Hass Avocado Alternate Bearing: Why Forecasting the Harvest Starts by Knowing Whether This Year Will Bear Fruit

According to Analdex, Colombia exported 201.479 tons of Hass avocado in 2025, 45.6 % more than the previous year. Behind every container there is a volume commitment with a buyer, a hired crew, and a packing plant on schedule. An error of dozens of tons in the forecast is paid for in empty trucks or unfulfilled orders.

That is why our TECDE research team built and tested a Machine Learning model on a commercial farm to forecast Hass harvests, and is presenting it at Territorio Aguacate, the avocado industry gathering in the country.The study was carried out on a commercial farm with about 42 hectares, 11,000 trees in production distributed across three blocks planted in 2020 and at 1,600 m above sea level. The same three Hass avocado blocks, with the same trees and the same management, delivered 892 tons in May 2025 and 117 tons in May 2026, a harvest more than five times smaller than that of May 2025. There was no pest outbreak, no frost, and no crew mistake. It was alternate bearing, the natural up-and-down of Hass between a heavy year and a light one, and it is the reason the average of previous years almost never gets the next harvest right.

That drop has two ingredients that are worth separating. The first is the calendar: in Colombia, Hass usually has a main harvest and a secondary harvest, and the secondary one is smaller. The second is alternate bearing: the natural swing between heavy and light crop years. The two add up, and neither is captured by the average of previous years.

A tree that does not bear the same way two years in a row

Next, we will make a slightly more theoretical introduction that will allow us to understand much better: what alternate bearing is and why it breaks averages, what other studies have shown and how ours differs, how we built the model, and what happened when the forecast went up against the scale in a real harvest.

Alternate bearing, also known in the field as biennial bearing, is the tendency of a tree to alternate a high-crop year with a low-crop year. The difference can be enormous: in a Hass trial in California (Garner et al., 2011), the same trees produced 391 fruits in the high-crop year and only 32 in the next one, twelve times fewer.

The cause lies in the tree itself. When a year sets a lot of fruit, that load slows the growth of summer shoots, and those shoots provide most of the flowering for the following year Kour et al., 2018. The heavily loaded tree is left with nowhere to flower in the next season, and the light year that follows gives it time to recover shoots and bear again. It is a kind of savings account that the tree manages on its own.

In Colombia, the swing also overlaps with the calendar. In a study on three Colombian farms, the main harvest accounted for 71.4 % of annual production and the secondary harvest for 28.6 % (Cano-Gallego et al., 2024). On farms above 2,000 meters, the main harvest averaged 48 kilos per tree and the secondary one just over 7 (Cano-Gallego et al., 2024). And alternate bearing is not even: in a four-hectare block, production fell 37 % from one year to the next, with large differences even between neighboring trees (Zuluaga Mejía et al., 2024). That is why the same farm can have blocks in different phases of the cycle at the same time.All of this explains why the historical average fails. The average assumes that each year looks like the previous one, and alternate bearing says exactly the opposite: if last year was a heavy crop year, the average overestimates this year; if it was a light one, it underestimates it. And the difference is not a few kilos, but multiples.

What was already known and what this work adds

Before getting into the model, it is worth placing it alongside other studies, because we did not start from zero and because each one answers a different question:

• Measure how much a tree swings (Garner et al., California, 2011). It documented the magnitude of alternate bearing in Hass. It describes the phenomenon, but does not forecast a harvest.
• See yield from satellite, tree by tree (Robson et al., Australia, 2017). With very high-resolution satellite images, vegetation indices were related to fruit weight per tree (coefficients of determination above 0.68), and the red-edge band proved especially useful. That work explains variability within a block at a given moment; it does not anticipate the next harvest.
• Classify the year as high- or low-crop (Rahman et al., South Africa, 2025). With 46 commercial Hass blocks, they combined Sentinel-2 images and weather during flowering, and correctly identified the type of year with 0.88 accuracy. It says whether the year will be high or low, but not how many kilos.
• Commercial forecasting systems (for example growAG, Australia). They combine satellite data with targeted field sampling and report accuracies close to 93 %. This is not a direct comparison with ours, because the definition of accuracy and the crop are not the same.

What our work adds is another piece: estimating the kilos for each month and each block, in high-altitude Hass avocado at 1,600 meters, with free medium-resolution images (Sentinel-2) and a test against a real harvest that had already been closed, not just against historical data.

First whether there is a harvest, then how much

A Hass block does not produce every month. It has concentrated harvest windows and long zero periods. A model that tries to guess the kilos for each month mixes the months when there is nothing to pick with the harvest months, and ends up spreading a little fruit into empty months and falling short in full ones.The solution that worked best was to split the problem into two questions. The first is whether there will be a harvest in that month and in that block: yes or no. The second, which is answered only when the first is yes, is how many kilos. It is the same thing a good manager would do: before calculating how many crates to buy, confirm whether that block enters this month. In statistics, this scheme is known as a hurdle model.

We worked with the monthly records of each block between April 2022 and April 2026, from the time the trees were already producing: 147 month-block observations. It is a small sample, and that conditions everything that follows.To answer the first question, the signal that carried the most weight was the block’s position within the alternate-bearing cycle, with about 40 % of the importance. It was followed by the block’s usual harvest window (about 15 %), the number of harvests it has already produced (13 %), and the average temperature (12 %). In other words, whether there is a harvest or not depends mainly on where the block is in the swing, more than on that month’s weather. This matches the conclusion of the South African study: alternate bearing is a pattern that can be anticipated, not chance.

The classifier got every harvest month right in validation and testing, but there were only three harvest months in each. In addition, the Hass calendar is fairly predictable, so we do not take that as proof that it is infallible.

How many kilos: a simple base and a correction

For the second question, the model starts from an agronomic base: how many productive trees each block has, how old they are, and how much a tree of that age usually produces on this farm. That base, by itself, is not enough: it had the highest error of all the versions tested, because it does not see alternate bearing or the conditions of each season. What makes it useful is a second layer that learns to correct it with what has happened in the block. This idea, starting from an agronomic base and letting the statistical model learn the remaining error on its own, comes from the hybrid-model literature, which in other crops has shown improved accuracy when data are scarce (for example, Shahhosseini et al., 2021, in corn).The signals that mattered most in that correction were three:

• The age of the block (about 36 %). The trees were planted in 2020 and are still moving toward full production, so every month of age matters. With trees this young, age also acts as the clock of the production curve, and that is why we cannot fully separate the effect of age from that of alternate bearing.
• The accumulated rainfall two months before harvest (about 19 %). In statistics this is called a lag: using data from a few months ago to explain what we see today. For a May harvest, it is the rainfall from March. That date falls right in the fruit-filling stage, when the fruit gains size and weight, and not in flowering, which happens many months earlier. That is why we read it as a signal of how the fruit arrives at the final stretch, not as the moment when the number of fruits is defined.
• The canopy vigor from the previous month (about 19 %). It comes from Sentinel-2, a free satellite that passes over the same point every five days and distinguishes details at ten meters. We used EVI2, an index of how green and active the foliage is. The best-known greenness index (NDVI) contributed almost nothing, because in a mature avocado orchard the canopy is so dense that the index stays stuck near the top and no longer distinguishes one tree from another. In Australia, the red-edge band was the one most strongly related to kilos per tree; in our farm, the red-edge index was not among the heaviest signals, and with three blocks and few harvests we cannot say for sure at this stage why.

This brings us to a limitation we prefer to state plainly: the model does not see flowering or fruit set. Our forecast was issued with data up to April for the May harvest, that is, one month ahead, when the fruit load is already defined. It is useful for planning harvest, packing, and logistics; forecasting six months in advance would require measuring flowering, and that is one of the lines we want to pursue.

To measure how much each piece contributes, we tested four versions and compared them over six test months the model had never seen. The base alone had the highest error; the seasonal historical average for each block, which is what many growers use, performed better; asking the model for kilos directly, without the prior question of whether there is a harvest, landed in the middle; and the full system had the lowest error. Compared with the historical average, the error fell by about 30 %, and compared with the version without the prior question it fell to about one third. There are 18 observations, so we read this as a trend, not a definitive figure.

What the satellite does not see: the fruit already hanging

One piece remains that the satellite does not see: the fruit already hanging on the tree. For that, the system uses a field count, with a video around a sample of trees. As we already covered in our blog Estimate your harvest with an algorithm, and know how much to trust it, which you can read here, a camera only counts what it sees and the leaves hide part of the fruit.That is why the count is not used raw: it is corrected with manually counted trees (60 plants in four blocks) and enters at the end as an adjustment to the forecast, not as its base. With that correction, the per-plant counting error fell 63 %, and in block totals it came in at 8.8 % on average, with a range from 4.2 % to 18.6 % depending on the block.

Knowing how much fruit will come out of your farm before committing it

How many kilos of Hass will you have in the next harvest, block by block, and with what margin of error can you promise it? On most farms that number comes from the average of past years or from a walk through the block, and alternate bearing takes care of making the promise not match what arrives at the packing house. Harvis helps you bring together in one place the harvests of each block, the weather, satellite images, and field counts, and delivers a forecast by block and for the whole farm, with its range and with a reading of where the error may be. Write to us and we will set it up on the blocks you already have.

May 2026, the forecast against the scale

A forecast model can look very good on paper because it is tested with data it already knew and fail on the farm. That is why this test was done the other way around: the model was closed with the information available up to April 2026, issued its forecast for the May secondary harvest, and then was compared with what was actually weighed. On the whole farm, the forecast was 184,000 kg and 177,232 kg were harvested: a deviation of 3.7 %, that is, 96.3 % accuracy on the total for that harvest. For anyone negotiating with an exporter or scheduling the packing plant, that is the number that matters.

Block by block, the story is different. The two large blocks, which together accounted for about 95% of production, stayed within less than five points: one came out 4.5% above the forecast and the other 2.2% above. The small block, with just 4.8% of the volume, produced less than half of what had been forecast.

Why did the total come out so well if one block failed so badly? Because the errors had opposite signs and the block that failed the most carried little volume. When the three were added together, some errors offset others. That compensation is not a model trick, but it is not guaranteed either: if all blocks miss in the same direction, the sum does not correct anything.

For block A we have two hypotheses: . The first is that it is a block that historically receives less light, and that there each stage of the fruit takes longer than in the others; our model only uses sunshine hours, an indirect measure that carried little weight, so it does not capture that effect well. The second is that it was the block with the lowest sampling coverage in the field count.Measuring the actual radiation by block and increasing the number of counted trees are the next steps to test it.

The field count also taught its lesson. Without the adjustment from the counted trees, the model projected about 16% more fruit than was harvested; with the adjustment, it came down to 3.7%. And the video count, corrected with the manually counted trees, reduced by 63% the error of counting only what the camera could see, and left an error of 8.8% in block totals. For comparison, an Australian tool that also combines satellite data and targeted field counting reports accuracies above 90% at block level.What we do know for sure is that: with only one harvest closure we cannot say whether that improvement came from the count information or from the fact that the recommended reading turned out to be more conservative in a moderate season.

What a farm forecast does not say about each block

First, the real test is only one harvest. The 96.3% came from the May 2026 secondary harvest and is a result for that season, not a guaranteed average for the ones to come. In previous seasons, especially in the face of very high production peaks, the model deviated much more, and that is why the test has to be repeated harvest after harvest.

Second, the model was designed to get the farm right, not each block. At the month-and-block scale it explains a moderate share of what happens (a coefficient of determination close to 0.37), which is common in perennial crops, where each tree responds in its own way.

Third, some parts of the system still have small samples. The video count was calibrated with 60 trees in four blocks, and in months with heavy cloud cover the satellite cannot get clean images of the farm. In addition, because the blocks are still entering production, age and alternate bearing are mixed in the data; testing the method on mature blocks is part of what comes next.

Fourth, these figures are from one farm: three blocks planted in 2020, at about 1,600 meters, with its own material and management. On another farm, with a different altitude, different tree age, or another harvest pattern, the model base and the correction have to be adjusted again with that farm’s data. What does carry over to any farm is the method: separate whether there is a harvest from how much, and measure the error on the total.

The forecast is judged on the whole farm, not on the tree

Alternate bearing makes Hass a difficult crop to forecast with averages, but not impossible to forecast. What changes is the way the question is asked: first whether each block will harvest, depending on where it is in the swing, and then how much, with block age, flowering rainfall, and canopy vigor as the main signals. And the result is measured where commercial decisions are made, on the whole farm, knowing that one block can fail even if the total is right.

The grower who does this exercise will arrive at the next negotiation knowing how much can be promised and with what margin; the one who keeps using the average of past years will discover at the packing house that alternate bearing had already decided something else.

Do you want to know how much fruit will come out of each block and your whole farm, with its range, before committing it to the buyer? With Harvis we bring together your harvests, the weather, satellite images, and field counts in one place, and deliver a forecast by block and for the farm, with its margin of error. Write to us and we will set it up on the blocks you already have.

References consulted

1. Ochoa Cortés, J. S., Ceballos Múnera, F. and Ceferino Marín, D. (2026). Biennial bearing and harvest forecasting in Hass avocado, a satellite-agronomic hybrid model validated in the field with an aggregate accuracy of 96%. TECDE Colombia S.A.S. Article in process of publication.
2. Analdex (2026). Avocado exports grew by 21.3% during 2025. National Foreign Trade Association. https://analdex.org/wp-content/uploads/2026/02/Nota-de-prensa-exportaciones-aguacate.pdf
3. Garner, L. C., Klein, G., Zheng, Y., Khuong, T. and Lovatt, C. J. (2011). Response of evergreen perennial tree crops to gibberellic acid is crop load-dependent, II. GA3 increases yield and fruit size of Hass avocado only in the on-crop year of an alternate bearing orchard. Scientia Horticulturae, 130(4), 753–761. https://doi.org/10.1016/j.scienta.2011.08.033
4. Córdoba-Gaona, O. J., Hernández-Arredondo, J. D., Loaiza-Ruiz, R. A., Barrera-Sánchez, C. F. and Cano-Gallego, L. E. (2024). Rootstock/scion compatibility effects on yield components in avocado cv. Hass in three locations in Colombia. Revista Colombiana de Ciencias Hortícolas, 18(1), e17074. https://doi.org/10.17584/rcch.2024v18i1.17074
5. Lovatt, C. J. (n.d.). Developing field strategies to correct alternate bearing (II). University of California Riverside and California Avocado Commission. https://www.californiaavocadogrowers.com/sites/default/files/Developing-Field-Strategies-to-Correct-Alternate-Bearing.pdf
6. Ramírez-Gil, J. G. (2017). Fruit quality of avocado with applications of NAA, boron, nitrogen, sucrose and girdling. Agronomía Mesoamericana, 28(3), 591–603. https://doi.org/10.15517/ma.v28i3.23688
7. Rahman, M. M., Robson, A. and Bekker, T. (2025). Machine learning approaches for assessing avocado alternate bearing using Sentinel-2 and climate variables, a case study in Limpopo, South Africa. Remote Sensing, 17(24), 3935. https://doi.org/10.3390/rs17243935
8. European Space Agency (n.d.). Sentinel-2, facts and figures. ESA. https://www.esa.int/Our%5FActivities/Observing%5Fthe%5FEarth/Copernicus/Sentinel-2/Facts%5Fand%5Ffigures
9. Robson, A., Rahman, M. M. and Muir, J. (2017). Using WorldView satellite imagery to map yield in avocado (Persea americana), a case study in Bundaberg, Australia. Remote Sensing, 9(12), 1223. https://doi.org/10.3390/rs9121223
10. Shahhosseini, M., Hu, G., Huber, I. and Archontoulis, S. V. (2021). Coupling machine learning and crop modeling improves crop yield prediction in the US Corn Belt. Scientific Reports, 11, 1606. https://www.nature.com/articles/s41598-020-80820-1
11. Dunn, C. (2021). World-first yield forecasting technology offers avocado growers 93% accuracy. growAG, AgriFutures Australia. https://www.growag.com/article/world-first-yield-forecasting-technology-offers-avocado-growers-93-accuracy
12. Cano-Gallego, L. E., Bernal-Estrada, J. P., Hernández-Arredondo, J. D., Correa-Londoño, G. A., & Córdoba-Gaona, O. J. (2024). Rootstock/scion compatibility effects on yield components in avocado cv. Hass in three locations in Colombia. Revista Colombiana de Ciencias Hortícolas, 18(1), e17074. doi.orghttps://www.researchgate.net/publication/381137302RootstockscioncompatibilityeffectsonyieldcomponentsinavocadocvHassinthreelocationsin_Colombia
13. Zuluaga Mejía, C., Bernal Estrada, J. A., Loaiza Ruiz, R. A., Monsalve García, D. A., Barrera Sánchez, C. F., Córdoba Gaona, O. J., Tamayo Vélez, Á. de J., Castaño Marín, Á. M., Cano Gallego, L. E., & Londoño Bonilla, M. de J. (2024). Productive model for Hass avocado cultivation in the department of Antioquia. AGROSAVIA Publishing. doi.org.https://repository.agrosavia.co/items/cf05a657-e936-4bb5-a400-202d76f75c39