with the full spectrum (Table 2, Supplementary Figure S2). The "full model", on the other hand, showed a consistent predictive ability on both the calibration dataset (R 2 = 0.69, RMSECV = 140 g/m 2 ) and validation dataset (R 2 = 0.72, RMSEP = 144 g/m 2 ) (Table 2, Figure 4). Again, the "bad weather" points were evenly distributed along the calibration line. This model performed better than the model with the full spectrum, when judged against both the calibration and validation datasets.
There are very strong limitations on using models developed under different environmental conditions, where predicting one dataset (e.g., the "good weather" dataset) with a model developed with contrasting environmental conditions (e.g., the "bad weather" dataset) will result in unreliable predictions, due to large error rates and nonlinearities in the predictions that render the predictions unusable (see Supplementary Figures S1 and S2). Thus, using models including all environmental conditions yielded more robust results that provide more reliable predictions. Table 2. Modeling parameters from the partial least squares (PLS) models developed for the full spectral range (350-2500 nm) and the limited range (350-900 nm). SNV stands for the standard normal variate treatment, k represents the number of latent variables used, m represents the differentiation order, p represents the polynomial order and w represents the window size for the smoothing function. RMSECV represents root mean squared error of the cross-validation and RMSEP represents root mean squared error of the prediction. * Note that the validation dataset for the "good weather" model is the "bad weather" data, and the validation for the "bad weather" model is the "good weather" data. unusable (see Supplementary Figures S1 and S2). Thus, using models including all environmental conditions yielded more robust results that provide more reliable predictions.

Satellite Simulations
The simulation of the Sentinel-2A satellite resulted in a similar model robustness compared to the full spectrum (i.e., 350-2500 nm) model with an R 2 of 0.57 and RMSECV of 163 in the calibration dataset and an R 2 of 0.46 and RMSEP of 192, showing that the band combination of this satellite results in models comparable to high-resolution hyperspectral measurements, although the lack of spectral resolution yields slightly less accuracy in the predictions (Table 2, Figure 5).
The model covering the full spectral range of EnMap (i.e., 420-2450 nm) showed slightly better validation parameters than the Sentinel-2A model, with an R 2 of 0.39 and RMSECV of 198 g/m 2 in the calibration dataset and an R 2 of 0.54 and RMSEP of 178 g/m 2 in the validation dataset. When limited to the 420-900 nm spectral region, the EnMap model showed a better fit with an R 2 of 0.63 and RMSECV of 155 g/m 2 in the calibration dataset and an R 2 of 0.62 and RMSEP of 162 g/m 2 in the validation dataset ( Figure 5, Table 2), indicating that limiting the spectral range results in more robust models.

Satellite Simulations
The simulation of the Sentinel-2A satellite resulted in a similar model robustness compared to the full spectrum (i.e., 350-2500 nm) model with an R 2 of 0.57 and RMSECV of 163 in the calibration dataset and an R 2 of 0.46 and RMSEP of 192, showing that the band combination of this satellite results in models comparable to high-resolution hyperspectral measurements, although the lack of spectral resolution yields slightly less accuracy in the predictions (Table 2, Figure 5).
The model covering the full spectral range of EnMap (i.e., 420-2450 nm) showed slightly better validation parameters than the Sentinel-2A model, with an R 2 of 0.39 and RMSECV of 198 g/m 2 in the calibration dataset and an R 2 of 0.54 and RMSEP of 178 g/m 2 in the validation dataset. When limited to the 420-900 nm spectral region, the EnMap model showed a better fit with an R 2 of 0.63 and RMSECV of 155 g/m 2 in the calibration dataset and an R 2 of 0.62 and RMSEP of 162 g/m 2 in the validation dataset ( Figure 5, Table 2), indicating that limiting the spectral range results in more robust models. Remote Sens. 2018, 10, x FOR PEER REVIEW 5 of 16  Table 2. Modeling parameters from the partial least squares (PLS) models developed for the full spectral range (350-2500 nm) and the limited range (350-900 nm). SNV stands for the standard normal variate treatment, k represents the number of latent variables used, m represents the differentiation order, p represents the polynomial order and w represents the window size for the smoothing function. RMSECV represents root mean squared error of the cross-validation and RMSEP represents root mean squared error of the prediction. * Note that the validation dataset for the "good weather" model is the "bad weather" data, and the validation for the "bad weather" model is the "good weather" data.

Discussion
The fieldwork and analyses in our study focused on overcoming challenges due to environmental conditions at high latitudes. Hyperspectral measurements resulted in robust models (Table 2), using both the full spectral range (350-2500 nm) and the limited spectral range (350-900 nm) (Figure 4). Using PLS on hyperspectral data results in reliable models [45] and ensures higher quality estimates of yield. Our results support the fact that optimal weather conditions to register hyperspectral data are dry and sunny days [46], as seen in the "good weather" models, and that challenging weather conditions result in less robust models (see the "bad weather" models) with lower predictive ability and high error rates, as well as non-linear responses, in estimates of the yield measurements. Our results show worse performance (R 2~0 .7) than previous work done in fields in lower latitudes, where remotely sensed crop estimation was highly accurate (R 2~0 .8) for corn and soybean [47] or wheat [48] under optimal weather conditions. Points with lowered production seem to either consist of several grass species in combination with weeds or with nuisance species like R. repens (Table 1), and these show increased error rates and reduced performance. Measures like weed detection and discrimination in combination with different pre-treatment procedures might increase the performance of PLS models [49]. Additionally, points with very low production (below approximately 50g/m 2 ) show high error rates. This could be due to bare soil affecting the spectra resulting from the early phenological stages of the crops: these fields should be included into the second step of our workflow ( Figure 2) and assessed in a more detailed level to assess if the field as a whole is in need of management, or if there are only small spots that need to be treated.
Our results show comparable results for PLS-models in Barmeier et al. [45] (R 2 = 0.71-0.95 for grain yield) and Fu et al. [13] (R 2~0 .8 for winter wheat in the range 0-1300 g/m 2 ) for winter wheat yield estimates at a 40 • N latitude. Other studies show a high variability on the performance of hyperspectral models to measure above-ground biomass in agricultural fields under different fertilizing regimes [50], with R 2 ranging from 0.7 to 0.9, in the range 74 to 592 g/m 2 . Studies reporting very high model performance (R 2 > 0.9), such as Biewer et al [12], are based on more limited ranges (dry weight range between 0 and 220 g/m 2 ). Our study, on the other hand, represents a higher range of yield (0-1212 g/m 2 ) most likely due to the small areas (0.25m 2 ) used as reference samples and the varying phenological stages of the grasses scanned for the model development (i.e., the sampling was performed through the whole growing season, including the early phenological stages). The inclusion of moisture-affected spectra into the model along with a majority of spectra captured under optimal weather conditions increases the resilience of the model against new variation: new plots will most likely be included in the yield range of our model, and noise caused by e.g., coastal spray or air moisture will not affect the model performance heavily. The apparent model performance may be reduced when including the moisture-affected spectra, but the model will be more robust against future variability given by weather variability. Yet, such resilience does not override the weather constraints for hyperspectral sampling: we recommend all measurements be performed under the best weather conditions possible at the sampling time: we recommend sampling on a clear day and after at least 24h without precipitation.
When cloud-free images are available, satellite images can provide field-level yield measurements that can be combined with in situ spectrometry with a hand-held spectrometer. The Sentinel-2A simulation showed that a multispectral satellite approach may give an acceptable model performance that can help detect fields with abnormal production (i.e., too low productivity), but the predictive ability is limited (R 2 validation =0.46, RMSEP=192). This results in biomass estimations that are mostly limited to coarse ordinal or qualitative assessments of productivity (i.e., low, mid or high productivity), rather than biomass measurements, in our studied fields. The EnMap simulation performed similarly to the Sentinel-2A simulation (R 2 > 0.39, RMSE < 198 g/m 2 ) in the full spectrum models, and both performed worse than the FieldSpec dataset (R 2 > 0.56, RMSE < 169 g/m 2 ). On the other hand, the 420-900nm range EnMap model (R 2 > 0.62, RMSE < 162 g/m 2 ) had a performance closer to the FieldSpec model (R 2 > 0.69, RMSE < 144 g/m 2 ), indicating a high potential on using that spectral range for grassland yield estimations. Unfortunately, this comparison is not possible with the Sentinel-2A satellite due to its limited number of bands. Such consistency in the 420-900 nm region opens avenues to develop models in that spectral range with a minimal loss in yield predictability. Satellite image processing techniques often rely on standardized toolboxes to convert digital numbers (DN) to reflectance which account for sun angle, atmospheric corrections and other preprocessing steps [20], which are necessary to successfully develop satellite-based predictive yield models. Using satellite images to predict total grass biomass production based leaf area index (LAI) showed a variable fit, with R 2 values between 0.4 and 0.55 when using spectra from different dates [32]. The addition of the ENMAP hyperspectral satellite will provide with new information at a high spectral but only a medium spatial (30 m) resolution [24], which can be combined with the presently available satellites (e.g., Sentinel-2, Landsat 8) to improve the monitoring of grasslands in high latitudes. Simulated Sentinel-2 images (10 m) fused (pan-sharpened) with simulated EnMAP data showed encouraging results (R 2 = 0.68) concerning wheat leaf area index [51,52]. Airborne AISA Eagle images fused with simulated EnMap images [51] showed even better results (R 2 = 0.75) giving the opportunity to pan-sharpen future EnMap with ordinary high spatial resolution UAV-imagery for assessment and prediction of yield using both the full spectrum (420-2500 nm) and the limited spectrum (420-900 nm). Although the satellite-based predictions in the full spectral range are mostly limited to qualitative assessments, they can be used as a general estimate on the state of the fields, especially during periods with high cloud cover: missing one or two EnMap satellite passes (with a re-visit time of 27 days at the Equator, or better at high latitudes) due to cloud cover results in long temporal gaps that may happen in critical periods (e.g., fertilization). However, an off-nadir pointing capability of up to 30 • enables a revisit time of better than four days at high latitudes [53], but this has to be programmed in advance.
In cases without available hyperspectral satellite data, it is preferable to have a qualitative biomass production estimate based on Sentinel-2 data (or other available multispectral satellites, if models are available) than having no estimate at all: general information on the status of a field is preferable to avoid unnecessary management actions e.g., fertilizing when unnecessary, which results in economic losses and ecological effects [54].
These promising results show that the two-step workflow we suggest here could improve the efficiency in managing agricultural lands in high latitudes by identifying fields with reduced fodder productivity. This can allow stakeholders to allocate their efforts to improving the productivity in fields that need special management and implement actions to increase productivity [36], or to prepare in advance to order grass from other regions or countries. Once the low-yielding fields have been identified, mapping the field with a portable spectrometer or a UAV-mounted hyperspectral system, and then applying the needed actions (e.g., increased fertilizer) to specific parts of the field, economic and environmental costs related to e.g., unnecessary fertilization [55,56] can be greatly reduced. Other actions may involve removal of weeds [57,58] or improvement of soil properties [59]: these alternatives require a careful assessment of costs and benefits derived from such invasive methods.

Conclusions
In conclusion, our study shows that models resulting from hyperspectral measurements can estimate grass yields with reasonable accuracy in high latitudes, though challenges are greater than in lower latitudes. The spectral region of 350-900 nm showed to be more robust than the full spectrum (350-2500 nm) against influencing factors such as moisture, facilitating the development of reliable models to measure yield of agricultural fields in high latitudes. The satellite measurements are expected to can provide information about fields with abnormal production and facilitate management actions, guided by in situ measurements taken by portable spectrometers.
Supplementary Materials: The following are available online at http://www.mdpi.com/2072-4292/11/4/400/s1, Figure S1: Calibration (left column) and validation (right column) plots of: first row, the model using the "good weather" (i.e., data captured under optimal environmental conditions) data, validated on the "bad weather" (i.e., data captured under challenging weather conditions) data and second row, the model using the "bad weather" data, validated on the "good weather" data using the full spectral range of the FieldSpec 2 (350-2500 nm). Figure  S2: Calibration (left column) and validation (right column) plots of: first row, the model using the "good weather" (i.e., data captured under optimal environmental conditions) data, validated on the "bad weather" (i.e., data captured under challenging weather conditions) data and second row, the model using the "bad weather" data, validated on the "good weather" data using the spectral range limited to the 350-900 nm region.