Learning to forecast vegetation greenness at ﬁne resolution over Africa with ConvLSTMs Claire Robin1Christian Requena-Mesa1Vitus Benson1Lazaro Alonso1Jeran Poehls1

2025-05-02 0 0 2.33MB 11 页 10玖币

侵权投诉

Learning to forecast vegetation greenness at ﬁne

resolution over Africa with ConvLSTMs

Claire Robin1,Christian Requena-Mesa1,Vitus Benson1,Lazaro Alonso1,Jeran Poehls1,

Nuno Carvalhais1,2, and Markus Reichstein1,2

1Biogeochemical Integration, Max-Planck-Institute for Biogeochemistry, Jena, Germany

2ELLIS Unit Jena, Michael-Stifel-Center Jena for Data-driven and Simulation Science, Jena,

Germany

Corresponding author: {crobin, crequ}@bgc-jena.mpg.de

Abstract

Forecasting the state of vegetation in response to climate and weather events

is a major challenge. Its implementation will prove crucial in predicting crop

yield, forest damage, or more generally the impact on ecosystems services rele-

vant for socio-economic functioning, which if absent can lead to humanitarian

disasters. Vegetation status depends on weather and environmental conditions that

modulate complex ecological processes taking place at several timescales. Inter-

actions between vegetation and different environmental drivers express responses

at instantaneous but also time-lagged effects, often showing an emerging spatial

context at landscape and regional scales. We formulate the land surface forecasting

task as a strongly guided video prediction task where the objective is to forecast

the vegetation developing at very ﬁne resolution using topography and weather

variables to guide the prediction. We use a Convolutional LSTM (ConvLSTM)

architecture to address this task and predict changes in the vegetation state in Africa

using Sentinel-2 satellite NDVI, having ERA5 weather reanalysis, SMAP satellite

measurements, and topography (DEM of SRTMv4.1) as variables to guide the

prediction. Ours results highlight how ConvLSTM models can not only forecast

the seasonal evolution of NDVI at high resolution, but also the differential impacts

of weather anomalies over the baselines. The model is able to predict different

vegetation types, even those with very high NDVI variability during target length,

which is promising to support anticipatory actions in the context of drought-related

disasters 1.

1 Introduction

Climate change is leading to an increase in extreme weather events, affecting both ecosystem and

human livelihoods. Africa is one of the most vulnerable continents to climate change with droughts

in the region expected to increase in severity according to the IPCC [

]. Given current projections of

population growth and impacts from climate change, many more people will be affected by extreme

drought events in the future [

]. Any insight into predicting their occurrences can help prepare

short-term solutions to alleviate potential impacts on local and regional communities [3].

The local scale response to extreme events is often not homogeneous [

]. Abiotic (e.g. soil type,

topography, water bodies) and biotic (e.g. vegetation type, plant rooting depth) factors and the

interactions between them affect how vegetation responds to atmospheric extreme events [

].In

1https://github.com/earthnet2021/earthnet-models-pytorch.git

Accepted in Artiﬁcial Intelligence for Humanitarian Assistance and Disaster Response: workshop at NeurIPS

2022, and in Tackling Climate Change with Machine Learning: workshop at NeurIPS 2022.

arXiv:2210.13648v2 [cs.LG] 30 Nov 2022

Figure 1:

Model prediction frame by frame

. The frames are temporally interpolated in a spatio-

temporal cube

(Left side)

(Top row)

RGB satellite imagery.

(Second row)

NDVI target.

(Third

row)

ConvLSTM predictions.

(Bottom row)

L1-norm of the difference between the target and the

prediction. Location: 13°16’34.8"N 16°07’10.4"W, The Gambia.

addition, weather impacts with temporal lagged responses can have a signiﬁcant inﬂuence on ecosys-

tem responses to weather variability [

–

]. This so-called ’memory effect’, arises from the complex

processes involved in vegetation dynamics leading to non-linear interaction on several time scales

[

] (e.g. the time-lagged effect of precipitation on vegetation due to the water soil recharge [

legacy effect of earlier drought [6, 13] or early starting season [14]).

Predicting the evolution and impacts of vegetation at the very local level is therefore a challenge

to support anticipatory action before a disaster In this paper, we aim to create a model capable

of learning the relationship between vegetation states, local factors and weather conditions, at a

very ﬁne resolution in order to characterise and forecast the impact of weather extremes from an

ecosystem perspective. We use the Normalized Difference Vegetation Index (NDVI) [

] as a proxy

for vegetation health monitoring. Our work can be easily extended to predict other vegetation indices,

depending on the ecological process of interest.

Contributions Our main contributions are summarized as follows:

•A proof-of-concept of forecasting vegetation greenness in Africa at high spatial resolution.

•A spatio-temporal analysis of the prediction, both of the dataset and of individual samples.

Application context

Anticipatory action, such as Forecast-based Financing (FbF) [

], implement

short-time action (e.g. commercial animal destocking or early procurement of food) during the period

of time between a warning and a disaster to reduce both the impact of the disasters and the ﬁnancial

cost of humanitarian aid [

]. One of the main obstacles to anticipatory action is the uncertainty of a

disaster actually taking place, what is its magnitude and where and when exactly respond [

] [

This context leads to protracted debates about response strategy and a reluctance to make decisions

on the part of donors and policymakers during these precious time windows [16].

Since drought vulnerability is very context speciﬁc and location speciﬁc [

], a forecast vegetation

evolution tool predicting at the very local scale from an ecosystem perspective can be impactful

for the drought vulnerability assessments in term of location and magnitude to support anticipatory

action. Additionally, our model is relevant for area not well covered by in situ meteorological and

vegetation monitoring instruments (which is notably the case in Africa [

]) since we use only on

satellite data.

Related work

At low spatial resolutions, machine learning models have been proposed for both

vegetation modeling [

–

] and crop forecasting [

–

]. Requena-Mesa et al.

[31]

introduced

Earth surface forecasting as modeling the future spectral reﬂectance of the Earth surface and provided

the ﬁrst dataset in this respect, EarthNet2021. Concurrent to our work, Diaconu et al.

[32]

and Kladny

et al.

[33]

have built ConvLSTM variants for EarthNet2021. Both study the inﬂuence of weather on

the predictions by providing artiﬁcial meteorological inputs.

2 Method

This paper follows the works of Requena-Mesa et al.

[31]

, who introduced EarthNet2021, focusing

instead on Africa and a speciﬁc deep learning model: the Convolutional LSTM (ConvLSTM).

Dataset

Similar to EarthNet2021 [

], our dataset contains Sentinel 2 satellite imagery (bands

red, blue, green and near-infrared), weather variables from ERA5 reanalysis (evaporation, surface

pressure, surface net solar radiation, 2m temperature, total precipitation, potential evaporation) [34]

and SMAP (latent and sensible heat ﬂux, rootzone and surface soil moisture, surface pressure) [

]

and topography from SRTMv4.1 DEM [

]. The data is collected at high spatial resolution for over

40000

locations in Africa, each resulting sample we call a minicube. For more information, see

supplementary materials 5.1 and 5.2.

Task

We deﬁne an Earth surface forecasting task [

] as a strongly guided video prediction task.

The objective is to forecast a length-k sequence of future NDVI satellite imagery (

[

Sn+1, ...,

[

Sn+k

)

based on previous length-n sequence of satellite imagery (

S1, ..., Sn

), topography

and guiding envi-

ronmental variables during both the context and prediction periods (

E1, ..., En, ..., En+k

). Formally,

the task is to learn a function fsuch that:

(

[

Sn+1, ...,

[

Sn+k) = f(S1, .., Sn, E1, ..., En, ..., En+k, T )

In this paper, deep learning models use a context period of one year and then forecast the next three

months of NDVI at a 10-daily timestepping.

Model

We use a ConvLSTM [

], that is an LSTM [

] using convolution to operate in the spatial

domain. We stack two ConvLSTM units (each designed as in Patraucean et al.

[39]

) into an encoder

and another two into a decoder (for forecasting). The encoder is used during the context period,

it gets as inputs the concatenated satellite imagery, environmental variables and topography. The

decoder uses only the environmental variables and the topography as inputs, but gets the hidden

states from the encoder to propagate information from the context period. The model is visualized in

supplementary material ﬁg. 3. The training procedure is detailed in supplementary material 5.4.

Baselines and ablation

We compare our model against a constant baseline (last valid pixel from

context period) and a previous season baseline (observations from previous year). Furthermore we

ablate our model by removing the environmental variables (ConvLSTM w/o weather), i.e. doing

an ablation without weather to see how much the model learns from just the ﬁrst order process

underlying vegetation dynamics: the seasonal cycle.

Evaluation

We evaluate the models using two mean squared error (

MSE

) derived scores: the

root mean squared error (

RMSE

) and the Nash-Sutcliffe model efﬁciency (

NSE

) [

]. The latter

rescales the MSE with the variance of the observations σ2

0, it is deﬁned as:

NSE = 1 −M SE/σ2

0,(1)

Both the MSE and the NSE can be decomposed into parts:

MSE =phase_err +var_err +bias_sq (2)

NSE = 2αr −α2−β2,(3)

Here we introduced the bias

bias_sq = (µ0−ˆµ)2

, the variance error

var_err = (σ0−ˆσ)2

, the

phase error

phase_err = (1 −r)∗2∗σ0∗ˆσ

, the correlation

, the rescaled bias

β= (ˆµ−µ0)/σ0

and a measure of relative variability

α= ˆσ/σ0

represents means,

σ2

represents variances,

·0

refers

to observations and

stands for predicted quantities. For the

NSE

decomposition, the components’

ideal values are r= 1,α= 1 and β= 0 [41].

3 Results

For our evaluation, we removed minicubes with more than

75%

missing pixels. The evaluation is

computed per pixel of the time series: we compute the metrics for each unmasked pixel of each

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

LearningtoforecastvegetationgreennessatneresolutionoverAfricawithConvLSTMsClaireRobin1,ChristianRequena-Mesa1,VitusBenson1,LazaroAlonso1,JeranPoehls1,NunoCarvalhais1,2,andMarkusReichstein1,21BiogeochemicalIntegration,Max-Planck-InstituteforBiogeochemistry,Jena,Germany2ELLISUnitJena,Michael-Stifel-C...

展开>> 收起<<

Learning to forecast vegetation greenness at ﬁne resolution over Africa with ConvLSTMs Claire Robin1Christian Requena-Mesa1Vitus Benson1Lazaro Alonso1Jeran Poehls1.pdf

共11页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Learning to forecast vegetation greenness at ﬁne resolution over Africa with ConvLSTMs Claire Robin1Christian Requena-Mesa1Vitus Benson1Lazaro Alonso1Jeran Poehls1

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: