Statistical impact models in agriculture: monitoring, seasonal, long-term and extreme forecastings
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Statistical impact models in agriculture:
monitoring, seasonal, long-term and extreme forecastings
J. MATHIEU and F. AIRES
LERMA/IPSL at Paris Observatory
December 7th 2018
Weather-sensitivity and impact models
Databases and methodology
Forecasting and estimation of the yield
The low yield extreme case
Conclusions and perspectives
2 / 34
Weather-sensitivity and impact models 2 / 34
• Estimate the weather impact on crop yield • Can we explain part of the crop yield variability based on indirect weather information? No satellite data used (will be considered in the futur, depending on application) 3 / 34
Weather-based impact models Model that represents the effects of the weather (observations, analysis, forecasts) on some socio-economic activity. 4 / 34
Weather-based impact models
Model that represents the effects of the weather (observations, analysis,
forecasts) on some socio-economic activity.
1. Physical, agronomic, dynamic, process-based:
◦ Advantages: description of the processes, can extrapolate to another
climate...
◦ Drawbacks: complex, auxiliary data necessary (e.g. soil type,
properties), numerous parameters
2. Statistical models:
◦ Advantages: data-driven, simplicity, flexibility, low development cost
◦ Drawbacks: less informative in terms of causal relationships
4 / 34
Weather-based impact models
Model that represents the effects of the weather (observations, analysis,
forecasts) on some socio-economic activity.
1. Physical, agronomic, dynamic, process-based:
◦ Advantages: description of the processes, can extrapolate to another
climate...
◦ Drawbacks: complex, auxiliary data necessary (e.g. soil type,
properties), numerous parameters
2. Statistical models:
◦ Advantages: data-driven, simplicity, flexibility, low development cost
◦ Drawbacks: less informative in terms of causal relationships
−→ Only a part of the crop yield variability can be explained by the
weather information (agriculture practice, deseases, irrigation are not
taken into account!!!)
4 / 34
Why forecasting crop yield?
• Monitoring: estimation at the end of the year
• Seasonal forecasting:
• Long-term forecasts:
Début de la
saison de croissance Récolte
Années
2060
Why forecasting crop yield?
• Monitoring: estimation at the end of the year
- Analyse weather sensitivities
- Monitoring at the global scale with no delays, yield insurance
- Less demanding than field measurements sampling strategies or surveys
• Seasonal forecasting:
• Long-term forecasts:
Début de la
saison de croissance Récolte
Années
2060
Estimations en
fin d’année
Why forecasting crop yield?
• Monitoring: estimation at the end of the year
- Analyse weather sensitivities
- Monitoring at the global scale with no delays, yield insurance
- Less demanding than field measurements sampling strategies or surveys
• Seasonal forecasting:
- Help field management at the shorter time scales
- Better water management
- Better stock management
• Long-term forecasts:
Début de la
saison de croissance Récolte
Années
2060
Prévisions saisonnières
Estimations en
fin d’annéeWhy forecasting crop yield?
• Monitoring: estimation at the end of the year
- Analyse weather sensitivities
- Monitoring at the global scale with no delays, yield insurance
- Less demanding than field measurements sampling strategies or surveys
• Seasonal forecasting:
- Help field management at the shorter time scales
- Better water management
- Better stock management
• Long-term forecasts:
- Anticipate the next 50 year’s evolution
- Choice of the crop and agriculture practices
- Investment on long-term infrastructures
Début de la
saison de croissance Récolte
Années
2060
Prévisions saisonnières
Estimations en Prévisions
fin d’année à long termeWhy forecasting crop yield?
• Monitoring: estimation at the end of the year
- Analyse weather sensitivities
- Monitoring at the global scale with no delays, yield insurance
- Less demanding than field measurements sampling strategies or surveys
• Seasonal forecasting:
- Help field management at the shorter time scales The application
- Better water management drives the choice
- Better stock management of predictors.
• Long-term forecasts:
- Anticipate the next 50 year’s evolution
- Choice of the crop and agriculture practices
- Investment on long-term infrastructures
Début de la
saison de croissance Récolte
Années
2060
Prévisions saisonnières
Estimations en Prévisions
fin d’année à long termeChallenges • Difficulties related to agriculture data: not enough data, spatially correlated −→ complex statistical learning • Stationarity of data not warranted • Do not include the impact of several factors not related to weather (complex or unknown information) Two objectives: • Obtain yield forecasts as precise as possible, obtain weather sensitivity • BUT assess realistic model quality!!! (careful with over-training) 6 / 34
Challenges • Difficulties related to agriculture data: not enough data, spatially correlated −→ complex statistical learning • Stationarity of data not warranted • Do not include the impact of several factors not related to weather (complex or unknown information) Complexifying the models might not be useful because the limitation comes from the data, not the algorithm Two objectives: • Obtain yield forecasts as precise as possible, obtain weather sensitivity • BUT assess realistic model quality!!! (careful with over-training) 6 / 34
Databases and methodology 6 / 34
La base de données de l’USDA (1920-2013)
USDA database (1920-2013)
200 200
Alabama Texas
1975
Corn yield (bushel/acre)
Corn yield (bushel/acre)
150 150
• the corn 100
100
• at the county scale
50 50
0 0
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010
Years Years
200 200
200 200
Minnesota Iowa
Rendement de maïs par cantons,150aux États-Unis
(bushel/acre) Alabama Texas
Corn yield (bushel/acre)
yield(bushel/acre)
150
Corn yield (bushel/acre)
150 150
100 100
100 100
Cornyield
9 / 47
Corn
50 50
50 50
0 00 0
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 1920
1920 1930
1930 1940
1940 1950
1950 1960
1960 1970
1970 1980
1980 1990
1990 2000
2000 2010
2010 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010
Years Years
Years Years
200 200
Minnesota Iowa
Corn yield (bushel/acre)
Corn yield (bushel/acre)
150 150
100 100
7 / 34
50 50La base de données de l’USDA (1920-2013)
USDA database (1920-2013)
200 200
Alabama Texas
1990
Corn yield (bushel/acre)
Corn yield (bushel/acre)
150 150
• the corn 100
100
• at the county scale
50 50
0 0
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010
Years Years
200 200
200 200
Minnesota Iowa
Rendement de maïs par cantons,150aux États-Unis
(bushel/acre) Alabama Texas
Corn yield (bushel/acre)
yield(bushel/acre)
150
Corn yield (bushel/acre)
150 150
100 100
100 100
Cornyield
9 / 47
Corn
50 50
50 50
0 00 0
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 1920
1920 1930
1930 1940
1940 1950
1950 1960
1960 1970
1970 1980
1980 1990
1990 2000
2000 2010
2010 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010
Years Years
Years Years
200 200
Minnesota Iowa
Corn yield (bushel/acre)
Corn yield (bushel/acre)
150 150
100 100
7 / 34
50 50La base de données de l’USDA (1920-2013)
USDA database (1920-2013)
200 200
Alabama Texas
2013
Corn yield (bushel/acre)
Corn yield (bushel/acre)
150 150
• the corn 100
100
• at the county scale
50 50
0 0
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010
Years Years
200 200
200 200
Minnesota Iowa
Rendement de maïs par cantons,150aux États-Unis
(bushel/acre) Alabama Texas
Corn yield (bushel/acre)
yield(bushel/acre)
150
Corn yield (bushel/acre)
150 150
100 100
100 100
Cornyield
9 / 47
Corn
50 50
50 50
0 00 0
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 1920
1920 1930
1930 1940
1940 1950
1950 1960
1960 1970
1970 1980
1980 1990
1990 2000
2000 2010
2010 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010
Years Years
Years Years
200 200
Minnesota Iowa
Corn yield (bushel/acre)
Corn yield (bushel/acre)
150 150
100 100
7 / 34
50 50La base de données de l’USDA (1920-2013)
USDA database (1920-2013)
200 200
Alabama Texas
2013
Corn yield (bushel/acre)
Corn yield (bushel/acre)
150 150
• the corn 100
100
• at the county scale
50 50
0 0
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010
Years Years
200 200
200 200
Minnesota Iowa
Rendement de maïs par cantons,150aux États-Unis
(bushel/acre) Alabama Texas
Corn yield (bushel/acre)
yield(bushel/acre)
150
Corn yield (bushel/acre)
150 150
100 100
100 100
Cornyield
9 / 47
Corn
50 50
50 50
0 00 0
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 1920
1920 1930
1930 1940
1940 1950
1950 1960
1960 1970
1970 1980
1980 1990
1990 2000
2000 2010
2010 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010
Years Years
Years Years
200 200
Very important long-term trend, not related to climate change
Minnesota Iowa
Corn yield (bushel/acre)
Corn yield (bushel/acre)
150 150
We focus on anomalies with respect to this trend
100 100
7 / 34
50 50Weather data (1979-2013)
• ERA-interim reanalyses from ECMWF
• Spatial resolution of ∼80km → projected to counties
Mean temperature, May 1979 Precipitation, May 1979
• Monthly data: Air temperature, precipitation, soil moisture,
normalized difference precip-evapotranspiration (SPEI)
Daily data: Tmin and Tmax
8 / 34Some agro-climatic indices 60 Chapitre III - Bases de données
Variables les plus courantes Unités
• Date of the last frozen day in spring
Pj Précipitation du jour j mm
Pmois Précipitations cumulées du mois "mois" mm
T avej Température moyenne du jour j oC
T min +T max
j
T avej = 2
Tbase Seuil de température de croissance oC
Tf rost Seuil de température gélive oC
• Cumul of growth degree days (DJ)
Tmois Température moyenne du mois "mois" oC
T minj Température minimale du jour j oC
T maxj Température maximale du jour j oC
Indices agroclimatiques (nom et formule)
CDF5 Cumul des degrés-froid durant la période d’endurcissement (po- Jour
• Number of days with higher than 30◦ C
tentiel d’endurcissement)
F PP
E−1
CDF 5 = CDF 5j
j=213
0 si j = 212
temperature
CDF 5j =
max{0, CDF 5j−1 + DFj } sinon
F P E = min{j|T minj 6 −10o C et j > 212}
CDj − DJj
DFj =
0 si T avej > 5o C
CDj = o
|T avej − 5| si T avej < 5o C
0 si T avej 6 5 C
DJj =
..
T avej − 5 si T avej > 5o C
DGP Date du dernier gel printanier Jour julien
Chapitre III - Bases de données 61
.
DGPTf rost = max {j|T minj 6 Tf rost }
j=1, ..., 212
FSC
DJ Date de
Cumul des findegré-jours
de la saisond’avril
de croissance
à octobre Jour
Degré-jours
Pmax {j|T M M P 5j > 5.5o C}
F SC 304
=
DJ = DJj
j=91 j=1, ..., 365
FU T M Date
DJ j = de
max fin de(T
{0, cumul
avej − des unités thermiques du maïs
T base)} Jour
DJOHIV Cumul
FU T M des
= min {j|T minjau6cours
degré-jours −2o C} de la période froide (perte d’en- Degré-jours
durcissement)j=213, ..., 365
• Total degree-days higher than 6◦ C during
LSC Longueur de FP
la saison de croissance
hiv
Jour
DJOHIV = max{0, T avej }
LSC = F SC − DSC
Dhiv
LSSG Dhiv = min{j|T
Longueur de la saisonmin j 6 −15 o
C}
sans gel Jour
the growing season
GAjTf6rost
o
FLSSG
hiv = Tmax{j|T
f rost
= Pmin −15 C} Tf rost
− DGP
DJsc
OCA Cumul des degré-jours
Fréquence durantsupérieures
des températures la saison deàcroissance
30o C Degré-
Jour jours
FP SC
P
365
DJsc
OCA=T SU P 30 = DDj T SU P 30j
j=DSC j=1
(
DSC Date de début de1 lasi saison de jcroissance
T max > 30o C Jour
T SU P 30 = o
• Thermal unity of corn
DSC = min {j|T0M M
j
si P 5Tjmax
> 5.56C} 30o C
j
PGA j=1, ..., gel
Date du premier 365automnal Jour julien
DU T M Date de début de cumul des unités thermiques du maïs Jour
P GA Tf rost = min {j|T minj 6 Tf rost }
5j > 12.8o C}
..
DU T M = minj=213,{j|T M...,M365
Psc Cumul des précipitations pendant la saison de croissance mm
FP
SC
.
Psc = Pj
j=DSC
TMA Température minimale annuelle oC
T M A = min {T minj }, j=1, ..., 365
T M M P 5j Moyenne mobile pondérée des températures moyennes quoti-
diennes sur 5 jours
1/16( T avej−4 + 4.T avej−3 + 6.T avej−2 + 4.T avej−1 + T avej )
,→ Linked to the agriculture practice
T M M 5j moyenne mobile sur 5 jours de la température moyenne quoti-
dienne
1/5( T avej−4 + T avej−3 + T avej−2 + T avej−1 + T avej )
,→ Requires daily weather data
UTM Cumul des unités thermiques du maïs CHU
P
FU TM
UTM = U T Mj
j=DU T M
U T Mj = 1/2(Y
maxj + Y minj )
0.33(T maxj − 10) − 0.084(T maxj − 10)2
Y maxj = si T maxj > 30o C
0 si T maxj 6 10o C
1.8(T minj − 4.44) si T minj > 4.44o C
Y minj =
0 si T minj 6 4.44o C
9 / 34
TABLEAU III.1 – Indices agroclimatiques utilisés dans cette thèse par ordre alphabétique.
Plus de détails sont disponibles dans [Côt12]. Les dates habituelles de plantation et de
1810 récolte par état aux États-Unis sont disponibles dans [USDA10].Databases and methodology Databases Yield estimation methodology 9 / 34
Yield estimation methodology
Rendement de maïs
Données météo
Analyse de la tendance
(Modèle mixte non linéaire)
Ajout des
Prédicteurs potentiels Comparaison Anomalie de rendement
meilleurs
(anomalies) (modèle linéaire)
predicteurs
Sélection de modèles
Si insuffisant
Information
N entrées sélectionnées
de groupe
Modèle linéaire / Modèle
Réseau de neurones mixte
anoyield = yieldtrend
−trend
Prévision du rendement
10 / 34Yield estimation methodology
Rendement de maïs
Données météo
Analyse de la tendance
(Modèle mixte non linéaire)
Ajout des
Prédicteurs potentiels Comparaison Anomalie de rendement
meilleurs
(anomalies) (modèle linéaire)
predicteurs
Sélection de modèles
Si insuffisant
Information
N entrées sélectionnées
de groupe
Modèle linéaire / Modèle
Réseau de neurones mixte
anoyield = yieldtrend
−trend
Prévision du rendement
10 / 34Yield trend
• a sigmoïd and a mixed-effect model (parametric)
Alabama county Texas county
200 200
Yield data Yield data
150 150
Corn yield (bu/acre)
Trend from ME logistic regression Trend from ME logistic
Yield (bu/acre)
regression
100 100
50 50
anoyield = yieldtrend
−trend
0 0
1920 1940 1960 1980 2000 2020 1920 1940 1960 1980 2000 2020
Years Years
1 1
0.5 0.5
Corn yield anomalies
Corn yield anomalies
0 0
-0.5 -0.5
-1 -1
1920 1940 1960 1980 2000 2020 1920 1940 1960 1980 2000 2020
Years Years
11 / 34Yield trend
• a sigmoïd and a mixed-effect model (parametric)
Alabama county Texas county
200 200
Yield data Yield data
150 150
Corn yield (bu/acre)
Trend from ME logistic regression Trend from ME logistic
Yield (bu/acre)
regression
100 100
50 50
anoyield = yieldtrend
−trend
0 0
1920 1940 1960 1980 2000 2020 1920 1940 1960 1980 2000 2020
Years Years
1 1
0.5 0.5
Corn yield anomalies
Corn yield anomalies
anoyield = −0.2
0 0
,→ is the variable to estimate
-0.5 -0.5
-1 -1
1920 1940 1960 1980 2000 2020 1920 1940 1960 1980 2000 2020
Years Years
11 / 34Yield estimation methodology
Rendement de maïs
Données météo
Analyse de la tendance
(Modèle mixte non linéaire)
Ajout des
Prédicteurs potentiels Comparaison Anomalie de rendement
meilleurs
(anomalies) (modèle linéaire)
predicteurs
Sélection de modèles
Si insuffisant
Information
N entrées sélectionnées
de groupe
Modèle linéaire / Modèle
Réseau de neurones mixte
Prévision du rendement
12 / 34Weather predictor’s selection
• Many redundant predictors
• Several considered criteria (COR, RMSE, AIC)
• An iterative and multivariate selection approach, avoiding colinearities
0.21
0.46 For the estimation at the year’s end:
0.205
0.44
0.42
Tjuly
0.2
SPEIjuly
RMSE
COR
0.4
0.38 0.195 SPEIjune
0.36
0.19
DJaugust
0.34 DJavril
0.32 0.185
au e
SP Ijuly
S
SP y
A
ril
D st
ay
SM er
oc er
D jun
ul
LG
C
ap
gu
b
D tob
m
Tj
O
to
E
EI
D
c
To
D
D
• Small number of selected predictors
13 / 34Weather predictor’s selection
• Many redundant predictors
• Several considered criteria (COR, RMSE, AIC)
• An iterative and multivariate selection approach, avoiding colinearities
0.21
0.46 For the estimation at the year’s end:
0.205
0.44
0.42
Tjuly
0.2
SPEIjuly
RMSE
COR
0.4
0.38 0.195 SPEIjune
0.36
0.19
DJaugust
0.34 DJavril
0.32 0.185
au e
SP Ijuly
S
SP y
A
ril
D st
ay
SM er
oc er
D jun
ul
LG
C
ap
gu
b
D tob
m
Tj
O
to
E
EI
D
c
To
D
D
• Small number of selected predictors
13 / 34Weather predictor’s selection
• Many redundant predictors
• Several considered criteria (COR, RMSE, AIC)
• An iterative and multivariate selection approach, avoiding colinearities
0.21
0.46 For the estimation at the year’s end:
0.205
0.44
0.42
Tjuly
0.2
SPEIjuly
RMSE
COR
0.4
0.38 0.195 SPEIjune
0.36
0.19
DJaugust
0.34 DJavril
0.32 0.185
au e
SP Ijuly
S
SP y
A
ril
D st
ay
SM er
oc er
D jun
ul
LG
C
ap
gu
b
D tob
m
Tj
O
to
E
EI
D
c
To
D
D
• Small number of selected predictors
13 / 34Yield estimation methodology
Rendement de maïs
Données météo
Analyse de la tendance
(Modèle mixte non linéaire)
Ajout des
Prédicteurs potentiels Comparaison Anomalie de rendement
meilleurs
(anomalies) (modèle linéaire)
predicteurs
Sélection de modèles
Si insuffisant
Information
N entrées sélectionnées
de groupe
Modèle linéaire / Modèle
Réseau de neurones mixte
Prévision du rendement
14 / 34Yield estimation methodology
Rendement de maïs
Données météo
Analyse de la tendance
(Modèle mixte non linéaire)
Ajout des
Prédicteurs potentiels Comparaison Anomalie de rendement
meilleurs
(anomalies) (modèle linéaire)
predicteurs
Sélection de modèles
Si insuffisant
Information
N entrées sélectionnées
de groupe
Modèle linéaire / Modèle
Réseau de neurones mixte
Prévision du rendement
14 / 34The mixed-effect models
The database can be divided into m groups.
Classical linear model
pooling no-pooling
50
45
40
35
30
Y = PT β + ε ∀i ∈ 1, m, Yi = PiT βi + εi
25
-130 -120 -110 -100 -90 -80 -70
• P: predictor’s matrix • ε ∼ N (0, Σ): noise
• β: parameter’s vector −→ fixed effect vector
15 / 34The mixed-effect models
The database can be divided into m groups.
Classical linear model
pooling no-pooling
50
45
40
35
30
Y = PT β + ε ∀i ∈ 1, m, Yi = PiT βi + εi
25
-130 -120 -110 -100 -90 -80 -70
• P: predictor’s matrix • ε ∼ N (0, Σ): noise
• β: parameter’s vector −→ fixed effect vector
Mixed-effect models
Yi = XiT β + ZiT bi + εi ∼ N (XiT β, Σi )
bi ∼ N (0, ∆i ) −→ these are the random
iid
εi ∼ N (0, σ2 Λi )
ind
15 / 34Forecasting and estimation of the yield Corn yield estimation at the end of the year Corn yield seasonal forecast Long-term impact: yield assessment through 2060 16 / 34
Forecasting and estimation of the yield
Corn yield estimation at the end of the year
Corn yield seasonal forecast
Long-term impact: yield assessment through 2060
Début de la
saison de croissance Récolte
Années
2060
Estimations en
fin d’année
16 / 34End-of-year estimation (monitoring) in the US
• Results exploitable on August
• Generalisation results only, on independent years
M
o
n Correlation
i between obser-
t vations and
o estimation of the
r yield anomaly
i
n Corr = 0.53
g
• Good spatial coherency, provides a weather-sensitivity information
17 / 34End-of-year estimation in Virginia
Model
Monitoring mode in a Virginia district for the yield anomaly estima-
tion (left) and yield (right).
18 / 34End-of-year estimation in Virginia
saturation
Model
Monitoring mode in a Virginia district for the yield anomaly estima-
tion (left) and yield (right).
18 / 34End-of-year estimation in Virginia
saturation
Model
CorrUSDA =0.95
Monitoring mode in a Virginia district for the yield anomaly estima-
tion (left) and yield (right).
18 / 34Forecasting and estimation of the yield
Corn yield estimation at the end of the year
Corn yield seasonal forecast
Long-term impact: yield assessment through 2060
Début de la
saison de croissance Récolte
Années
2060
Prévisions saisonnières
19 / 34J
on (W) Seasonal forecast over the US
u
Configuration (I)
n
Configuration (W) e Configuration (I)
• Different predictors depending on the
M
forecasting month
a
y J Correlation map between the observed
u and forecasted yield anomalies, from
May to August.
l
y
J
u
n
e
J A
u u
g
l u
y s
t
20 / 34J
on (W) Seasonal forecast over the US
u
Configuration (I)
n
Configuration (W) e Configuration (I)
• Different predictors depending on the
M
forecasting month
a
y J Correlation map between the observed
u and forecasted yield anomalies, from
May to August.
l
y
J
u
n
e
Corr = 0.30
J A
u u
g
l u
y s
t
Corr = 0.53
Corr = 0.50
20 / 34J
u
on (W) Seasonal forecast over the US
n
Configuration (I)
e
Configuration (W) Configuration (I)
• Different predictors depending on the
M forecasting month
a J
Monthly forecasts, but methodol-
y u
l
y ogy applicable at the weekly level
J
u
n
e
Corr = 0.30
A
J u
u g
l u
s
y t
Corr = 0.53
Corr = 0.50
20 / 34Seasonal forecasts over Virginia
Model
21 / 34Seasonal forecasts over Virginia
Model
Corr 2 = 58%
,→ Forecasts possible starting in July, for weather-sensitive States like
Virginia
21 / 34Two published papers: • Mathieu, J.A. and Aires, F. (2016) Statistical weather impact models: an application of neural network and mixed-effects forn corn production over the United-States, Journal of applied Meteorology and Climatology, 55 (11) • Mathieu, J.A. and Aires, F. (2018) Impact of agro-climatic indices to improve crop yield forecasting, Agricultural and forest Meteorology, 15 (30) 22 / 34
Forecasting and estimation of the yield
Corn yield estimation at the end of the year
Corn yield seasonal forecast
Long-term impact: yield assessment through 2060
Début de la
saison de croissance Récolte
Années
2060
Prévisions
à long terme
23 / 34Long-term yield forecasts through 2060
• Impact on agriculture potentially important
• Some important limitations:
◦ Uncertainty on the climate evolution,
◦ Uncertainty on the agriculture practice (we do not extrapolate
long-term trend),
◦ Uncertainty of the joint evolution of the non-weather factors
24 / 34Long-term yield forecasts through 2060 • Impact on agriculture potentially important • Some important limitations: 24 / 34
Long-term yield forecasts through 2060 • Impact on agriculture potentially important • Some important limitations: • Describe the global trend of the yield anomalies and their spatial distribution. • Use of a mixed-effect model with six simple weather inputs (Tmay, Tjune, Tjuly, Taugust, Pjuly and Paugust) easy to obtain from climate models. 24 / 34
contrast, there is a decrease in radiative forcing, for RCP4.5 and RCP2.6, of 0.07 and 0.2 W/m2,
Estimate the future climate: RCP scenarios of the GIEC
1000 5000 500
900
N O concentrations (ppb)
CO2 concentration (ppm)
CH4 concentration (ppb)
4000 400
800
3000 300
700
600 2000 200 RCP2.6
• RCP4.5
500 RCP6
1000 100 • RCP8.5
2
400
300 0 0
2000 2025 2050 2075 2100 2000 2025 2050 2075 2100 2000 2025 2050 2075 2100
Trends in greenhouse gases contraction [Clarke et al. 2010]
Fig. 9 Trends in concentrations of greenhouse gases. Grey area indicates the 98th and 90th percentiles
(light/dark grey) of the recent EMF-22 study (Clarke et al. 2010)
,→ Compare the evolution of the yield anomalies following the RCP
4.5 & RCP 8.5 scenarios
25 / 34contrast, there is a decrease in radiative forcing, for RCP4.5 and RCP2.6, of 0.07 and 0.2 W/m2,
Estimate the future climate: RCP scenarios of the GIEC
1000 5000 500
900
N O concentrations (ppb)
CO2 concentration (ppm)
CH4 concentration (ppb)
4000 400
800
3000 300
700
600 2000 200 RCP2.6
• RCP4.5
500 RCP6
1000 100 • RCP8.5
2
400
300 0 0
2000 2025 2050 2075 2100 2000 2025 2050 2075 2100 2000 2025 2050 2075 2100
Trends in greenhouse gases contraction [Clarke et al. 2010]
Fig. 9 Trends in concentrations of greenhouse gases. Grey area indicates the 98th and 90th percentiles
(light/dark grey) of the recent EMF-22 study (Clarke et al. 2010)
,→ Compare the evolution of the yield anomalies following the RCP
4.5 & RCP 8.5 scenarios
• Global simulations of the climate CMIP 5 of IPSL
• Weather conditions too different after 2060:
- unreliable extrapolation of the long-term trend (agriculture practice)
- temperatures too different compared to record used to calibrate model
25 / 34Yield temporal evolution
Alabama Indiana Missouri
rcp45 rcp45 rcp45
0.25 0.25 0.25
Anomalies de rendement (%)
Anomalies de rendement (%)
Anomalies de rendement (%)
0 0 0
-0.25 -0.25 -0.25
-0.5 -0.5 -0.5
-0.75 -0.75 -0.75
-1 -1 -1
2010 2020 2030 2040 2050 2060 2010 2020 2030 2040 2050 2060 2010 2020 2030 2040 2050 2060
South Carolina South Dakota Virginia
Années Années Années
rcp45 rcp45 rcp45
0.25 0.25 0.25
Anomalies de rendement (%)
Anomalies de rendement (%)
Anomalies de rendement (%)
0 0 0
-0.25 -0.25 -0.25
-0.5 -0.5 -0.5
-0.75 -0.75 -0.75
-1 -1 -1
2010 2020 2030 2040 2050 2060 2010 2020 2030 2040 2050 2060 2010 2020 2030 2040 2050 2060
Années Années Années
26 / 34Yield temporal evolution
Alabama Indiana Missouri
rcp45 rcp45 rcp45
0.25 0.25 0.25
Anomalies de rendement (%)
Anomalies de rendement (%)
Anomalies de rendement (%)
0 0 0
-0.25 -0.25 -0.25
-0.5 -0.5 -0.5
-0.75 -0.75 -0.75
-1 -1 -1
2010 2020 2030 2040 2050 2060 2010 2020 2030 2040 2050 2060 2010 2020 2030 2040 2050 2060
South Carolina South Dakota Virginia
Années Années Années
rcp45 rcp45 rcp45
0.25 0.25 0.25
Anomalies de rendement (%)
Anomalies de rendement (%)
Anomalies de rendement (%)
0 0 0
-0.25 -0.25 -0.25
-0.5 -0.5 -0.5
-0.75 -0.75 -0.75
-1 -1 -1
2010 2020 2030 2040 2050 2060 2010 2020 2030 2040 2050 2060 2010 2020 2030 2040 2050 2060
Années Années Années
• Non-homogeneous impact, close scenarios
26 / 34Spatial distribution of the yield evolution
2020-2030 2030-2040
2040-2050 2050-2060
Mean yield anomaly per decade (RCP 4.5)
27 / 34Spatial distribution of the yield evolution
2020-2030 2030-2040
2040-2050 2050-2060
Mean yield anomaly per decade (RCP 4.5)
,→ stress the need for adaptation strategies
27 / 34The low yield extreme case The objectives Extreme classification results 27 / 34
The objectives • Increased difficulties, classification instead of regression • 5% of the years are extreme (anoyield < −0.45) • We focus on yield extreme probability estimation • Classification by neural network 28 / 34
The objectives
• Increased difficulties, classification instead of regression
• 5% of the years are extreme (anoyield < −0.45)
• We focus on yield extreme probability estimation
• Classification by neural network
• Confusion matrix as quality measure for the neural network
Prédits Taux
Positifs (1) Négatifs (0) TVP TFN
Observés Positifs (1) VP FN TFP TVN
Observés Négatifs (0) FP VN
Extreme=1, Non-extreme=0
28 / 34The objectives
• Increased difficulties, classification instead of regression
• 5% of the years are extreme (anoyield < −0.45)
• We focus on yield extreme probability estimation
• Classification by neural network
• Confusion matrix as quality measure for the neural network
Prédits Taux
Positifs (1) Négatifs (0) TVP TFN
Observés Positifs (1) VP FN TFP TVN
Observés Négatifs (0) FP VN
Extreme=1, Non-extreme=0
Maximisation of TVP
• Maximize TVP without too many false alarms ⇒
s.c. TFPExtreme=1, Non-extreme=0
29/36 Extreme=1, Non-extreme=0
29/36 Extreme=1, Non-extreme=0
29/36 NN output information richer than its threshold decision
The NN output
Alabama Illinois
0.5 0.5
corr = 0.81 corr = 0.85
0 0
-0.5 -0.5 Times series of
the yield anoma-
Yield anomaly
NN output x(-1)
lies and of the
-1 -1
1980 1990 2000 2010 1980 1990 2000 2010
0.5
New Jersey
0.5
South Carolina
NNoutput×(−1) for
5 distant counties.
corr = 0.76 corr = 0.77
0 0
-0.5 -0.5
-1 -1
1980 1990 2000 2010 1980 1990 2000 2010
Tennessee Years
0.5
corr = 0.84
0 TFP = 15%
-0.5 TVP = 71%
-1
1980 1990 2000 2010
Years
30 / 34The NN output
Alabama Illinois
0.5 0.5
corr = 0.81 corr = 0.85
0 0
The NN behaves like
-0.5 -0.5
Yield anomaly
NN output x(-1)
-1
1980 1990 2000 2010
-1
1980 1990 2000 2010 a yield loss index.
New Jersey South Carolina
0.5 0.5
corr = 0.76 corr = 0.77
0 0
-0.5 -0.5
-1 -1
1980 1990 2000 2010 1980 1990 2000 2010
Tennessee Years
0.5
corr = 0.84
0 TFP = 15%
-0.5 TVP = 71%
-1
1980 1990 2000 2010
Years
30 / 34The classification in terms of distributions
The PDF of the yield anomalies for the samples classified as:
• (1) "extremes" with high confidence (yellow),
31 / 34The classification in terms of distributions
The PDF of the yield anomalies for the samples classified as:
• (1) "extremes" with high confidence (yellow),
• (2) "non-extremes" with high confidence (blue),
31 / 34The classification in terms of distributions
The PDF of the yield anomalies for the samples classified as:
• (1) "extremes" with high confidence (yellow),
• (2) "non-extremes" with high confidence (blue),
• (3) ambiguous (red)
31 / 34Conclusion on the extreme case
• Classification is difficult, less studied in the litterature
• Simple classificateur (SPEIjuly, Tjuly, SPEIjune, Taugust),
non-linear
• 71% of yield extremes well classified
• The NN offers a yield loss extreme index
• This model can anticipate low yield extremes in August
• Some moderate negative anomalies are too ambiguous to be classified (false
alarms)
——————–
Mathieu, J.A. and Aires, F. (2018) : Using Neural Network classifier approach for
statistically forecasting extreme corn yield losses in Eastern USA, Earth and Space
Sciences, in press.
32 / 34Conclusions and perspectives 32 / 34
Conclusions • Measuring the weather-sensitivity → which weather variables are related to yield in a quantitative way → where and when weather sensitivity → why: need expert-knowledge 33 / 34
Conclusions • Measuring the weather-sensitivity • Real assessment of the model ability → low number of agriculture data → true independence of learning, testing and generalisation datasets → specific methodologies: Leave-one-out, Monte-Carlo cross-validation → need for metrics!! impose standard practice 33 / 34
Conclusions • Measuring the weather-sensitivity • Real assessment of the model ability • The limitation is the data, not the algorithms: Simple models! → low number of agriculture data, the weather explains only a part of yield variability → classical bias-variance dilemma → complex agro-climatic indices not particularly necessary → avoid over-training: limited number of parameters and of inputs → adapt statistical methodology to avoid over-training 33 / 34
Conclusions
• Measuring the weather-sensitivity
• Real assessment of the model ability
• The limitation is the data, not the algorithms: Simple models!
• The seasonal forecasts and the year-end estimations
→ estimations using mixed-effect models: need to use local specificities
→ the temperature and precipitation information explain 1/3 of the
variability of corn yield in the US
→ for 1 state over 3, > 60%
→ results coherent with recent studies at global scale:
Deepak et al. 2015 (Nature),
Lobell et al. 2007 (ERL)
→ seasonal forecast possible starting in July (e.g. 58% in Virginia)
33 / 34Conclusions • Measuring the weather-sensitivity • Real assessment of the model ability • The limitation is the data, not the algorithms: Simple models! • The seasonal forecasts and the year-end estimations • Long-term impact → information about the yield evolution → localisation of the states that will have biggest impact → yield reduction by 50% for the most sensitive states in 2060 (RCP 4.5) (assuming constant practice!) → Northern States not really impacted (-2% for the RCP 4.5 scenario) 33 / 34
Conclusions • Measuring the weather-sensitivity • Real assessment of the model ability • The limitation is the data, not the algorithms: Simple models! • The seasonal forecasts and the year-end estimations • Long-term impact • Extreme yield case → classification and detection of the extreme years → only 4 predictors used → 71% of the low extremes are well detected → again, the information content of the selected predictors is the limiting factor, not the methodology 33 / 34
Perspectives
Methodological improvements
• Improvement of the yield model
additional inputs: satellites (NDVI, soil moisture, fluorescence)
use of mechanistic model information (e.g. hybrid model)
collaboration with climate modelers (SPEI, ensemble runs,...)
• Improvement of the extreme classification
choice of the extreme threshold depending on location
choice of the decision threshold
application to the next 50 years (climate simulations) to assess frequency trends
Futur studies
• Application to other continents (Europe), other crops
• Finer temporal resolution (from monthly to weekly)
• Development of an online automatic platform
• Links with financial or insurance solutions to mitigate risks, protection agains adverse
conditions
34 / 34You can also read
