High Resolution Forecasting of Heat Waves impacts on Leaf Area Index by Multiscale Multitemporal Deep Learning

by   Andrea Gobbi, et al.
Fondazione Bruno Kessler

Climate change impacts could cause progressive decrease of crop quality and yield, up to harvest failures. In particular, heat waves and other climate extremes can lead to localized food shortages and even threaten food security of communities worldwide. In this study, we apply a deep learning architecture for high resolution forecasting (300 m, 10 days) of the Leaf Area Index (LAI), whose dynamics has been widely used to model the growth phase of crops and impact of heat waves. LAI models can be computed at 0.1 degree spatial resolution with an auto regressive component adjusted with weather conditions, validated with remote sensing measurements. However model actionability is poor in regions of varying terrain morphology at this scale (about 8 km at the Alps latitude). Our deep learning model aims instead at forecasting LAI by training multiscale multitemporal (MSMT) data from the Copernicus Global Land Service (CGLS) project for all Europe at 300m resolution and medium-resolution historical weather data. Further, the deep learning model inputs integrate high-resolution land surface features, known to improve forecasts of agricultural productivity. The historical weather data are then replaced with forecast values to predict LAI values at 10 day horizon on Europe. We propose the MSMT model to develop a high resolution crop-specific warning system for mitigating damage due to heat waves and other extreme events.



There are no comments yet.


page 2

page 6

page 7


EarthNet2021: A novel large-scale dataset and challenge for forecasting localized climate impacts

Climate change is global, yet its concrete impacts can strongly vary bet...

TRU-NET: A Deep Learning Approach to High Resolution Prediction of Rainfall

Climate models (CM) are used to evaluate the impact of climate change on...

Heat Waves – a hot topic in climate change research

Research on heat waves (periods of excessively hot weather, which may be...

EarthNet2021: A large-scale dataset and challenge for Earth surface forecasting as a guided video prediction task

Satellite images are snapshots of the Earth surface. We propose to forec...

A Scalable Machine Learning System for Pre-Season Agriculture Yield Forecast

Yield forecast is essential to agriculture stakeholders and can be obtai...

Impact of heat waves and cold spells on cause-specific mortality in the city of Sao Paulo, Brazil

The impact of heat waves and cold spells on mortality has become a major...

Cloud Computation and Google Earth Visualization of Heat/Cold Waves: A Nonanticipative Long-Range Forecasting Case Study

Long-range forecasting of heat/cold waves is a topical issue nowadays. H...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Intro

The growing concern for heat waves impacts on human and animal health is matched by the urgent need of mitigating long and short term damage to crops at all latitudes worldwide. The Leaf Area Index (LAI) has been widely used in modeling the growth phase of crops (Ewert (2004); Heuvelink et al. (2005); Malone et al. (2002); Eik and Hanway (1966)) and for monitoring the impact of the heatwaves on vegetation (Albergel et al. (2019)

). Notably, besides decreasing yield, heat waves impacts are relevant modifiers of phenology phases and then of crop quality, thus requiring to a timely implementation with mitigation procedures, ideally based on forecasts valid at the cultivar scale. Computation of LAI is usually based on an autoregressive component adjusted with weather conditions. The LDAS-Monde platform has recently shown the potential of assimilating enhanced estimates of land surface conditions and remote sensing measurements at medium resolution scales (

) (Jarlan et al. (2007); Albergel et al. (2019)). Unfortunately, such scale is unusable in case of highly variable terrain morphology, e.g. in the Alps and far for practical actionability in general.

The availability of LAI as a product of the Copernicus Global Land Service (CGLS) for all Europe at 300m resolution has opened a novel option for downscaling LAI at more actionable spatial scales. The CGLS PROBA-V’s wide view and polar orbit provide a revisit of every spot on Earth’s land every two days, thus enabling to build a new global composite every 10 days. In this context, we have developed MSMT, a Multi Scale Multi Temporal neural network architecture that estimates the CGLS LAI maps as a novel forecast product. The network is fed with medium-resolution historical weather data integrated with high-resolution Land Data conditions as extracted by a Geographical Analyis System (abbreviated here as GIS features). The MSMT model can compute a daily LAI raster at high-resolution for all the Europe. We then replace the historical weather data with medium-resolution forecast values in order to obtain a high resolution LAI forecast at European scale. Although in an initial phase, this research work opens the potential for developing high resolution predictive maps of LAI at world scale, thus enabling the implementation of warning system in agriculture to mitigate damage due to extreme events, such as heat waves.

2 Data

We used LAI maps from the CGLS PROBA-V (vegetation) mission for all Europe at 300m resolution, corresponding to about 200Mb every 10 days. Weather data (4 measures per day of temperature, pressure and precipitation) are from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5-Land database at a resolution for five seasons (2014-2018), all in the vegetation periods beginning of March to end of October. Inputs include a set of 12 raster data layers defining GIS features, as described in Tab.1 (Appendix). All the data sets have been aligned and harmonized both in time and space in the bounding box . Daily predictions on pixels for which LAI or weather inputs are missing have been filtered out in this study. In summary, one observation per day with spatial resolution equal to 300m has been produced for 25-50 Mln of valid data points per year, depending on the considered season (about 1.2 TB in total).

Figure 1: Model architecture

3 The MSTM Model

The MSMT model in this study computes for each pixel in Europe ( resolution) a LAI value for each day of the season. For each pixel, the model is initialized with the first LAI observation (around the 10th of March), the GIS features and the trajectories of the weather variables during the season, as sketched in Fig. 1. Inspired to a recursive neural network architecture, the output of the MSMT network at time is part of the input at time . In particular, the output of the network is composed by nine values: the predicted LAI at time and 8 values representing a memoryvector in which the network can buffer information about the time evolution of the LAI and weather data. The algorithm aims to capture some well known mechanisms of vegetation growth such as accumulation of temperatures and precipitations. After seasonal initialization, the network is rerun for each day until the end of October. The result is the daily prediction for the LAI: since CGLS data are available each 10 days, only data in such timestamps are used to compute the loss and back propagate the error. Two embedding layers have been used to convert categorical variables (land use and soil) into numerical variables. After that, all the features (GIS, weather plus the memory features described above) pass through a fully connected neural network producing a the following value of the LAI and a new memory state. The network is then used in inference with the ERA-5 data to obtain forecasts.

4 Experiment and Results

Several configurations of MSMT hyperparameters were tested (not reported here for lack of space) with experiments deployed in a CPU-based Azure VM with 64CPUs with 126Gb of RAM. Three seasons (2014, 2015 and 2016) were used as development sets while seasons 2017-18 were held out as external validation. The best model reaches RMSE = 0.738 (0.756) in season 2017 (2018), comparable with the results obtained by

Albergel et al. (2019). In this model the two embedding layers have output size equal to 8 as the dimension of the memory state. We used a simple fully connected neural network (FCNN) with 3 hidden layer (64-128-64); the output size is 9 (1 LAI + 8 memory states). The model is accurate along most of the season, especially in the second part, even if it is based only on the LAI value at 10th of March and on the weather trajectories (Fig.2). In particular, MSMT correctly manages the heatwaves occurring during July 2018 producing accurate prediction during and after this period of interest.

Figure 2: MSMT predictions for the validation set compared to LAI CGLS data (mean values). Orange dots:CGLS; blue line:MSMT. June 2017 and 2018 are highlighted in green.

Land characterization (e.g. Corine class) seems to be significant for the model error. The distribution of the RMSE over Corine classes are shown in Fig.3 (Appendix). We are experimenting how to control land usage class and other GIS features to improve the results of the model, e.g. by excluding specific terrain patterns or weighting more others. This is relevant to address climate change impacts, in particular to optimize risk models such as the crop-specific LAI estimates. Examples of LAI maps, comparing state of art models and MSMT at 300m res, model error and errors with respect to CGLS are reported in Fig.4 (Appendix). Error maps for South-West Europe are also reported in Fig.5 (Appendix).

5 Conclusions

We have introduce the MSMT deep learning architecture to obtain a novel forecast LAI map at the same high resolution (300 m) of the Copernicus product. This model is applicable at global scale, here demonstrated for whole Europe in combination with land surface conditions (GIS features), whose assimilation is being investigated to improve the model. Notably the MSMT network, coupled with a state of art weather forecasting system can be used to implement warning services in agriculture. In particular, we aim to support strategies to mitigate damage to crops during extreme events such as heatwaves. Future works will include experiments with other core networks and the development of crop specific prediction based on the LAI. For the fine tuning of the model we will explore training strategies dependent of Corine over land use classes and terrain morphology patterns.


This work has been supported by the grant ’Microsoft AI for Earth” and by the ”Partenariato Europeo per l’Innovazione” project ”Clima ed Agricoltura in area montana" .


  • [1] C. Albergel, E. Dutra, B. Bonan, Y. Zheng, S. Munier, G. Balsamo, P. de Rosnay, J. Muñoz-Sabater, and J.-C. Calvet (2019) Monitoring and Forecasting the Impact of the 2018 Summer Heatwave on Vegetation. Remote Sensing 11, pp. 520. Cited by: §1, §4.
  • [2] K. Eik and J. J. Hanway (1966) Leaf Area in Relation to Yield of Corn Grain. Agronomy Journal 58, pp. 16–18. Cited by: §1.
  • [3] F. Ewert (2004) Modelling Plant Responses to Elevated CO2: How Important is Leaf Area Index?. Annals of Botany 93 (6), pp. 619–627. Cited by: §1.
  • [4] E. Heuvelink, M. J. Bakker, A. Elings, R.C. Kaarsemaker, and L.F.M. Marcelis (2005) Effect of leaf area on tomato yield. Acta Horticulture 691, pp. 43–50. Cited by: §1.
  • [5] L. Jarlan, G. Balsamo, Lafont,S., A. Beljaars, J-C. Calvet, and E. Mougin (2007) Analysis of Leaf Area Index in the ECMWF land surface scheme and impact on latent heat and carbon fluxes: Applications to West Africa. Journal of Geophysical Research Atmospheres 113 (D24), pp. D24117. Cited by: §1.
  • [6] S. Malone, Jr. Ames Herbert, and D. L. Holshouser (2002) Relationship Between Leaf Area Index and Yield in Double-Crop and Full-Season Soybean Systems. Journal of Economic Entomology 95 (5), pp. 945–951. Cited by: §1.


Name Res. Source Type Freq.
Corine land use 25m https://land.copernicus.eu/pan-european/corine-land-cover 44 classes static
Soil WRB 250m https://data.isric.org/geonetwork/srv/eng/catalog.search#/metadata/5c301e97-9662-4f77-aa2d-48facd3c9e14 118 classes
DTM 25m https://land.copernicus.eu/imagery-in-situ/eu-dem/eu-dem-v1.1/view float
lat, lon 25m DTM
pcurv*, tcurv*, slope* and aspect* 25m DTM + r.slope.aspect default parameters
relief* 25m DTM + r.relief default parameters
strength* and fisher* 25m DTM + r.roughness.vector windows size = 7
min/max temp., pressure, sum/avg precip. https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-land?tab=overview float 4 per day
LAI CGLS 300m https://land.copernicus.eu/global/products/lai float every 10 days
Table 1: Resolution, source, type and frequency of input features. Part of GIS features has been processed with GRASS 7.6.
Figure 3: RMSE (for season 2017) distribution over 18 selected land use classes. Names and percentage of coverage are reported inside each box.
Figure 4: Analysis of LAI variability in a mountainous area. Polygons are apple fields. Top left: LAI value at medium resolution (ECMFW ) for the 10th August 2018. Top right: LAI at high resolution (CGSL 300m). Bottom left: RMSE error for the MSTM for the season 2018. Bottom right: aspect map (N: 90; E: 0=360) computed with GRASS GIS r.slope.aspect function.
Figure 5: MSMT Error maps (zoom on South-West Europe) for the season 2017. Top RMSE, bottom: normalized L1. Normalized L1 (pixel-wise operation: L1 divided by the mean value of the LAI) captures the variation of the prediction respect of the real values in terms of magnitude respect to absolute values.