Integrating Agronomy and Machine Learning to Analyze Yield Gap Magnitudes and Causes from Field to Global Levels.

Fernando Aramburu Merlos, Research Assistant Professor, Department of Agronomy and Horticulture, University of Nebraska Lincoln Author

05/25/2024 Added

5 Plays

Description

Identifying untapped opportunities for crop production increase is crucial to guide food security interventions. In this presentation, Fernando will show how agronomic knowledge, big data, and machine learning can be integrated to map yield potentials at high resolution and identify agronomic practices that promptly deliver large on-farm yield gains.

Searchable Transcript

Search:

[00:00:00.750]The following presentation
[00:00:02.220]is part of the Agronomy and Horticulture seminar series
[00:00:05.790]at the University of Nebraska Lincoln.
[00:00:09.760]Alright, well, welcome everybody,
[00:00:11.749]thank you for being with us this morning.
[00:00:15.180]Always exciting for these seminars.
[00:00:17.520]So, a few housekeeping items as usual as we get going here.
[00:00:21.840]First, there's plenty of snacks and coffee in the back.
[00:00:25.320]Make sure to get some,
[00:00:26.520]make sure nothing's left by the time we're all done.
[00:00:29.910]Second item is please hold your questions
[00:00:33.060]until the end of the seminar for the Q&A session.
[00:00:36.120]And then third item is when we do get to the Q&A session,
[00:00:40.111]remember to wait for the microphone
[00:00:42.720]before asking your questions
[00:00:44.070]so that our online members have a chance
[00:00:46.140]to participate in the discussion.
[00:00:49.980]So our speaker today
[00:00:51.150]is Dr. Fernando Aramu Murlow here with us.
[00:00:55.620]Fernando is a research assistant professor
[00:00:57.660]here in our Department of Agronomy and Horticulture.
[00:01:00.300]His expertise is in crop modeling,
[00:01:01.950]spatial analysis and crop ecology.
[00:01:04.680]He received his bachelor's degree
[00:01:06.270]from the National University
[00:01:08.490]of Mar de Plata Argentina in 2010,
[00:01:11.670]and his master's degree from the same institute in 2016.
[00:01:15.960]During his master's, Fernando had his first contact
[00:01:20.730]with UNL when he worked with Dr Patricio Grissini
[00:01:25.470]at the very beginning
[00:01:26.490]of the Global Yield Gap Analysis project
[00:01:29.430]to quantify the yield gaps in Argentina.
[00:01:32.437]Then Fernando got a Fulbright scholarship
[00:01:36.990]and went to UC Davis for his PhD
[00:01:39.665]where he used concepts from ecology
[00:01:42.390]and bio-geography to analyze the spatial distributions
[00:01:46.920]of crops and the consequences on agro-ecosystem diversity.
[00:01:52.110]After completing his doctoral studies,
[00:01:54.210]Fernando joined our department in July of 2022.
[00:01:58.620]So here in the last year and a half, really,
[00:02:01.170]coming up on two years.
[00:02:02.640]Yeah.
[00:02:03.473]As a postdoc in Patricio's lab.
[00:02:06.900]Today, Fernando will present the results
[00:02:08.700]from two of his papers he's been working on
[00:02:10.740]during this last year and a half here at UNL.
[00:02:14.340]So it's my great honor to introduce Fernando,
[00:02:16.980]and it's all yours.
[00:02:18.360]All right, thank you.
[00:02:20.130]It's working?
[00:02:20.963]All right.
[00:02:21.930]So, thank you, Aaron, for the wonderful introduction
[00:02:24.413]and the committee for inviting me to the seminar.
[00:02:27.810]It's my pleasure to be here.
[00:02:29.580]And today I'm gonna talk about how we can integrate agronomy
[00:02:33.540]and machine learning to analyze yield gap magnitudes
[00:02:37.140]and causes from field to global levels.
[00:02:41.370]So, an overview of this presentation,
[00:02:45.360]first I'm gonna introduce
[00:02:47.280]the Global Yield Gap Atlas project.
[00:02:49.200]I think most of you might be familiar with it,
[00:02:51.120]but just in case to have everyone on the same page.
[00:02:54.150]And then I'm gonna present the two papers
[00:02:57.270]I have been working in the first year,
[00:02:59.340]year and a half here.
[00:03:01.260]The first one relates GYGA,
[00:03:04.725]the Global Yield Gap Atlas project,
[00:03:05.700]with machine learning to,
[00:03:07.190]the idea is to gain coverage
[00:03:08.970]and spatial resolution and get the better product for GYGA.
[00:03:13.946]And the second paper, we analyzed yield gap causes
[00:03:17.815]in main systems in Sub-Saharan Africa.
[00:03:21.450]So we submitted the first paper to "Natural Food"
[00:03:24.090]in November of 2023 and yesterday I got the first revision.
[00:03:28.320]It was very positive, so I'm happy with it.
[00:03:30.750]I need to work on those revisions
[00:03:32.730]and hopefully it will be published in a month or two.
[00:03:37.830]And for a second paper that was submitted
[00:03:39.600]to "Nature Communications" in July last year,
[00:03:42.714]we are working to get it published soon.
[00:03:45.930]We are in R3, I mean the third revision
[00:03:48.690]and hopefully will be published within the couple,
[00:03:52.650]I don't know, next week or so.
[00:03:54.390]I don't know, let's see. (chuckles)
[00:03:57.600]So, talking about the yield potential,
[00:04:00.330]oh now I have the microphone.
[00:04:01.590]Okay, it's kind of loud.
[00:04:05.070]All right.
[00:04:08.670]So oh, I put this down a little.
[00:04:16.170]Okay.
[00:04:17.880]So, some concepts about yield potential and yield gaps.
[00:04:22.680]Probably you have seen this field many times,
[00:04:25.260]but first let's start defining yield potential.
[00:04:28.350]So the yield potential is the maximum yield
[00:04:30.180]of a locally adapted cultivar
[00:04:32.460]and system mine by solar radiation,
[00:04:34.290]temperature, carbon dioxide consideration.
[00:04:37.416]And if we're talking about rain-fed crops,
[00:04:41.130]then we're also considering the water supply
[00:04:43.590]and the soil type because we're gonna be more interested
[00:04:45.960]in the water limited yield potential
[00:04:48.870]rather than on the yield potential under irrigation.
[00:04:55.680]We know that, I mean, it is impossible for our farmers
[00:04:59.816]to reach the yield potential
[00:05:02.490]because there are many limitations that they have to face.
[00:05:04.890]And also because it won't be economically optimal
[00:05:09.176]to reach this because you need perfect management
[00:05:12.300]of nutrients, pests and everything.
[00:05:15.840]So, we have seen that the reasonable target
[00:05:18.328]for most farmers that have access to inputs
[00:05:22.336]and good markets is to attain 80% of the yield potential.
[00:05:27.277]And we call this difference as exploitable yield gap.
[00:05:30.707]And then the difference between farmer's yield
[00:05:34.770]and this attainable yield will be related
[00:05:39.540]to neutral deficiencies, poor plant materials,
[00:05:41.970]sub-optimal plant management and so on.
[00:05:48.990]And sometime ago, like 10 or 15 years ago,
[00:05:57.300]some researchers here in UNL work together
[00:06:00.390]with people from Wageningen to create
[00:06:04.710]this Global Yield Gap Atlas.
[00:06:06.630]And this project has been carried out
[00:06:10.590]like for the last 15 years,
[00:06:12.570]and right now it has data for more
[00:06:15.060]than 75 countries and counting.
[00:06:19.138]And it accounts for 60 to 90%
[00:06:22.890]of the global production of rice, maize, wheat and soybean.
[00:06:28.410]They use a global, I mean, a local estimates
[00:06:32.870]of yield potential and yield gaps for each country
[00:06:36.210]but with a global protocol and is very-
[00:06:41.919](teacher chatters)
[00:06:46.141]The idea is to partner with local farmers
[00:06:48.600]and local experts in each country
[00:06:51.300]to get very reliable estimates of yield potential
[00:06:54.210]and yield gaps for each of these countries.
[00:06:57.570]And as a Aaron said, a long time ago,
[00:07:00.210]more than 10 years ago,
[00:07:01.710]I worked to add Argentina to the atlas.
[00:07:06.660]It was one of the first countries added to the atlas
[00:07:09.390]and then it was repeated all over the world.
[00:07:16.290]So, why do we care about yield gaps?
[00:07:20.400]Well, the first reason is to understand
[00:07:24.510]what is the core production potential
[00:07:26.700]at local to national spatial scales
[00:07:28.710]with available land and water resources.
[00:07:31.740]Also to guide investment in our cultural research
[00:07:33.810]and development and to monitor the impact
[00:07:36.270]of our time because if we have a good benchmark
[00:07:38.580]of the potential productivity
[00:07:40.080]that we can achieve in a place,
[00:07:42.000]then we can know how our investments help
[00:07:45.030]to fill those yield gaps.
[00:07:48.210]We can also use this framework to identify
[00:07:50.220]on-farm yield constraints and associated management practice
[00:07:54.630]and also to assess scenarios of food security,
[00:07:57.672]irrigation expansion and climate change,
[00:07:59.820]land use, and environmental footprint.
[00:08:01.830]So, this project has multiple purposes
[00:08:05.520]and it has been very productive in terms of
[00:08:11.970]the reach out of these estimates in each of the countries.
[00:08:17.040]I mean, they led to a big number of publications
[00:08:21.240]and also follow up projects trying
[00:08:23.160]to understand what are the cases of yield gaps
[00:08:25.260]and how to sustainably intensify crop production
[00:08:29.280]in different parts of the world.
[00:08:34.560]So, how do we measure yield gaps?
[00:08:36.600]Well, ideally we would like to measure the yield potential
[00:08:41.670]of each crop in each particular place
[00:08:44.430]via a field experiment that has no limitations of nutrients,
[00:08:49.530]pests, diseases and so on.
[00:08:51.990]But doing that in every field everywhere will impossible.
[00:08:56.455]So it is impossible to think
[00:08:58.740]about these experiments at large scale.
[00:09:02.280]So, what we can do alternatively
[00:09:04.320]is to use crop simulation models coupled
[00:09:06.900]with high quality water, soil, and cropping system data
[00:09:10.020]to estimate the potential and capture
[00:09:12.690]its spatial and temporal variability.
[00:09:16.770]And in principle we can do this in two ways.
[00:09:18.690]We can use a bottom-up approach,
[00:09:20.670]as the one used about the global atlas, that we use.
[00:09:24.720]We do estimates of potential for specific sites
[00:09:28.027]and then we have scaled those estimates
[00:09:30.330]or we can use a top-down approach,
[00:09:31.950]that is take some graded products
[00:09:34.320]for all over the world, for my origin of interest,
[00:09:37.020]and apply the same model to every place
[00:09:41.705]without really worrying about the quality
[00:09:44.190]of my data and the quality of the model calibration.
[00:09:48.930]So, this is a summary of the two types
[00:09:51.660]of spatial frameworks to estimate yield gaps.
[00:09:58.440]So, on one hand we have the bottom-up approach
[00:10:00.510]that starts with site selection
[00:10:03.123]in order to get a good representation
[00:10:05.760]of the crop yield distribution of each crop.
[00:10:08.730]And then for each of these sites we gather information
[00:10:10.830]on weather, cropping systems and soil,
[00:10:13.080]trying to get the best available information
[00:10:15.060]for each of these variables.
[00:10:17.280]And we use a locally calibrated crop model
[00:10:19.530]to get estimates of yield potential for each site
[00:10:22.290]and then we upscale that to the whole region
[00:10:26.190]or using some special framework
[00:10:28.650]of climate zone or ecological zone.
[00:10:32.550]And on the other hand,
[00:10:34.350]a top-down approach will be to get grid data.
[00:10:40.173](classroom chatter)
[00:10:53.700]So it has been used in the past.
[00:10:56.708]One is the Global Agricultural Zone developed as FAO
[00:11:00.708]and the other one is AgMIP, for example.
[00:11:04.337]and in this paper here conducted by
[00:11:07.200]who was also working with Patricio Grissini
[00:11:11.040]sometime ago here at UNML, he compare both approaches.
[00:11:14.991]And the conclusion was that this top-down approach
[00:11:19.650]led to potential estimates that lack agronomy rigor
[00:11:24.000]and local relevance.
[00:11:30.240]So, here is an example of some results
[00:11:34.050]from the global GYGA atlas for rain-fed maize in the US.
[00:11:39.776]So we have results at three levels of variation.
[00:11:44.657]First we have results at weather station level
[00:11:47.755]or site level.
[00:11:49.140]Different colors represent different levels
[00:11:51.435]yield potential, the darker or greener the color,
[00:11:54.016]the higher the yield potential.
[00:11:56.352]So as you can see we have many sites all around
[00:11:59.868]the Mid-West and of course the yield potential
[00:12:02.971]is higher, I mean this is for rainfall conditions
[00:12:06.577]and lower to the west.
[00:12:09.780]Then we can upscale that,
[00:12:11.579]so aggregate that at the climate zone level.
[00:12:14.038]So a climate zone solution with similar climatic conditions
[00:12:16.872]in terms of growing rates and temperature seasonality.
[00:12:22.440]And we can expect that because
[00:12:24.755]it has similar climatic conditions,
[00:12:26.130]the rainfall potential and the and inter is gonna be similar
[00:12:31.470]within this climate zone.
[00:12:35.550]So we average climate zone estimate at (indistinct) level
[00:12:41.490]and then we can continue up scaling to the country level
[00:12:44.490]and give an estimate of the yield potential
[00:12:46.737]and for each country in the atlas.
[00:12:50.310]So what are the pros and cons
[00:12:51.143]of the global GYGA spatial framework?
[00:12:53.670]Well, the first main advantage is that
[00:12:56.187]it's really good that prioritizing resources
[00:12:59.361]to get the best new potential estimate in each site.
[00:13:03.675]And these sites can be selected
[00:13:05.548]to represent the distribution.
[00:13:09.495]And the protocol is good enough and also easy to,
[00:13:12.149]I mean, easy to be consistent through time and space.
[00:13:16.220]So the same protocol must apply
[00:13:18.000]all over the world in different countries
[00:13:20.061]with different levels of data quality
[00:13:22.627]and expertise in modeling.
[00:13:27.570]But some of the challenges
[00:13:28.800]that we face when we use this protocol is that
[00:13:31.331]despite we have this tiered approach of
[00:13:33.990]what will be the best available data and so on,
[00:13:37.430]it still has high data requirements.
[00:13:39.870]So there are some places where it will be really hard
[00:13:43.029]to use this bottom-up approach.
[00:13:46.800]And also we don't have an estimate of new potential
[00:13:49.500]for our specific sites being one use that were simulated.
[00:13:53.957]So we have a limited capacity to link the results
[00:13:57.990]of the atlas with other green databases
[00:14:01.093]that are used very often for different kind of analysis.
[00:14:08.220]I mean if we want to link things the results
[00:14:10.953]of the global GYGA atlas with another,
[00:14:14.307]like a more rational or study
[00:14:19.560]in which we use really data,
[00:14:21.968]the results that we pull such at the climate zone level
[00:14:26.670]and with results I mean now like a unique value
[00:14:29.584]for each climate zone that not really follow
[00:14:32.430]the natural environmental gradient
[00:14:33.690]because you can have big cliffs,
[00:14:35.552]I mean, this is a very rough representation
[00:14:37.771]but you can have like a high potential here
[00:14:40.140]and then you may want move a little bit
[00:14:43.770]to the east or south something
[00:14:45.894]and you have a big jump in yield potential.
[00:14:49.360]And also that's not take full advantage
[00:14:50.653]of all the information that is already available
[00:14:53.679]in the atlas because there may be some
[00:14:56.277]climate zone that was not included in the atlas,
[00:14:58.650]but you have information all around that zone
[00:15:01.170]and want to try to make some kind of interpolation
[00:15:05.100]and try to get an estimate from this place.
[00:15:07.110]So with the current protocol in the atlas,
[00:15:10.157]we are not taking full advantages
[00:15:12.090]of all the data that is available.
[00:15:18.240]So, now going to the first paper
[00:15:23.070]that I've been working on in the last year or two years.
[00:15:29.880]So this is an idea that we have with Patricio
[00:15:33.720]that we wanted to use all information in the atlas
[00:15:38.580]to get some global estimates of yield potential,
[00:15:42.060]but first we need to fill some data gaps that we have
[00:15:44.417]because we have data from most of the country
[00:15:46.800]but not for whole country.
[00:15:48.420]And also within each country there was some places
[00:15:52.094]where we still don't have data.
[00:15:54.240]So, we developed this protocol to combine the results
[00:16:00.510]from the atlas with machine learning
[00:16:03.000]to get maps of yield potential at the higher resolution
[00:16:06.390]and for most of the crop lines around the world.
[00:16:10.980]So this framework starts with the estimates
[00:16:13.486]of the global GYGA atlas at the site level.
[00:16:18.360]So it relies on this bottom-up approach
[00:16:21.510]to get very reliable estimates for each of these sites.
[00:16:27.240]So here we're using crop moedling that was specifically
[00:16:31.005]calibrated for each of these countries
[00:16:33.960]and measure weather, soil and crop system data
[00:16:36.300]to generate new potential estimates
[00:16:38.760]that are also locally evaluated
[00:16:41.337]and for the sites that are more important
[00:16:43.170]for crop production in each of these countries.
[00:16:47.160]And here is a map or there's an another map here.
[00:16:52.830]So these are all the sites that are included in the atlas
[00:16:56.232]for maize, wheat and rice by middle of last year.
[00:17:01.817]Now we have more sites also for Mexico
[00:17:05.160]for rainfed and irrigated maize.
[00:17:07.293]And so the atlas is continually updating.
[00:17:11.910]But as you can see we have very good coverage
[00:17:14.914]with a lot of sites for all over the world
[00:17:17.516]for the main most important cropping systems
[00:17:20.610]for each of these crops.
[00:17:24.030]But still there are some countries
[00:17:26.430]that are not included yet and are gonna be very hard
[00:17:29.550]to include using the bottom-up approach, for example,
[00:17:32.940]within Russia is a big one that would be really hard
[00:17:36.180]to evaluate right now.
[00:17:39.540]And there are a few more.
[00:17:43.170]So the idea is to combine these site specific
[00:17:49.500]potential estimates with gridded cropping system,
[00:17:52.320]climate data in a machine learning model.
[00:17:58.710]But then we use that machine learning model
[00:18:01.222]to create estimates of yield potential
[00:18:05.030]or create estimates of yield potential
[00:18:06.613]at the global level to gain spatial coverage
[00:18:09.678]and also to increase resolution.
[00:18:14.460]Of course if we want to that we need the best available
[00:18:18.280]recovery system and come
[00:18:20.056]and sell data that we have
[00:18:22.410]that we can have at the global level.
[00:18:26.010]And so this is the data that we use in this study.
[00:18:31.740]Of course before there is not perfect,
[00:18:33.540]there are large uncertainties, especially when we want
[00:18:39.480]to get crop and system specific data.
[00:18:42.240]So we wanted to gather the best
[00:18:44.900]of all data on crop calendars around the world.
[00:18:48.390]And this is a big data app
[00:18:49.890]that if someone is very interested on creating global
[00:18:52.427]crop calendar data that will be a very useful project.
[00:18:59.102]But we combine it from different sources
[00:19:03.414]and with these core calendars with the right
[00:19:06.330]and we were information, we arrived
[00:19:08.607]by like there was the crop cycle lens
[00:19:11.250]of in each place they accumulate precipitation
[00:19:13.774]within each path of the crop cycle
[00:19:17.070]and also the photosynthetic active radiation.
[00:19:21.780]And then because we know that this data is not perfect
[00:19:24.519]and there are many data labs here,
[00:19:26.190]we also use some bio climatic measures that can be related
[00:19:30.360]to crop performance, like the ones that are used
[00:19:34.920]to create these climate zones
[00:19:37.623]that are used in the global GYGA atlas,
[00:19:39.866]but also some other variables that are related
[00:19:42.600]to the seasonality, temperature
[00:19:44.370]and precipitation in each of these places.
[00:19:48.600]And then also include the soil value to create estimates
[00:19:52.590]of landfill crop potential.
[00:19:55.829]But it is not perfect
[00:19:57.294]but it's a risk available at the global level
[00:19:59.310]but it's an estimate
[00:20:01.280]of the plant available salt water hauling capacity
[00:20:03.807]between zero and one meter
[00:20:05.700]and it in the second meter of F.
[00:20:12.810]Then we know this information where we did is we
[00:20:17.760]train a random forest model
[00:20:20.700]using the yield potential estimates of the atlas
[00:20:26.419]as training data
[00:20:29.400]and all these environmental values as predictors
[00:20:33.450]and then we create this global yield potential map
[00:20:40.306]for the main regions around the world.
[00:20:44.883]And something important here is we ensure that the,
[00:20:50.190]our predictions are within the environmental range
[00:20:54.270]that is explored by the atlas.
[00:20:56.100]So we don't want to extrapolate,
[00:20:57.900]I mean we extrapolate in space
[00:20:59.790]but we don't want to make any extrapolations in
[00:21:01.920]environmental space.
[00:21:03.000]We don't want to make extrapolation on
[00:21:05.160]for very different environmental conditions than the ones
[00:21:09.060]that are already included in the atlas.
[00:21:11.160]The real thing is that because we have so many sites
[00:21:13.410]and so with coverage of the atlas,
[00:21:15.531]we can get still very high coverage
[00:21:17.490]with these grid predictions.
[00:21:22.811]Another tricky part is how
[00:21:24.203]to rate this machine learning model operations
[00:21:26.550]because we have, I mean usually
[00:21:29.068]for machine learning we do cross validation, that means
[00:21:32.313]that we exclude some sites from the trainee data
[00:21:36.139]and then we test the ability of the model to
[00:21:38.883]perform on those excluded sites.
[00:21:43.110]But when you have special data
[00:21:44.963]that is very cluster in some particular issues,
[00:21:48.720]if you wanna exclude one point
[00:21:50.250]or if you exclude random points, then it might happen
[00:21:54.360]that you have a training site
[00:21:55.977]that is very close to that training point.
[00:21:58.560]So it's not like a very valid testing.
[00:22:02.722]So we, we do a spatial cross validation
[00:22:08.820]that, so for example here in this map
[00:22:14.670]we have the prediction grid that means all area
[00:22:19.380]where we want to make estimates and this is for
[00:22:22.247]length and width.
[00:22:23.640]So as you can see a lot of area where we want
[00:22:25.770]to make estimates of (indistinct) potential in Russia
[00:22:29.250]and also here in Turkey and some other places.
[00:22:32.430]But most of our observations are from Europe, Argentina,
[00:22:36.990]United States and Canada and Australia.
[00:22:39.000]And we have a few here in East Africa.
[00:22:43.260]So we want to test the ability of the model,
[00:22:46.337]to predict you here
[00:22:48.607]and also here, but mostly the ability
[00:22:51.900]to predict you in places
[00:22:52.733]where we don't have estimates of potential.
[00:22:56.490]So to do that we use this special cross validation in
[00:22:59.970]which we train the model.
[00:23:02.250]For example with the points in room
[00:23:04.134]we excluded all the points in blue here
[00:23:07.620]and we tested the ability of the model
[00:23:09.652]to predict here in this point in red
[00:23:11.160]and we repeated the process for all the points
[00:23:14.311]by excluding some sites train the model with the points
[00:23:18.893]outside or far from that side and testing on that.
[00:23:29.490]And something else is that we repeated this process also
[00:23:33.027]for the current (indistinct) upscaling protocol
[00:23:38.430]that is used in the atlas.
[00:23:39.810]So in the atlas we have like these
[00:23:41.730]estimates at the climate zone level.
[00:23:43.830]So we do this we scale some sites from the
[00:23:46.740]climate zone average.
[00:23:49.380]We complete the average without that site
[00:23:52.182]and we compare the average at the climate zone with
[00:23:54.450]the value in that particular place.
[00:24:03.450]So here is the comparison of the performance
[00:24:07.820]of the machine learning with this
[00:24:10.230]climate zone upscaling protocol.
[00:24:14.040]So currently the, when we did this for the climate
[00:24:17.963]for the climate zone in the atlas of excluding some size,
[00:24:21.630]we completely rubbish
[00:24:22.463]and comparing the average of the climate zone without
[00:24:25.650]that side with climate zone or the average or
[00:24:28.013]or the, sorry, with the value of that side, we get this kind
[00:24:32.010]of elimination between the yield potential
[00:24:35.053]of the predictive value,
[00:24:39.630]yield potential in the global yield atlas.
[00:24:44.204]So the currencies are around for, this is for wheat
[00:24:47.310]on the bottom and for maze on the top.
[00:24:50.909]And as you can see the maze is around on average 20
[00:24:54.897]or 14%
[00:24:57.350]but it, it is pretty good, it's pretty decent,
[00:24:59.504]decent move points, (indistinct) values
[00:25:01.710]and wind points are data values
[00:25:04.890]and the global crop area coverage that will reach
[00:25:07.175]with the climate zone in the atlas is around 51
[00:25:09.720]to 54% for these two crops.
[00:25:16.200]And our option is to extrapolate the estimates
[00:25:19.980]at the climate zone level to countries
[00:25:21.570]that are not included in the atlas.
[00:25:23.520]That is to say, okay, I'll compare the yield
[00:25:28.914]in this site, in this climate zone with the average yield
[00:25:31.950]of the same climate zone in another country.
[00:25:35.853]And when we do that, the (indistinct) drops to 19% for wheat
[00:25:40.260]and 27% for maze.
[00:25:46.200]But we can gain a lot of coverage
[00:25:47.730]because we're including climate zones
[00:25:49.680]that are already available in one country
[00:25:51.750]but not in our country of interest.
[00:25:55.830]So we increase the college to 85 to 87%
[00:26:01.356]and we, and with this machine learning more,
[00:26:03.629]what we can do is we can keep
[00:26:05.100]or even increase the accuracy that we reach
[00:26:08.730]with the climate zone of scaling protocol.
[00:26:12.332]So we can see here that this cross valuation,
[00:26:16.983]the average error for wheat is 14%
[00:26:21.142]and for maze is 18%.
[00:26:23.606]So it's very similar than for the
[00:26:25.749]current (indistinct) country at current protocol.
[00:26:29.130]But we can gain a lot of coverage,
[00:26:30.570]we can reach 95% of coverage
[00:26:34.301]of global around the world.
[00:26:40.920]And another thing about this machine protocol,
[00:26:45.954]this machine learning tool,
[00:26:47.370]is that we can use the same model to try
[00:26:50.790]to quantify where is the uncertainty of these predictions.
[00:26:56.790]And this uncertainty is based on the relationship
[00:26:59.790]between the errors that we observe in the cost validation
[00:27:03.600]with the similarity
[00:27:04.733]between the training site and the testing sites.
[00:27:07.620]So the more different that we have are the training site
[00:27:11.280]with the testing sites in terms
[00:27:12.750]of environmental conditions, the higher the level.
[00:27:16.560]So when we start relation between the similarity
[00:27:19.534]and error to get an estimate of uncertainty
[00:27:23.520]for every pixel in which we make a
[00:27:26.220]prediction of your potential.
[00:27:29.370]And as you can see in these two maps,
[00:27:33.314]dark blue colors are very low uncertainty,
[00:27:35.970]yellowish colors are high uncertainty.
[00:27:38.760]So we have low uncertainty for
[00:27:40.646]the most important crop production emissions for each
[00:27:45.000]of the crops, like maize, related maize,
[00:27:49.518]and wheat, and the uncertainty is much greater
[00:27:53.220]for places that are maybe less important
[00:27:56.556]or very different from the ones
[00:27:57.389]that are already included in the atlas
[00:28:00.510]and everything of places that we can help us to
[00:28:04.140]define the next target in the Global Yield Gap Atlas,
[00:28:07.140]we can say okay,
[00:28:08.190]there is a lot of uncertainty in Mexico
[00:28:10.320]and it's still a very important area for wheat and maize,
[00:28:13.530]let's focus on that crop in that region.
[00:28:18.210]So it can give us a lot of insight in that sense.
[00:28:25.763]And here we have an example of these
[00:28:30.600]maps of the potential maps for rain fed
[00:28:33.810]and related maze and wheat.
[00:28:37.140]We also created estimates for rice
[00:28:39.570]but I didn't go here to get some real maps.
[00:28:44.872]This information is gonna be available
[00:28:47.217]one day in the atlas I think.
[00:28:50.190]And you see well the with darker
[00:28:53.398]blue colors you have higher potential
[00:28:56.940]red colors lower in potential.
[00:29:02.070]And another way to try to evaluate the accuracy
[00:29:05.718]of this new potential estimates is to check
[00:29:08.742]for negative yield gaps.
[00:29:10.680]So as we say the yield gap is a difference
[00:29:13.320]between the good potential and farmers yields.
[00:29:16.410]So we know that the yield gap has to be positive
[00:29:20.250]because the yield potential should be higher
[00:29:22.880]and in Google vehicle, but usually we be higher than
[00:29:26.190]farmers average use.
[00:29:28.980]So in know that these gaps are clear evidence
[00:29:30.770]of potential and estimation.
[00:29:33.720]So we compare the,
[00:29:37.200]I mean we calculated the year gap for the US Midwest for
[00:29:41.760]that we know that it's a very intensive main system that
[00:29:46.930]we, it has very low yield gaps.
[00:29:51.420]So we need that analysis for our data
[00:29:54.660]for our GYGA machine learning meta model
[00:29:57.708]and for a very popular approach that is
[00:30:02.689]the global (indistinct) science of FAO.
[00:30:06.769]And here in blue colors you have different levels
[00:30:10.140]of heatmaps,
[00:30:10.973]and in red colors you have different levels
[00:30:12.480]of negative build gaps.
[00:30:14.070]So blue positive, red negative.
[00:30:18.127]So in the machine learning model we only have a few counties
[00:30:22.050]where we have a slightly negative big gaps
[00:30:26.206]but it, it can happen but it's less than 1% of area.
[00:30:31.707]And in contrast for this FAO estimates of yield potential,
[00:30:36.973]we can see that in all these western region there is a place
[00:30:41.400]where it is kind of hard to
[00:30:42.870]but real estimates of heat potential
[00:30:44.640]because of length of new potential
[00:30:47.850]because it is subject to water stress
[00:30:51.210]but still have really high yield.
[00:30:54.180]You can see that the model
[00:30:55.997]that they apply is underestimating yield range
[00:30:58.620]of yield potential in all this area here.
[00:31:05.010]And I find comparison about these two
[00:31:08.603]approaches or the three approaches, a top line approach
[00:31:13.830]like the one of FAO, the GYGA climate zones,
[00:31:17.460]that is are currently available in the website
[00:31:20.310]and our machine learning meta model for East Africa
[00:31:24.043]or for you can see
[00:31:25.498]that there's maybe little variability potential in this
[00:31:30.360]across East Africa where (indistinct)
[00:31:32.975]are very valuable, we can tell. (chuckles)
[00:31:35.940]And with the estimates from (indistinct),
[00:31:40.590]you can see that it captures virtually the viability.
[00:31:44.610]There are some places where we don't have any estimates yet.
[00:31:48.763]And for example here in Zambia, you can see that,
[00:31:52.152]I mean you follow environmental grades
[00:31:53.807]but you have like these big jumps from one mission
[00:31:55.710]to another in terms of new potential
[00:31:59.794]and the meta model, we can do, we can expand the college
[00:32:02.580]to reach some countries that are
[00:32:04.590]not yet included in the atlas.
[00:32:07.157]And we have more smooth environmental gradients
[00:32:10.620]in different parts of East Africa.
[00:32:14.670]So what are advantages of combining
[00:32:19.426]GYGA with this machine learning model?
[00:32:21.600]Well just to summarize, we're having talking so far we are
[00:32:26.543]gaining coverage without losing accuracy.
[00:32:30.180]So the coverage of the climate zone level estimates of GYGA
[00:32:35.311]by now is around 55% for these crops
[00:32:38.040]and we can gain up to 95% of global (indistinct) coverage.
[00:32:42.690]And the accuracy is more
[00:32:43.680]or less the same for these two examples.
[00:32:48.600]Also we get high resolution.
[00:32:50.760]So I mean we have now the capacity to pinpoint
[00:32:54.080]to specific sites
[00:32:55.500]and to know the new potential of that particular place
[00:33:01.350]and we can do that with no uncertainty
[00:33:03.912]and we know that uncertainty things gonna be much lower in
[00:33:06.432]places where we already have estimates
[00:33:08.103]of the lower GYGA atlas.
[00:33:10.080]That happens to be the main breadbaskets.
[00:33:12.510]So of course we want to get the new potential for site
[00:33:17.431]that is very far from current sites including GYGA,
[00:33:20.880]then uncertainty is gonna be higher.
[00:33:22.650]But if we want to move this for I know for example
[00:33:25.770]for the US where we have good estimates,
[00:33:27.780]then we can get the higher resolution yield potential
[00:33:30.257]across the US,
[00:33:34.277]and if all messages is that we get the good estimate
[00:33:37.290]that this still has local relevance
[00:33:39.570]and that overcomes the measure limitations
[00:33:41.430]of these top approaches that actually I have shown,
[00:33:44.520]I have shown so before.
[00:33:48.487]Okay so now let's take a look at the
[00:33:53.760]yield gaps of maize, or length of maize
[00:33:56.692]in different parts of the world.
[00:34:00.148]So as suspected, I mean the Europe are very small
[00:34:05.708]in the US, Europe
[00:34:07.857]and East China where agriculture is more,
[00:34:12.030]it's more developed and identified,
[00:34:14.580]we have intermediate gaps in South America
[00:34:17.280]or south-west America and Southeast Asia
[00:34:21.780]and very large yield gaps in sub-Saharan Africa and India.
[00:34:30.360]And with that I will jump to my second paper.
[00:34:34.716]I don't know what to tell you right now
[00:34:36.810]but it's a different topic so we can start from scratch.
[00:34:41.490]Well we want to, well we wanted to analyze
[00:34:44.430]what are the causes of sub-Saharan Africa yield gaps.
[00:34:48.870]So we know that the gaps in this area are huge
[00:34:55.110]but in addition to that if we analyze
[00:34:57.903]what have happened in this region in the last 20 years.
[00:35:03.240]We can see that maize area has doubled
[00:35:08.394]while in the red, the area of (indistinct)
[00:35:12.208]have remained the same.
[00:35:14.013]So there is a lot of increasing demand in this area,
[00:35:18.300]sorry, in mass in this, in this region
[00:35:22.550]because people among the diet from sorghum based
[00:35:28.672]to maize based is not, they are using
[00:35:31.720]maize to feed animals.
[00:35:33.780]They are directly consuming maize
[00:35:35.070]but they are switching crops
[00:35:38.310]and it is expected that
[00:35:40.210]the maize demand will increase 2.1 time
[00:35:43.170]in the next 27 years.
[00:35:45.000]So right now the maize demand is about 80 million tons
[00:35:48.180]and it's gonna be 184 millions by 2050.
[00:35:55.080]And if we also see
[00:35:56.100]what have happened in yield in in this place,
[00:35:59.250]well the union rates are very low,
[00:36:02.280]only about 27 clearance per data per year,
[00:36:06.240]with yields that are currently around two (indistinct).
[00:36:09.600]It was 1.5 20 years ago.
[00:36:12.548]And if we compare that human rate with the union rate in our
[00:36:15.812]tropical places or subtropical places like Southeast Asia
[00:36:20.550]or South America, we can see
[00:36:22.380]that there is a big difference in the union rate
[00:36:24.877]and there is still a lot
[00:36:26.580]that could in Sub-Saharan Africa, (indistinct).
[00:36:30.733]So what are the best perspectives
[00:36:32.640]of Sub-Saharan Africa like 2050?
[00:36:35.340]Well stepping in the current situation we have analysis
[00:36:39.270]of two types of data.
[00:36:41.574]They produce around 80 million tons
[00:36:42.897]and the demands is pretty much the same.
[00:36:45.690]So they are currently self-sufficient for maize.
[00:36:49.470]They dont need to bother with maize.
[00:36:52.046]There is movement between countries in sub-Saharan Africa
[00:36:54.840]but as a reason you'll need to import a lot.
[00:37:00.030]But if we have the same rate of fuel gain
[00:37:03.677]with a special decrease in mass demand,
[00:37:06.328]we'll get another year of 2.7 tons by 2050.
[00:37:11.924]That means to a production assuming
[00:37:13.500]that we have the same area of maize of 108 ton,
[00:37:18.690]million tons and the demand is
[00:37:21.515]between 184 million tons.
[00:37:24.510]So the self sufficiency ratio is 0.6.
[00:37:27.300]This means that they may have to
[00:37:28.620]import 40% of their local produce.
[00:37:32.126]The balance is area 76 million tons.
[00:37:36.011]That is a lot of maize to import
[00:37:37.697]or the option will be to continue increasing maize area.
[00:37:41.610]But that will be at the expense
[00:37:42.690]of probably natural ecosystem so,
[00:37:45.090]or national land that are not very suitable
[00:37:47.010]for mass production.
[00:37:49.162]And assuming that new is the same in this new land
[00:37:53.780]that might not be the case then they, they will need
[00:37:57.730]an extra 28 million acres of maize.
[00:38:04.770]So the real message here
[00:38:05.840]is that sub-Saharan Africa urgently needs
[00:38:09.060]to intensify mass production sustainably.
[00:38:11.898]And we have seen in many important, I mean like
[00:38:16.890]how impact channels where there is a instance that call
[00:38:20.940]for approaches that have very
[00:38:23.440]critical areas on their capacity to increase yield.
[00:38:27.191]When we talk about sub-Saharan Africa,
[00:38:29.400]like you can see purpose about nature
[00:38:31.050]resolutions, the general agriculture
[00:38:33.390]and so on and so forth.
[00:38:35.653]But we're, we're in the focus
[00:38:38.670]towards promoting green basic agronomy in this region
[00:38:41.760]to close these yield gaps.
[00:38:47.940]So we know that Sub-Saharan Africa has large yield gaps,
[00:38:50.903]but we, what we want to do is to try to quantify
[00:38:53.683]what are the main management practices
[00:38:56.490]behind these yield gaps.
[00:38:59.130]So whether it is poor soil nutrition
[00:39:01.686]or bad quality seeds, poor management
[00:39:03.240]and the other factors to give a number
[00:39:05.970]to each of these variables.
[00:39:10.563]And in order to do that what we can do
[00:39:13.083]and we have done, I mean (indistinct) group have done
[00:39:16.991]in other regions of the world
[00:39:19.110]and previous work is to work with farmer yield data trying
[00:39:24.060]to understand the variability behind these yields.
[00:39:29.281]We know that despite having
[00:39:33.330]a big large gap,
[00:39:34.806]there is gonna, there is gonna be variability
[00:39:36.693]among farmers in terms of practices and yield.
[00:39:40.530]So we want to compare the yield of high yield fields
[00:39:43.960]with the yield of low fields while considering
[00:39:48.847]what's the environmental background
[00:39:51.570]of each of these fields.
[00:39:54.960]So our questions here was well
[00:39:58.830]what management practices lead to higher yields?
[00:40:01.620]What is the need of farmers adopting good agronomy?
[00:40:03.810]And what will happen if all farmers
[00:40:05.430]adopt better agronomic practices available right now?
[00:40:11.247]And to answer these questions we partner with people from
[00:40:15.075]one the families and she is working in East Africa
[00:40:18.600]and also a little bit in West Africa
[00:40:21.720]and they have this amazing database of farmer fields
[00:40:25.093]where they went
[00:40:26.670]and measured the yield by themselves doing crop cuts.
[00:40:30.690]So it's not, and they got this yield measurements
[00:40:35.125]and the location of those fields
[00:40:38.880]and they also asked the farmer a set of questions related
[00:40:42.600]to the management practices in terms of neutral inputs
[00:40:45.900]and application methods, planting dates, density,
[00:40:48.930]spacing, which kind of equipment they use,
[00:40:54.420]what kind of waste management practices
[00:40:56.550]and the damage level that they observe in the fields
[00:41:01.078]and where they apply soil amendments.
[00:41:06.240]So we benefit a lot from this large database
[00:41:10.607]and we work with them to clean the data
[00:41:13.890]and pre process the data and it was a lot of work
[00:41:17.415]but we ended up very clean data set of
[00:41:22.668](indistinct) of maize
[00:41:24.360]because they also have data for (indistinct) maize
[00:41:27.420]and poor maize stands, we only use data
[00:41:30.145]for poor maize.
[00:41:35.370]And after working to clean
[00:41:37.500]and pre process data we this analysis
[00:41:41.430]of causes of field gaps and we did this
[00:41:45.977]but first stratifying fields based on climate zones
[00:41:49.530]in order to make fair comparison among fields.
[00:41:51.720]So we know that we're comparing fields at subject
[00:41:54.513]to similar hematic conditions
[00:41:56.310]and then we use advanced statistical methods
[00:41:58.830]to handle the uncertainty of this observational data that
[00:42:03.533]where we don't have a formal experimental design
[00:42:06.351]and we don't have a really good control about all the
[00:42:08.662]environmental variables affecting yields.
[00:42:14.061]So this is an overview of the data,
[00:42:19.059]the different climate source that we have.
[00:42:21.035]We have data from Nigeria, Rwanda, Kenya, Rwanda,
[00:42:23.640](indistinct), Tanzania and Zambia
[00:42:27.900]and we classify the data based on the climate zones.
[00:42:33.240]We have here the different varieties
[00:42:37.080]for each of these countries.
[00:42:38.940]I mean all the main variables,
[00:42:40.470]and as you can see,
[00:42:41.910]yes the average year was around two tons per data
[00:42:44.216]but there was a large variability in countries
[00:42:46.890]and between countries and there was also a large variability
[00:42:50.520]in all the management practices that we analyzed
[00:42:53.940]like nitrogen inputs, (indistinct) and plant density.
[00:42:58.560]Also most of the farmers use hybrid seeds
[00:43:01.860]but still 20% use open combination varieties
[00:43:05.850]or seed that they say from previous seasons,
[00:43:10.106]in the soil (indistinct) in each climate zone,
[00:43:14.078]and only 35% apply only pesticide.
[00:43:16.140]That was mostly an insecticide for only one.
[00:43:21.210]Okay, so we have
[00:43:26.580]two main research,
[00:43:28.770]three research question here,
[00:43:31.380]to answer the following.
[00:43:32.550]The first research question about what management
[00:43:35.580]practices lead to higher yields, we use two approaches.
[00:43:38.657]One is a conditional inference tree analysis
[00:43:41.100]at each climate zone,
[00:43:42.870]and on the other hand we use a gradient boosting machine
[00:43:45.630]that is a machine learning model
[00:43:47.579]with many environmental co-variables, (indistinct).
[00:43:52.995]So this is the
[00:43:55.140]results from this conditional inference tree analysis
[00:43:58.020]across the climate zones.
[00:44:00.057]Here we have the relative importance of different
[00:44:03.670]management practices
[00:44:06.605]and this relative importance is how many times
[00:44:08.670]or the frequency
[00:44:10.140]where a management practices have a positive effect on yield
[00:44:16.260]in each of the sites or climate zones.
[00:44:19.680]And here you can see also of the importance
[00:44:23.423]of these variable in different countries.
[00:44:30.232]Of course nitrogen rate was the most important variable,
[00:44:34.665]after all we need,
[00:44:36.170]but also hybrid seed use was very important
[00:44:38.021]and fertilizer replacement,
[00:44:39.396]Pesticide use, (indistinct).
[00:44:45.880]And fertilizer replacement is whether
[00:44:47.657]fertilizer in the whole as opposed
[00:44:50.642]to just (indistinct) of fertilizer.
[00:44:55.470]And then so and also very important
[00:44:58.067]and we have press damaged pesticide use
[00:45:00.582]and fertilization, (indistinct) fertilization rate.
[00:45:05.250]With the machine learning model we can create this kind
[00:45:07.710]of plot where we see the, in the impact
[00:45:11.718]or average impact
[00:45:13.130]of each management practice across all observations.
[00:45:19.680]So these are added effects.
[00:45:21.420]So for example, when the machine learning model
[00:45:24.810]tell us is that
[00:45:28.710]while keeping all things constant
[00:45:31.200]when we increase emission fertilization rate, the effect
[00:45:34.665]that will have on new we,
[00:45:37.980]around 400 clearance all of things considered
[00:45:43.680]and then the serious effects for
[00:45:45.246]like fertilizer rate, plant density
[00:45:47.970]which has a very strong effect
[00:45:50.580]and so they also have very strong effect
[00:45:52.677]and here we the effect of our practices.
[00:45:55.620]So all these are additive values
[00:45:58.412]that were I mean believed by the machine learning model
[00:46:03.242]in terms of the fact that we have across
[00:46:05.370]all observations in Sub-Saharan Africa.
[00:46:11.520]Then the second question gradient (indistinct),
[00:46:13.110]was what the yield of farmers adopting good agronomy,
[00:46:15.990]and again we did this in two ways,
[00:46:17.950]first in classifying farmers by
[00:46:19.980]technology adoption,
[00:46:21.485]and we use (indistinct) models to with technology level
[00:46:24.650]as a fixed effect and climate zone and random effects to
[00:46:29.594](indistinct) of farmers electing
[00:46:31.440]different technology levels.
[00:46:33.270]And then we also use the machine learning model
[00:46:35.310]to produce the, that will be achieved by helping set of
[00:46:38.455]with management practice across Africa.
[00:46:42.120]So this is a result for, for the miss effect model
[00:46:47.340]here in red you have the average yield
[00:46:49.607]across sub-saharan Africa from FAOSTAT.
[00:46:54.270]So it's little less than two tons per data,
[00:46:58.230]and that's very similar to the average yield
[00:47:00.406]of the baseline management of the farmer fields we have,
[00:47:06.960]So this is showing the farmers that use local seed
[00:47:11.670]or even valuation varieties, many low nitrogen rates,
[00:47:15.480]low densities and average to late sowing dates.
[00:47:20.310]And then you can see how the yield increases as we
[00:47:22.890]incorporate more technology in this field
[00:47:27.887]and a high nitrogen fertilization rate
[00:47:32.040]of about 50 kilograms per hectare,
[00:47:34.470]and 20 kilograms of phosphorus.
[00:47:36.570]We gain one ton of yield
[00:47:39.965]so we jump to three tons per hectare
[00:47:44.610]but of course this is are expensive of high cost
[00:47:46.590]because we need to buy nitrogen and seeds.
[00:47:50.520]But then we have our management company
[00:47:52.590]that are relatively low cost
[00:47:55.140]and then we keep like keep increasing yield.
[00:47:58.380]For example, if we move a plant density from three
[00:48:00.630]to five pounds per square meter, we gain another ton
[00:48:05.338]of yield and if on top of that we planted
[00:48:10.498]properly, we gained another 400 kg of yield.
[00:48:16.333]So between the higher technology yield
[00:48:19.440]that we are serving in our farmers fields compared
[00:48:22.620]with the base now we have a 2.2 times increase
[00:48:26.820]and this is still very low compared to the yield that is
[00:48:29.220]obtaining station on station trials.
[00:48:32.430]So this is the average unit of station trials
[00:48:35.335]for many maize varieties in Rwanda
[00:48:37.740]and this still (indistinct) to the potential yield.
[00:48:42.274]So we are managing yield significantly
[00:48:44.667]but we still have a long way to go to increase yield
[00:48:48.376]to the 80% of the yield potential.
[00:48:53.040]And if we compare the results with the one done
[00:48:55.920]with the machine learning model where we yield
[00:48:59.100]across sub-Saharan Africa
[00:49:00.840]and we also the same kind of increase from a baseline
[00:49:04.560]of about two tons per hectare to one intensified yield,
[00:49:08.430]I mean with best management practices
[00:49:10.350]of open to (indistinct).
[00:49:16.290]So finally the last thing we need, okay, so we know
[00:49:20.040]that if farmers here allowed this set
[00:49:23.610]of management practices they will need to close
[00:49:25.893]the limit gap this much,
[00:49:28.380]what will happen if all the farmers
[00:49:29.970]in Sub-Saharan Africa close
[00:49:31.616]the yield gap to the same level.
[00:49:36.456]So we call this scenario acceleration of yield gain
[00:49:40.663]and extrapolating this yield gap
[00:49:42.834]closer level to the rest of sub-Saharan Africa,
[00:49:47.580]we brought the average yield, 4.2 (indistinct),
[00:49:52.525]so this is a little lower than the 4.4 that
[00:49:54.410]we show before because conditions across
[00:49:56.450]sub-Saharan Africa might be a
[00:49:58.277]little less beneficial for maize production
[00:50:02.100]but still very close, 4.2.
[00:50:06.600]And if we keep the same maize area that we have
[00:50:10.050]in the last years in this region
[00:50:12.900]with this increase we reach 168 million tons of production.
[00:50:17.910]That is not yet still the demand that is paid
[00:50:22.629]by 2050 but it's 90% of it.
[00:50:25.680]So it's very close to maize demand
[00:50:29.460]and the balance is a negative 16 million tons.
[00:50:32.370]That means that you will have to import 16 million tons
[00:50:35.957]of maize or incorporate four more million
[00:50:40.770]tons of maize area.
[00:50:44.160]So it's a much, a much more confident,
[00:50:49.890]good situation than
[00:50:51.242]if we keep the same developing rate.
[00:50:56.083]So the message here is that we pledge for good agronomy
[00:51:00.510]for sub-Saharan Africa, for security.
[00:51:03.330]We have shown here that with simple combination
[00:51:05.850]of basic good agronomy practices we can deliver large
[00:51:10.938]and quick yield increases in the region.
[00:51:14.928]And yeah we think
[00:51:18.960]that agricultural research and development programs
[00:51:21.180]and policies should be tuned to facilitate farmer access
[00:51:23.691]to input and technical information.
[00:51:26.699]This is the kind of work that the (indistinct)
[00:51:29.760]has been doing the last years in this region.
[00:51:33.261]We hope that most of these kind of initiatives
[00:51:37.320]are in place and can be scaled in all the regions
[00:51:42.180]so we have much better protective
[00:51:46.920]for the future and really the current high demand
[00:51:50.247]and reliance on imports with argue
[00:51:53.306]that the region should not be used
[00:51:54.333]as a testing ground
[00:51:55.350]for approaches that have been largely improved under the
[00:51:57.810]capacity to increase yields
[00:51:59.400]and that we must focus on good agronomy practices
[00:52:02.160]that we know that can increase yields.
[00:52:05.040]And with that I want to finish
[00:52:06.960]and thank you for listening me.
[00:52:09.467]It was a long presentation,
[00:52:11.216]how you have any-
[00:52:16.826]Well thank you very much for that great presentation,
[00:52:19.170]very insightful
[00:52:20.220]and a lot of great informative information there.
[00:52:23.280]Any questions from the crowd?
[00:52:26.940]All right.
[00:52:28.680]I'll start. So great presentation.
[00:52:30.690]Thank you and full appreciate everything we've done.
[00:52:34.770]Okay great. And so I think one of our issues
[00:52:37.980]and challenges going forward in agriculture period is
[00:52:40.560]to move the conversation from just yield to economic yield
[00:52:43.980]and then sustainable yield.
[00:52:45.060]So as you think about where you have the most data points,
[00:52:48.270]so most information, what would it take
[00:52:52.260]to use these tools to start to develop into saying
[00:52:55.920]what is your economic yield potential?
[00:52:59.730]That's one thing. And then you can start to say
[00:53:01.978]what is your economic sustainable yield potential given
[00:53:05.658]you know, environmental degradation,
[00:53:08.520]water quality, those sorts of things.
[00:53:10.590]And then ultimately I think where we wanna start
[00:53:12.870]to use these tools is
[00:53:14.670]to think about economic sustainable yield over
[00:53:17.460]five seasons, right?
[00:53:18.450]So what is the right crop, what's the right crop rotation,
[00:53:21.480]how do you manage it and start
[00:53:23.130]to use all these powerful tools to do that.
[00:53:27.060]So what is your feeling?
[00:53:29.580]Number one, do you read what I said is important
[00:53:31.770]and number two, do you believe some of that is doable now
[00:53:35.640]or how long do we have to wait?
[00:53:38.400]Yeah, okay. So yes, I mean,
[00:53:43.387]going from like to the beginning I think it's global.
[00:53:46.800]I really think it's global. I think,
[00:53:48.450]I mean there are some roles that are exploring this kind
[00:53:51.450]of analysis right now.
[00:53:53.430]Always having the economical component
[00:53:56.160]to this analysis is a little challenging when you work at
[00:53:59.221]traditional or global level.
[00:54:01.697]So I think it's this kind of analysis are more suited for
[00:54:05.664]a local, I mean it could be at the
[00:54:08.863]solution in the country or the country level
[00:54:12.880]because then you will have many more variables that are
[00:54:18.660]playing a role there relating to different prices
[00:54:23.040]that you could have valuation prices between (indistinct)
[00:54:27.935]or even if the policy changes then will have an
[00:54:30.180]impact, a big impact.
[00:54:31.710]So I think that these kind
[00:54:33.870]of tools can give a first step toward this analysis
[00:54:37.350]and then it will be more like work I find,
[00:54:42.630]yeah I refine work at the local level to try
[00:54:46.440]to make a better recommendation for farmers
[00:54:51.019]in terms of what the economical optimum level
[00:54:55.320]but it will, yeah it will depends a level of local context.
[00:54:58.760]So yeah, I think that this is like a more broad global
[00:55:03.240]or regional assessment
[00:55:04.140]and then you can move in at the local place.
[00:55:07.260]Yeah.
[00:55:12.600]Early in your seminar you talked about weather stations.
[00:55:16.530]What is the source of this weather information
[00:55:19.170]and does that include Langley's
[00:55:22.230]or like quantity information,
[00:55:26.400]do you use that and that sort of thing?
[00:55:29.010]Yeah, so we use well
[00:55:33.510]information in like two source, I mean
[00:55:38.040]twice in the first analysis, let's go back,
[00:55:43.800]I mean first we have the bottom up approach
[00:55:46.445]of the lower (indistinct)
[00:55:47.278]to get estimate the potential for a specific site.
[00:55:51.570]So there the great information is usually
[00:55:56.850]when possible is measured where data,
[00:56:01.860]they where data from weather stations
[00:56:06.269]and they're subject to a lot of quality control as well.
[00:56:09.600]So for the bottom up global legal optimized
[00:56:13.710]estimates a few potential,
[00:56:15.420]they use metal measurable information.
[00:56:19.380]Then for this,
[00:56:21.420]for the machine learning let's say interpolation
[00:56:23.940]or extrapolation,
[00:56:26.160]to get good estimates of your potential,
[00:56:31.500]this is a product I use,
[00:56:33.409](indistinct) that is a product where
[00:56:35.790]very interrelated climatic records
[00:56:40.050]from weather stations.
[00:56:41.640]But these are like long-term climatic records.
[00:56:44.850]So it's not like daily information
[00:56:48.163]from a weather station it's more like
[00:56:50.700]what is the average precipitation during summer,
[00:56:53.370]what is the average precipitation during this month?
[00:56:56.580]So it's two kind of sources.
[00:56:59.520]For (indistinct) modeling,
[00:57:01.350]we know that it's much better to,
[00:57:03.060]to have measure daily weather data.
[00:57:06.690]So then what we do is,
[00:57:10.350]okay we got a really good, let's say grand truth
[00:57:14.400]of the yield potential using this measure validate,
[00:57:17.850]measure weather data.
[00:57:20.400]And we use this machine learning model to interpolate
[00:57:23.728]and extrapolate those estimates to other sites.
[00:57:26.190]But we know that this is the grand truth
[00:57:27.990]and we rely on that to avoid making big mistake or take
[00:57:32.840]or making big biases when we use the gridded approaches.
[00:57:37.350]All right, yeah.
[00:57:40.290]I'm sorry I didn't make the question very clear I guess.
[00:57:43.680]Is this national weather service data
[00:57:46.050]or is it university or other?
[00:57:49.647]For the weather map is mostly national weather.
[00:57:52.960]Yeah. Yeah.
[00:57:53.910]So it'll depend on country
[00:57:55.620]to country depending on the availability of each country,
[00:57:58.830]for most of the countries national weather data, yeah.
[00:58:03.371]But yeah, it depends,
[00:58:05.484]but most of the countries are national, yeah.
[00:58:09.120]Thanks, Ferando.
[00:58:09.953]We have a common question from the online audience
[00:58:13.110]regarding the second part of your work.
[00:58:15.030]They're mentioned if, if you include that database
[00:58:17.700]or survey examples of big farms, private farms
[00:58:21.480]that they are actually applying good agronomic practices
[00:58:25.050]and if you can comment on that.
[00:58:27.330]Yeah, thanks for the question.
[00:58:28.203]I mentioned that no this is all data from
[00:58:31.740]smallholder farmers.
[00:58:33.600]I never mentioned smallholders, I should have done that.
[00:58:36.780]So (indistinct) goes with the smallholder farmers only.
[00:58:41.550]So yeah, I mean it would be interesting
[00:58:44.400]to include that kind of data.
[00:58:45.720]Probably we'll see our average yield very much higher.
[00:58:51.303]But for the purpose of this situation in most
[00:58:54.990]of us are in Africa where smallholders are the
[00:58:57.300]biggest challenge.
[00:59:00.030]I think. (chuckles)
[00:59:02.160]That's what we used. Yeah.
[00:59:07.440]Yeah Fernando, one question,
[00:59:08.730]you sure you can double yields
[00:59:10.530]in Africa with a bag of hybrid seed,
[00:59:12.870]with a bag of (indistinct)
[00:59:15.090]and a few more things there
[00:59:16.350]that we all know that works pretty well.
[00:59:19.440]That sounds like good revolution. (chuckles)
[00:59:21.780]So why?
[00:59:23.460]Why Africa has not gone
[00:59:24.810]through a green revolution if it is so easy to do it.
[00:59:29.850]Okay, yeah.
[00:59:31.740]So of course there are many challenges for farmers
[00:59:34.560]to adopt these practices.
[00:59:38.057]We have in the data we show, we have farmers
[00:59:42.180]that were subscribed to this one farm program
[00:59:45.660]where essential provides the farmer with access
[00:59:48.720]to credit and inputs
[00:59:52.080]and I mean they don't give the seeds
[00:59:55.693]and the fertilization for free but they provide credits
[01:00:00.283]and they provide recommendations and management practices.
[01:00:04.860]So not all farmers have access to that.
[01:00:07.933]And also, and I mean
[01:00:12.480]farmers usually is a situation where you have farmers
[01:00:16.980]that usually have another source of income where maybe
[01:00:22.454]agriculture is not the main activity.
[01:00:24.510]So they're more interested in the,
[01:00:26.237]of the income coming from other source.
[01:00:29.580]Also, usually they don't have extension services
[01:00:33.210]or access to a (indistinct) environment
[01:00:37.037]or about how to use this inputs
[01:00:41.010]and we also don't have good credits
[01:00:43.350]or there's a lot of uncertainty when they
[01:00:47.843]deal close about the price that are really good when they
[01:00:51.573]sell the seed at the end of the season,
[01:00:53.940]most of these farmers
[01:00:56.160]use much of the production for to,
[01:00:59.730]I mean they, their own food.
[01:01:03.420]So it is like a combination of many things that are,
[01:01:08.700]yeah, making that is really helpful for farmers to,
[01:01:13.170]to intensify the carbon system
[01:01:16.170]or there is not so much interest in on them
[01:01:18.360]depending on the situation.
[01:01:20.040]Yeah.
[01:01:21.150]Alright, well thank you very much Fernando
[01:01:24.346]for a great seminar today.
[01:01:26.667]We make sure to come check out the seminar
[01:01:29.040]next week, we'll have Dr. Desmond Lane
[01:01:31.305]from Auburn University,
[01:01:33.870]who's the department head of the Department
[01:01:35.566]of Horticulture there.
[01:01:37.410]It's gonna be here talking about their "Feed you program".
[01:01:40.290]See you all next week.
[01:01:42.233](audience applauds)

The screen size you are trying to search captions on is too small!

You can always jump over to MediaHub and check it out there.

Comments

0 Comments

Integrating Agronomy and Machine Learning to Analyze Yield Gap Magnitudes and Causes from Field to Global Levels.

Description

Searchable Transcript

Comments icon comment

Related Channels

Comments